The Application of Unmanned Aerial Vehicle to Precision ...
Transcript of The Application of Unmanned Aerial Vehicle to Precision ...
Utah State University Utah State University
DigitalCommons@USU DigitalCommons@USU
All Graduate Theses and Dissertations Graduate Studies
5-2016
The Application of Unmanned Aerial Vehicle to Precision The Application of Unmanned Aerial Vehicle to Precision
Agriculture: Chlorophyll, Nitrogen, and Evapotranspiration Agriculture: Chlorophyll, Nitrogen, and Evapotranspiration
Estimation Estimation
Manal Elarab Utah State University
Follow this and additional works at: https://digitalcommons.usu.edu/etd
Part of the Civil and Environmental Engineering Commons
Recommended Citation Recommended Citation Elarab, Manal, "The Application of Unmanned Aerial Vehicle to Precision Agriculture: Chlorophyll, Nitrogen, and Evapotranspiration Estimation" (2016). All Graduate Theses and Dissertations. 4891. https://digitalcommons.usu.edu/etd/4891
This Dissertation is brought to you for free and open access by the Graduate Studies at DigitalCommons@USU. It has been accepted for inclusion in All Graduate Theses and Dissertations by an authorized administrator of DigitalCommons@USU. For more information, please contact [email protected].
THE APPLICATION OF UNMANNED AERIAL VEHICLE TO PRECISION
AGRICULTURE: CHLOROPHYLL, NITROGEN, AND
EVAPOTRANSPIRATION ESTIMATION
by
Manal Elarab
A dissertation submitted in partial fulfillment of the requirements for the degree
of
DOCTOR OF PHILOSOPHY
in
Civil and Environmental Engineering
Approved:
______________________ ____________________ Dr. Mac McKee Dr. Niel Allen Major Professor Committee Member
______________________ ____________________ Dr. David Stevens Dr. Douglas Ramsey Committee Member Committee Member
______________________ ____________________ Dr. David Rosenberg Dr. Mark McLellan Committee Member Vice President for Research and Dean of the School of Graduate Studies
UTAH STATE UNIVERSITY Logan, Utah
2016
ii
Copyright © Manal Elarab 2016
All Rights Reserved
iii
ABSTRACT
The Application of Unmanned Aerial Vehicle to Precision Agriculture:
Chlorophyll, Nitrogen, and Evapotranspiration Estimation
by
Manal Elarab, Doctor of Philosophy Utah State University, 2016
Major Professor: Dr. Mac McKee Department: Civil and Environmental Engineering
Precision agriculture (PA) is an integration of a set of technologies aiming to
improve productivity and profitability while sustaining the quality of the surrounding
environment. It is a process that vastly relies on high-resolution information to enable
greater precision in the management of inputs to production. This dissertation explored
the usage of multispectral high resolution aerial imagery acquired by an unmanned aerial
systems (UAS) platform to serve precision agriculture application. The UAS acquired
imagery in the visual, near infrared and thermal infrared spectra with a resolution of less
than a meter (15 - 60 cm). This research focused on developing two models to estimate
cm-scale chlorophyll content and leaf nitrogen. To achieve the estimations a well-
established machine learning algorithm (relevance vector machine) was used. The two
models were trained on a dataset of in situ collected leaf chlorophyll and leaf nitrogen
measurements, and the machine learning algorithm intelligently selected the most
appropriate bands and indices for building regressions with the highest prediction
accuracy. In addition, this research explored the usage of the high resolution imagery to
iv
estimate crop evapotranspiration (ET) at 15 cm resolution. A comparison was also made
between the high resolution ET and Landsat derived ET over two different crop cover
(field crops and vineyards) to assess the advantages of UAS based high resolution ET.
This research aimed to bridge the information embedded in the high resolution imagery
with ground crop parameters to provide site specific information to assist farmers
adopting precision agriculture. The framework of this dissertation consisted of three
components that provide tools to support precision agriculture operational decisions. In
general, the results for each of the methods developed were satisfactory, relevant, and
encouraging.
(148 Pages)
v
PUBLIC ABSTRACT
The Application of Unmanned Aerial Vehicle to Precision Agriculture:
Chlorophyll, Nitrogen, and Evapotranspiration Estimation
by
Manal Elarab, Doctor of Philosophy Utah State University, 2016
Major Professor: Dr. Mac McKee Department: Civil and Environmental Engineering
Precision agriculture (PA) is an integration of a set of technologies aiming to
improve productivity and profitability while sustaining the quality of the surrounding
environment. It is a process that vastly relies on high-resolution information to enable
greater precision in the management of inputs to production. This dissertation explores
the usage of multispectral high resolution aerial imagery acquired by an unmanned aerial
systems (UAS) platform to estimate three fundamental crop parameters (plant
chlorophyll, leaf nitrogen and crop water demand). The UAS acquired imagery in the
visual, near infrared and thermal infrared spectra with a resolution of less than a meter
(15 - 60 cm). This research applied a well-established machine learning algorithm
(relevance vector machine) to the five band imagery (R, G, B, NIR, and TIR) to estimate
plant chlorophyll content and leaf nitrogen respectively. In addition, this research
explored the usage of the high resolution imagery to estimate crop evapotranspiration
(ET) at 15 cm resolution. A comparison was also made between the high resolution ET
and Landsat derived ET over two different crop cover (field crops and vineyards) to
vi
assess the advantages of UAS based high resolution ET. This research aims to bridge the
information embedded in the high resolution imagery with ground crop parameters to
provide site specific information to assist farmers adopting precision agriculture.
Manal Elarab
vii
I dedicate this work to Mohammad, Fairouz, Nada and Ahmad.
Your unconditional love and support created the person I am today, You made me believe in myself when I was in doubt.
I want to make you proud always.
… In the loving memory of Musa Nimah
Love you Manal
viii
ACKNOWLEDGMENTS
I am very grateful to my adviser and mentor, Dr. Mac McKee, for giving me this
opportunity, providing me with such a great project and team to work with, and above all
teaching me how to be a better researcher and person on daily basis. I would like to
extend my gratitude to all my committee members. I would like to thank Dr. Alfonso
Torres-Rua and Dr. Andres Ticlavilca for their guidance, help and support throughout my
education at USU. I appreciate all those who helped me collect and process the data
needed in my research (Roula Bachour, Shannon Syrstad, Mark Winkelaar, Alfonso
Torres, Ian Gowing, etc.). Thank you to all the AggieAir team that helped in the Scipio
project.
I also acknowledge all the members of AggieAir over the past few years who
have contributed indirectly to this work. Most of all, I owe this work to my loving
husband for his endless sacrifices and relentless support.
ix
CONTENTS
Page
ABSTRACT……………………………………………………………………………. iii PUBLIC ABSTRACT…………………………………………………………………… v DEDICATION…………………………………………………………………………. vii ACKNOWLEDGMENTS ............................................................................................... viii LIST OF TABLES ............................................................................................................. xi LIST OF FIGURES .......................................................................................................... xii CHAPTER 1: INTRODUCTION ....................................................................................... 1 Objectives ........................................................................................................................................ 3 Dissertation Organization ................................................................................................................ 4 Contributions ................................................................................................................................... 5 References ........................................................................................................................................ 6 CHAPTER 2: ESTIMATING CHLOROPHYLL FROM THERMAL AND BROADBAND MULTISPECTRAL HIGH RESOLUTION IMAGERY FROM AN UNMANNED AERIAL SYSTEM USING RELEVANCE VECTOR MACHINES FOR PRECISION AGRICULTURE ........................................................................................... 7 Abstract ............................................................................................................................................ 7 Introduction ...................................................................................................................................... 8 Materials and Methods ................................................................................................................... 14
RVMs ................................................................................................................... 14 Study Area ........................................................................................................... 18 Instrumentation: Remote Sensing Platform AggieAir ......................................... 19 Data Collection .................................................................................................... 23 Model Potential Inputs and Performance ............................................................. 24
Result and Discussion .................................................................................................................... 28 Conclusion ..................................................................................................................................... 34 Acknowledgment ........................................................................................................................... 36 References ...................................................................................................................................... 36 CHAPTER 3: USE OF HIGH RESOLUTION MULTISPECTRAL IMAGERY TO ESTIMATE PLANT NITROGEN IN OATS (AVENA SATIVA). ................................ 46 Abstract .......................................................................................................................................... 46 Introduction .................................................................................................................................... 47
x
Material and Methods .................................................................................................................... 51 Study Area ........................................................................................................... 51 Instrumentation: Remote Sensing Platform AggieAir ......................................... 52 AggieAir Data Processing ................................................................................... 53 Ground Data Acquisition ..................................................................................... 54 Relevance Vector Machine (RVM) ..................................................................... 54 Generating the Model Inputs and Targets ............................................................ 55 Model Configuration and Performance ............................................................... 57
Results and Discussion .................................................................................................................. 59 Conclusion ..................................................................................................................................... 66 References ...................................................................................................................................... 67 CHAPTER 4: ASSESSMENT ON THE USE OF AGGIEAIR UNMANNED AERIAL SYSTEM REMOTE SENSING CAPABILITIES TO ESTIMATE CROP EVAPOTRANSPIRATION AT HIGH SPATIAL RESOLUTION................................. 73 Abstract .......................................................................................................................................... 73 Introduction .................................................................................................................................... 74 Data Collection ..................................................................................................................................
Site Description.................................................................................................... 79 UAV Description ................................................................................................. 80 Sensors Description ............................................................................................. 82 Flight Details ....................................................................................................... 82
Processing Chain of AggieAir ....................................................................................................... 83 Mission Planning ................................................................................................. 83 Image Processing and Orthorectification ............................................................. 84 Radiometric Calibration ....................................................................................... 85 Thermal Calibration ............................................................................................. 86
Models: METRIC and METRIC High Resolution (METRIC-HR) ............................................... 87 METRIC .............................................................................................................. 88 METRIC-HR ....................................................................................................... 90 Albedo .................................................................................................................. 94
Results and Discussion .................................................................................................................. 96 Hot and Cold Pixel Selection ............................................................................... 96 METRIC Results.................................................................................................. 99 METRIC-HR Results ......................................................................................... 101 Comparison in Site S ......................................................................................... 104 Comparison in Site G ......................................................................................... 107
Conclusion ................................................................................................................................... 111 References .................................................................................................................................... 112 CHAPTER 5: SUMMARY.............................................................................................. 120 APPENDIX ................................................................................................................................ 123 CURRICULUM VITAE .......................................................................................................... 130
xi
LIST OF TABLES
Table Page
2.1 AggieAir Unmanned Aerial System Specifications .............................................. 20
2.2 Statistical Description of the Dataset Used for Plant Chlorophyll Estimation ..... 25
2.3 Vegetative Indices Formulation Used for Plant Chlorophyll Estimation ............. 26
2.4 Potential “Best Scenarios” .................................................................................... 29
3.1 Vegetative Indices Formulation Used for Leaf Nitrogen Estimation ................... 56
3.2 Statistical Description of the Dataset Used for Leaf Nitrogen Estimation ........... 57
3.3 Selected Model Characteristics ............................................................................. 60
4.1 AggieAir Unmanned Aerial System Features Description ................................... 81
4.2 The Developed Correction Equations for Each Band ........................................... 94
4.3 Details on the Selected Hot and Cold Pixels from the AggieAir and Landsat Imagery ................................................................................................................. 98
xii
LIST OF FIGURES
Figure Page
2.1 The location of the study area in Utah (left), the quarter used in plant chlorophyll estimation model (right) . .................................................................. 18
2.2 AggieAir airframe layout. ..................................................................................... 19
2.3 Relative spectral response of the VIS-NIR (left) and thermal camera (right). ..... 21
2.4 Raw natural color images from the unmanned aerial system (left), accurate orthorectified mosaic image from EnsoMOSAIC (center), and radiometric calibration of visible spectrum image (right). ....................................................... 22
2.5 Measured versus predicted chlorophyll concentration in the three flights for Model 4. Vertical yellow lines separate the three flight dates. ............................. 31
2.6 Model 4: Residual plot of the three flights (left) and one-by-one plot excluding the bare soil–zero chlorophyll points and reflecting the chlorophyll meter accuracy (right). .................................................................................................... 31
2.7 True-color maps, normalized difference vegetation index (NDVI) maps, leaf area index (LAI) maps, and the estimated chlorophyll concentration (μg.cm-2) map for the three different dates representing early growth, midgrowth, and early flowering. ..................................................................................................... 32
2.8 True-color image of flight zero (left), leaf area index (LAI) map (middle), and estimated chlorophyll concentration map (μg.cm-2, right). ................................... 33
3.1 The location of the study area in Utah (left), the quarter used in plant chlorophyll estimation model (right) .................................................................... 52
3.2 Measured versus predicted leaf nitrogen in the two flights used to build the model (above) and residual plot of the selected mode (below). ........................... 62
3.3 True-color maps (left), the estimated plant leaf nitrogen estimate map. .............. 63
3.4 June 17 image in false color (left), and estimated leaf nitrogen map (right). ....... 64
3.5 The six manageable areas as defined in this study. ............................................. 65
3.6 Average leaf nitrogen per sector during the crop growth. .................................... 66
4.1 The location of the study areas in Scipio, Utah, and Lodi, California. ................. 80
4.2 Details of AggieAir airframe . .............................................................................. 81
xiii
4.3 Relative quantum response of the VIS-NIR (left) and thermal camera (right). .... 82
4.4 An example of the summary report, the number of overlapped images range from 1 to 9 per frame. The black dots represent the center of the frame collected over the four flight lines . .............................................................................................. 85
4.5 An example of a ground control point (GCP), high vegetation, needed for calibration of thermal infrared (TIR) data. A picture of the GCP is captured right, a TIR imagery from a 3 m elevation in the middle, and the Blackbody reading on the left. ................................................................................................ 87
4.6 Landsat 8 and AggieAir relative spectral response. .............................................. 92
4.7 Landsat 8 point spread function (PSF) along the blue, green, and red (BGR) bands (left); Landsat 2D PSF representation on the right. .................................... 93
4.8 1:1 plot of the albedos from AggieAir and Landsat. ............................................ 96
4.9 Site S: Landsat reflectance in false color (left), Mapping Evapotranspiration with Internalized Calibration (METRIC) reference evapotranspiration fraction (EtrF) output (right). ...................................................................................................... 100
4.10 Site G: Landsat reflectance bands in true color (left), Mapping Evapotranspiration with Internalized Calibration (METRIC) reference evapotranspiration fraction (ETrf) output (right). ........................................................................................... 101
4.11 Site S: AggieAir reflectance in false color (left), Mapping Evapotranspiration with Internalized Calibration (METRIC) reference evapotranspiration fraction (ETrf) output (center), TIR (Celsius) map (right) ............................................... 102
4.12 Site G: AggieAir reflectance bands in true color (left), Mapping Evapotranspiration With Internalized Calibration (METRIC) reference evapotranspiration fraction (ETrf) output (center), TIR (celcuis) map (right). .. 103
4.13 Site S: Mapping Evapotranspiration With Internalized Calibration high resolution (METRIC-HR) reference evapotranspiration fraction (ETrf) on left, METRIC ETrf on right, both showing the 48 managable arcs as defined in this study. ........................................................................................................ 104
4.14 Northern center pivot reference evapotranspiration fraction (ETrf) estimates from Landsat and AggieAir inputs. .................................................................... 105
4.15 Southern center pivot reference evapotranspiration fraction (ETrf) estimates from Landsat and AggieAir inputs. .................................................................... 106
xiv
4.16 Site S: Comparison between reference evapotranspiration fraction (ETrf) estimates derived from Landsat and AggieAir data in the northern and southern center pivots. ....................................................................................................... 107
4.17 Mapping Evapotranspiration With Internalized Calibration (METRIC) reference evapotranspiration fraction (ETrf) on left, METRIC high resolution ETrf in middle, and 19 blocks of individual manageable areas on right. ........................ 109
4.18 The average reference evapotranspiration fraction (ETrf) estimates from Landsat and AggieAir inputs for each block. ..................................................... 110
4.19 Site G: Comparison between reference evapotranspiration fraction (ETrf) estimates derived from Landsat and AggieAir data. ........................................... 111
CHAPTER 1
INTRODUCTION
Unmanned air systems (UAS) have been around for a century, and for a long time
their sole application was in military practice. Then, UAS started branching into other
applications like archeological site assessment and mineral exploration. In 2004, only
approximately 2% of the UAS were operating merely in the civil market (Newcome,
2004). Since then, use has been increasing quickly. A report discussing the economic
impact of UAS published by the Association for Unmanned Vehicle Systems
International (2013) estimated an economic impact of $82 billion by 2025. Those
estimates are rationalized by a large, emerging application for UAS: precision
agriculture.
Precision agriculture (PA), the intelligent crop production system, is a scientific
and modern approach to agriculture production in the 21st century. Building on
traditional knowledge, this production farming system integrates new technologies. The
three key technological components are (a) a remote sensing platform like UAS that
collects data; (b) a geographic information system, where data analysis and visualization
are performed using various techniques and tools; and (c) modern precision farming tools
like the variable rate applicator that allows the implementation of site-specific
recommendation. Site-specific in PA is a term that refers to the treatment of the smallest
possible area as a single element. Precision agriculture is designed to increase long-term,
site-specific production efficiency, productivity, and profitability while avoiding the
2
undesirable effects of excess chemical loading to the environment or productivity loss
due to insufficient input application.
Precision agriculture requires remote sensing platforms with unique features that
are not found in conventional platforms like satellites and airborne platforms. These
features prevail in UAS and are the following: (a) cost-effectiveness, with a much lower
operational cost compared to satellite and manned aircraft; (b) high spatial resolution that
could reach centimeters; and (c) practicality, such that the UAS can be flown any time,
under reasonable weather condition. In precision agriculture applications, UAS collects
aerial imagery of agricultural fields to monitor the health of crops, estimate nutrient
status, quantify crop water demand, and estimate yield and many other practices. All of
these help the farmer to achieve the “three glory Ms”: maximize yield, minimize resource
utilization, and minimize adverse environmental footprint.
The capability to increase the adoption of precision agriculture among farmers
relies on five specific points, which need to be addressed by the research community:
1. Identify the management challenges that the producers encounter daily,
including being agronomical, environmental, or economical.
2. Adopt cost-efficient, high spatial resolution platforms that can collect data to
address the identified issues.
3. Enhance the data-processing procedures from automatic photogrammetric
software, to instrument calibration and atmospheric correction.
4. Build algorithms that can extract and interpret the information in the processed
data.
3
5. Communicate these findings by building decision-support tools that can
support a better informed management decision to best management practices
for operations.
This research is a step in this direction. This dissertation explores remotely sensed
data acquired from a UAS to serve precision agriculture application. The UAS, named
AggieAir, was developed in the Utah Water Research Laboratory. This battery-powered
aircraft has onboard a payload computer that controls three sensors acquiring imagery in
the visible, near-infrared and thermal infrared spectra. The three sensors are ideal because
of their small size, light weight, and low cost. The imagery produced from AggieAir is of
a spatial resolution of 5–60 cm. After processing these images, they are used to develop
models that estimate crop chlorophyll content, plant leaf nitrogen content, and crop water
use. Platform and sensor description, image processing, radiometric calibration, and
estimation models are described in this research. The information generated by this
research is directly beneficial to both small- and large-scale producers and indirectly to
the environment.
Objectives
The chapters of this research each address specific objectives of this study.
Chapter 2 was written to address the following objectives: (a) investigate the suitability
of high spatial resolution data acquired by AggieAir to estimate chlorophyll
concentration, (b) develop a model that can estimate chlorophyll concentration from
remotely sensed crop surface reflectance, and (c) determine which bands in the
reflectance spectrum are sensitive to estimation of leaf chlorophyll concentration. The
4
objectives of Chapter 3 were the following: (a) determine whether leaf nitrogen content
could be predicted from reflectance data collected by AggieAir, and (b) develop a model
that can estimate leaf nitrogen of oat crop with a 15 cm spatial resolution. The objectives
of Chapter 4 were the following: (a) explore the applicability of AggieAir 15 cm
resolution data collected in the visible, near-infrared, and thermal infrared spectra to
estimate crop evapotranspiration (ET); (b) propose modification on existing model inputs
to accommodate AggieAir data; (c) compare ET estimations derived from Landsat data to
those from AggieAir; and (d) develop a 15 cm ET estimate with recommendations on
farm manageable units.
Dissertation Organization
This dissertation’s objectives, outlined above, were met in a three-manuscript
format. In Chapter 2, the retrieval of cm-scale chlorophyll content from thermal and
multispectral optical imagery collected by a UAS is presented. A relevance vector
machine is trained on a dataset of in situ collected leaf chlorophyll measurements, and the
machine learning algorithm intelligently selects the most appropriate bands and indices
for building regressions with the highest prediction accuracy. Chapter 3 applies the same
methodology used in Chapter 2 to estimate leaf nitrogen content. The model recommends
a set of inputs that are needed to estimate the spatial distribution of nitrogen at a 15 cm
spatial resolution. Chapter 4 discusses the application of AggieAir data to map ET at a 15
cm spatial resolution. A detailed description of the pre- and post-processing procedure of
the AggieAir data is presented in this chapter. Chapter 4 also suggests adjustments done
on a surface energy balance algorithm with inputs to accommodate AggieAir imagery. A
5
comparison per the smallest manageable area is presented for better irrigation
management practices over the two study sites. Chapter 5 provides a summary of this
work, draws the major conclusions, and presents recommendations for further research.
The structure of this dissertation is based on the multiple-paper format. As a result, some
redundancies and repetition of parts of the material presented occur, especially the
description of the study area and the UAS platform. Additionally, each chapter is edited
in a stand-alone format, with acronym usage and references specific to each chapter.
Contributions
The findings of this dissertation contribute towards a greater understanding of
high spatial resolution data acquired by a UAS in precision agriculture application.
Specifically, this dissertation contributes the following:
The research demonstrates the capability of a UAS system named AggieAir to
successfully determine important agronomical parameters to be used in a
precision agriculture farming system (Chapter 2, 3, 4).
The research introduces Bayesian-based learning machine algorithms in precision
agriculture research (Chapter 2, 3).
The research identifies sensitive bands and indices for building regressions with
the highest prediction accuracy that can estimate plant chlorophyll content and
nitrogen leaf content (Chapter 2, 3).
Whereas other documents may contain some of these steps, this is the first
document to include a detailed description of the AggieAir platform (sensors,
6
thermal calibration, radiometric calibration, flight planning, and imagery
orthorectification) (Chapter 4).
The research demonstrates the potential of using data-mining learning machine
algorithms to build regressions that relate spectral information and ground
samples in a spatial context (Chapter 2, 3).
Findings demonstrate that crop water demand estimates using UAS high-
resolution data were comparable to the estimates generated from Landsat 8 data
(Chapter 4).
The research identifies the spectral difference of consumer-grade cameras and
Landsat 8 sensors and applies an adequate correction procedure (Chapter 4).
The paper presents a novel approach of comparing estimates that replaces the
traditional pixel-by-pixel comparison with a unit based on the irrigation
management system and design (Chapter 3, 4).
References
Association for Unmanned Vehicle Systems International. (2013). The economic impact of unmanned aircraft systems integration in the United States. Retrieved from http://www.auvsi.org/auvsiresources/economicreport
Newcome, L. R. (2004). Unmanned aviation: A brief history of unmanned aerial vehicles. Reston, VA: American Institute of Aeronautics and Astronautics.
7
CHAPTER 2
ESTIMATING CHLOROPHYLL FROM THERMAL AND BROADBAND
MULTISPECTRAL HIGH RESOLUTION IMAGERY
FROM AN UNMANNED AERIAL SYSTEM USING RELEVANCE
VECTOR MACHINES FOR PRECISION AGRICULTURE1
Abstract
Precision agriculture requires high-resolution information to enable greater
precision in the management of inputs to production. Actionable information about crop
and field status must be acquired at high spatial resolution and at a temporal frequency
appropriate for timely responses. In this study, high spatial resolution imagery was
obtained through the use of a small, unmanned aerial system called AggieAirTM.
Simultaneously with the AggieAir flights, intensive ground sampling for plant
chlorophyll was conducted at precisely determined locations. This study reports the
development of a relevance vector machine (RVM) coupled with cross validation and
backward elimination to a dataset composed of reflectance from high-resolution
multispectral imagery (visible/near-infrared, or VIS-NIR), thermal infrared (TIR)
imagery, and vegetative indices, in conjunction with in situ SPAD measurements from
1 Reprinted with some editorial corrections from “Estimating Chlorophyll With Thermal and Broadband Multispectral High Resolution Imagery From an Unmanned Aerial System Using Relevance Vector Machines for Precision Agriculture,” by M. Elarab, A. M. Ticlavilca, A. F. Torres-Rua, I. Maslova, and M. McKee, 2015, International Journal of Applied Earth Observation and Geoinformation, 43, 32-42. © 2015 with permission from Elsevier.
8
which chlorophyll concentrations were derived, to estimate chlorophyll concentration
from remotely sensed data at 15 cm resolution. The results indicate that an RVM with a
thin plate spline kernel type and kernel width of 5.4, having leaf area index (LAI),
normalized difference vegetation index (NDVI), thermal and red bands as the selected set
of inputs, can be used to spatially estimate chlorophyll concentration with a root mean-
squared error (RMSE) of 5.31 μg.cm-2, efficiency of 0.76, and 9 relevance vectors.
Keywords: Remote Sensing, High Spatial Resolution Imagery, Relevance Vector
Machine, Precision Agriculture, Chlorophyll Concentration
Introduction
Increasing world population levels will bring increased demand for food, water,
and agricultural inputs. Various agricultural farming strategies are being reevaluated to
determine how to improve food production, minimize environmental impact, and reduce
costs. Among many, precision agriculture has evolved as a viable system to improve
profitability and productivity (Daberkow, McBride, Robert, Rust, & Larson, 2000;
Lambert & Lowenberg-De Boer, 2000; Swinton & Lowenberg-DeBoer, 1998). Precision
agriculture is a process of finely adjusting agricultural inputs (e.g., water, nutrients) and
in-field practices (e.g., irrigation, fertilization), through the use of site-specific
information and spatial imagery, to improve measures of agricultural productivity (e.g.,
yield, net farm income; Pierce & Nowak, 1999).
Use of spatial imagery in agriculture has been the focus of many studies for the
past five decades (Bauer, 1985; Benedetti & Rossini, 1993; Franke & Menz, 2007; Idso,
Jackson, & Reginato, 1977; MacDonald & Hall, 1980; Mathur &Foody, 2008; Shanahan
9
et al., 2001; Stone et al., 1996), requiring increased investments in relevant research and
technologies (Schellberg, Hill, Gerhards, Rothmund, & Braun, 2008) that indicate remote
sensing can be a valuable tool to enhance precision agriculture (Haboudane, Miller,
Tremblay, Zarco-Tejada, & Dextraze, 2002; Lamb & Brown, 2001; Seelan, Laguette,
Casady, & Seielstad, 2003). However, remote sensing has yet to reach its full capability
in precision agriculture applications. Lack of fine spatial resolution and near real-time
data, compounded by high costs, has hindered remote sensing applications at the field
scale (Brisco, Brown, Hirose, McNairn, & Staenz, 1998; Kalluri, Gilruth, Bergman, &
Plante, 2002; Liaghat & Balasundram, 2010; M. S. Moran, Inoue, & Barnes, 1997).
Thirty years ago, Jackson (1984) envisioned an autonomous remote sensing platform that
could overcome most of the limitations; this is becoming a reality with the introduction of
affordable unmanned aerial systems (UAS). UAS, a potential substitute for satellite-based
remote sensing, are gaining attention and recognition in the scientific community as a
potential technology that can generate high spatial resolution imagery (< 1 m) and at a
temporal frequency appropriate for timely responses in the production of actionable
information about crop and field status.
One such UAS, named AggieAirTM, was developed by the Utah Water Research
Laboratory at Utah State University. AggieAir is designed to carry camera payloads to
acquire high-resolution, georeferenced aerial imagery to be used in various water, natural
resources, and agricultural applications, including precision agriculture. AggieAir holds
three sensors: Sensors 1 and 2 are consumer-grade cameras (personal point-and-click
cameras) that capture imagery, depending on flight elevation above ground, of 6–25 cm
10
resolution in the VIS (red, green, blue spectrum) and NIR spectrum, respectively. Sensor
3 is a micro bolometer thermal camera that captures images of 30–150 cm resolution in
the TIR spectrum. The three sensors are ideal because of their small size, light weight,
low cost, and high resolution. The use of high-resolution imagery (< 1 m) can potentially
improve the ability to evaluate the spatial dynamics of chlorophyll and detect its temporal
variation. In this study, the use of multispectral VIS-NIR-thermal high-resolution
imagery is investigated as a tool to estimate plant chlorophyll concentration to provide
time-critical information for precision agriculture.
Chlorophyll concentration, measured in mass per unit leaf area (μg cm-2), is an
important biophysical parameters retrievable from reflectance data. Chlorophyll is a vital
pigment primarily responsible for harvesting light energy used in photosynthesis (Evans,
1989; Niinemets & Tenhunen, 1997; Sims & Gamon, 2002) and is therefore an excellent
indicator of a crop’s overall physiological status (Evans, 1989; Yoder & Pettigrew-
Crosby, 1995), stress or disease (Chaerle & Van Der Straeten, 2000; Peñuelas & Filella,
1998; Zarco-Tejada Miller, Morales, Berjón, & Agüera, 2004), and yield predictions
(Dawson, North, Plummer, & Curran, 2003; Gitelson et al., 2006). Chlorophyll can
potentially provide an assessment of leaf nitrogen, an essential plant nutrient, due to the
close relationship between leaf chlorophyll and leaf nitrogen (Daughtry, Walthall, Kim,
De Colstoun, & McMurtrey, 2000; J. A. Moran, Mitchell, Goodmanson, & Stockburger,
2000; Wood, Tracy, Reeves, & Edmisten, 1992). Chlorophyll concentration varies with
vegetation growth, thus estimating chlorophyll across the field at different growth stages
could offer the farmer time- and location-specific critical information ideal for assisting
11
decision makers in monitoring their crops and managing farming activities to achieve
maximum production.
Several leaf scale studies have focused on estimating chlorophyll concentration
from VIS-NIR reflectance data. These studies indicated that the green and far-red regions
of the visible spectrum are sensitive to variations in chlorophyll concentrations (Datt,
1999; Demarez & Gastellu-Etchegorry, 2000; Gitelson & Merzlyak, 1994; Kim, 1994;
Zarco-Tejada Miller, Noland, Mohammed, & Sampson, 2001). Various successful
indices have been formulated to estimate chlorophyll concentration (Broge & Leblanc,
2001; Haboudane et al., 2002; le Maire, François, & Dufrêne, 2004). Some of these
indices are ratios of reflectance in individual narrow visible wavebands (Blackburn,
1998; Carter & Spiering, 2002) or ratios of reflectance in VIS and NIR (Gitelson,
Kaufman, & Merzlyak, 1996), whereas others are red-edge reflectance ratio indices
(Gitelson & Merzlyak, 1994; Kim, Daughtry, Chappelle, McMurtrey, & Walthall, 1994;
Zarco-Tejada & Miller, 1999) or first and second derivatives of reflectance spectra
(Miller, Hare, & Wu, 1990). Composites of indices have been developed (Haboudane et
al., 2002) in an attempt to correct for distortions in the reflectance data caused by soil
background effect and canopy architecture.
Detailed discussions and thorough reviews concerning appropriate optimal
wavelengths and various chlorophyll indices can be found in the literature (Broge &
Leblanc, 2001; Haboudane, Miller, Pattey, Zarco-Tejada, & Strachan, 2004). However,
most of the studies have had low spatial and coarse spectral resolution characteristics;
therefore, the applicability of those indices to high spatial resolution airborne data cannot
12
be evaluated. Regarding thermal imagery, it is mainly explored when information on
plant water status is in question, for example when screening drought-tolerance
genotypes (Blum, Mayer, & Gozlan, 1982), detecting crop water stress levels (Berni,
Zarco-Tejada, Suárez, & Ferereset, 2009), or estimating soil moisture and
evapotranspiration (Hassan Esfahani, Torres-Rua, Jensen, & McKee, 2014a, 2014b;
Jackson, Idso, Reginato, & Pinter, 1981; Wallace, Lucieer, Watson, & Turner, 2012).
However, TIR data have not been investigated in estimating chlorophyll yet. Exploring
thermal data in this study is rationalized by the close relationship between heat stress and
the photosynthetic capacity of the leaves (Raison, Roberts, & Berry, 1982; Sharkey,
2005) and consequently the chlorophyll concentration. The mechanism by which
moderate heat stress reduces photosynthetic capacity has been debated since the 1980s,
when researchers attributed the photosynthesis inhibition to different factors such as the
impairment of electron transport activity or the inactivation of Rubisco (Berry &
Bjorkman, 1980; Murakami, Tsuyama, Kobayashi, Kodama, & Iba, 2000; Salvucci &
Crafts‐ Brandner, 2004; Weis, 1981).
Estimating chlorophyll at a canopy level from optical remotely sensed data can
generally be carried out by several methodologies. The simplest methodology that is
widely accepted is the empirical method, such as those based on vegetation indices
(Johnson, Hlavka, & Peterson, 1994). Nevertheless, indices generated in this context are
inclined to unstable performance when applied to images that differ from the designed
method (Verrelst, Schaepman, Malenovský, & Clevers, 2010). Physical behavior based
methods are another approach to formulating estimates from remotely sensed data. This
13
method is based on physical laws that describe the transfer and interaction of radiation
within the atmospheric column and canopy, such as radiative transfer models (Myneni et
al., 1995). This approach has become more promising with advances in atmospheric
radiative transfer modeling. The biggest drawback for such a model is that it requires
site-specific information for proper model parameterization, which is not always
available. As a result, methods based on vegetation indices or physical models may be
either too simple or too complex to deliver accurate estimates (Baret & Buis, 2008).
Several books and published papers have reviewed these methodologies and highlighted
the advantages and disadvantages associated with the complexity of the modeling
approach selected, and the degree of general or local applicability of the methodology in
remote sensing (Baret & Buis, 2008; Zarco-Tejada et al., 2001).
Considerable research has been carried out to explore advanced computational
methods that are both accurate and robust. Machine learning regression algorithms
present a potential approach for generating adaptive; robust; and, once trained, fast
estimates (Hastie, Tibshirani, & Friedman, 2009; Knudby, LeDrew, & Brenning, 2010).
Recent studies have demonstrated successful performance of a very well-known machine
learning algorithm in estimating biophysical parameters using neural network models
(Cipollini, Corsini, Diani, & Grasso, 2001; De Martino et al., 2002; González Vilas,
Spyrakos, & Torres Palenzuela, 2011; Hassan Esfahani et al., 2014a; Verrelst, Alonso,
Camps-Valls, Delegido, & Moreno, 2012). In recent studies, neural networks are being
replaced by more advanced regression-based methods that are simpler to calibrate, like
support vector machines (SVMs; Camps-Valls, Bruzzone, Rojo-Álvarez, & Melganiet,
14
2006; Moser & Serpico, 2009; Pal & Mather, 2005) and RVMs (Camps-Valls, Gómez-
Chova et al., 2006). SVMs have been widely used in various remote sensing applications;
nevertheless, their large computational complexity is a major drawback. This complexity
of SVM models is due to their liberal use of basis functions that typically grow linearly
with the size of the training set (Tipping, 2001). Studies have shown that the behavior of
RVMs is often superior to that of SVMs (Demir & Erturk, 2007). The results given by
Tipping (2001) demonstrated that the RVM has a comparable generalization performance
to the SVM, while requiring dramatically fewer kernel functions or model terms. RVM,
in a statistical learning method proposed by Tipping, constitutes a Bayesian
approximation for solving nonlinear regression models and is often used for classification
and pattern recognition. RVMs offer excellent sparseness characteristics, are robust, and
can produce probabilistic outputs that permit the capture of uncertainty in the predictions
(Gómez-Chova, Muñoz-Marí, Laparra, Malo-López, & Camps-Valls, 2011;
Thayananthan, Navaratnam, Stenger, Torr, & Cipolla, 2008).
The main purposes of this study were to (a) introduce AggieAir as a successful
tool for use in precision agriculture; (b) explore the use of VIS, NIR, and TIR in
estimating chlorophyll concentration, and (c) use RVM algorithms to formulate spatially
distributed chlorophyll concentration estimates.
Materials and Methods
RVMs
This section presents a brief description of RVMs relevant to this study. Tipping
introduced the RVM in 2001. The RVM was developed with a Bayesian framework to
15
find sparse solutions in classification and regression studies based on acquiring relevance
vectors and weights by maximizing a marginal likelihood. In RVM regression models,
the weight of each input is governed by a set of hyperparameters that describe posterior
distribution of the weights and are estimated iteratively during the machine learning
training step (Tipping, 2001). This paper adopts the RVM introduced by Tipping (2004),
which resembles the 2001 model. The main feature in the 2004 model is that the inferred
predictors are even sparser, with relatively few relevance vectors. This model also offers
good generalization performance (Yuan, Wang, Yu, & Fang, 2007).
To build the model, input-output vector pairs {𝑥𝑖, 𝑦𝑖}𝑖=1𝑁 are sampled from a data
set of N input vectors {𝑿𝒏}𝒏=𝟏𝑵 with corresponding N output values {𝐲𝒏}𝒏=𝟏𝑵 . From these
vector paired data, we generate a training data subset from which the model learns the
dependence between inputs and the output target, with the purpose of making accurate
predictions of for previously unseen values of shown in Equation 1:
𝑦 = 𝑤𝜑(𝑥) + 𝜀 (1)
where is a vector of weight parameters, 𝜑(𝑥) = [1, 𝑓(𝑥, 𝑥1,… , 𝑓(𝑥, 𝑥𝑁)]^𝑇 is a
design matrix of 𝑁 + 1 vectors of kernel basis functions , is the error that for
algorithmic simplicity is assumed to be zero-mean Gaussian with variance .
The kernel or basis function is a method that detects embedded patterns in the
data by transforming or extending linear algorithms into nonlinear ones. Kernel methods
map the data into higher dimensional spaces to increase the computational power of the
machine (Cristianini & Shawe-Taylor, 2000; Genton, 2001; Souza, 2010; Vapnik, 2000).
Kernel functions could be linear, polynomial, and Gaussian kernel. However, choosing
y x
w
f ε
2
16
the most appropriate one highly depends on the nature of the relationship between the
inputs and outputs. Six kernel types, f, were considered: Gauss, Laplace, spline, Cauchy,
thin plate spline, and bubble (Bachour, Walker, Ticlavilca, McKee, & Maslova, 2014;
Ticlavilca, McKee, & Walker, 2013; Torres, Walker, & McKee, 2011). The process of
selecting the kernel type in this paper was conducted by trial and error.
The Gaussian likelihood of the data set can be written as in Equation 2:
2
22/2
2exp)2(),|(
wywyp NN
(2)
One of the classic approaches to estimating the parameters and in Equation
2 is using the method of maximum likelihood. However, with many parameters used as
training observations, the maximum likelihood estimation would lead to severe
overfitting (Tipping, 2004). To overcome this complexity, Tipping (2001) proposed
adding a “prior” to constrain the selection of parameters by defining an explicit zero-
mean Gaussian prior probability distribution over them as shown in Equation 3:
2exp)2()|(
2
1
2/12/ mmM
mm
M wwp
(3)
Where M is the number of independent hyperparameters TM ),...,( 1 . Each is
associated independently with every weight to moderate the strength of the prior and
provide the sparsity of the model (Tipping, 2001). How far each weight is allowed to
deviate from zero is controlled by the hyperparameter vectors (Yuan et al., 2007).
Consequently, using Bayes’s posterior inference, the posterior over W could be computed
as shown in Equation 4:
w 2
17
),|()|(),|(),,|( 2
22
ypwpwypywp (4)
Here, ),|( 2yp is the normalizing factor; ),|( 2wyp and )|( wp are both
Gaussian priors, so the posterior is also Gaussian with ),,|( 2ywp ~ ),( wN . The
posterior mean and covariance are defined as:
12 )( TA And (5)
yT 2 , (6)
where is .
A fast marginal likelihood optimization algorithm is used to obtain the optimal set
of hyperparameters, . This optimization algorithm uses an efficient sequential
addition and deletion of candidate basis functions described by Tipping and Faul (2003).
Given an unseen input vector, , the predictive distribution for the
corresponding target can be computed. This search for optimal hyperparameters is
learned using a type II maximum likelihood method coupled with iterative estimation
(Tipping, 2001) as shown in Equation 7:
)dw),(σp(w|y,α))p(y*|w,(σ)=),(σp(y*|y,α optoptoptoptopt 222 (7)
=> )σ*,N(y*|)),(σp(y*|y,α o p to p t 22 *)(
Where is the predictive mean of the output of the unseen data, and the posterior
mean weight of , TM ],...,[* **
1 ; and TM ])(,...,)[(*)( 2*2*
12 is the predictive
variance. This predictive variance is the sum of variances associated with both the noise
of the data and the uncertainty in the prediction of the weight parameters (Tipping, 2004).
A ), . . . ,( 1 Mdi ag
o p t
*x
*y
* *x
18
In this optimization process, the vectors from the training set associated with nonzero
weights are called the relevance vectors. The theory behind RVM, mathematical
formulation, likelihood maximization, and optimization procedure is discussed in detail
in Tipping (2004) and Tipping and Faul (2003).
Study Area
The field study was carried out in the summer of 2013 on privately owned
agricultural land in Scipio, Utah (39°14'N 112°6'W; see Figure 2.1). The plot, mainly
composed of loamy clay soil, was equipped with a center pivot sprinkler for irrigating
and fertigation oats (Avena sativa). The study area was restricted to the northwest quarter
of the center pivot so that samples could be collected within a close time frame relative to
the AggieAir flight. AggieAir aircraft were flown four times over the area, covering the
entire growth cycle of oats. The flights on 05/16, 06/01, 06/09, and 06/17 reflected the
four stages of growth: 10 days after planting, early growth, mid growth, and early
flowering. Oats were harvested after the fourth flight to be used as forage.
Figure 2.1. The location of the study area in Utah (left), The quarter used in plant chlorophyll estimation model (right).
19
Instrumentation: Remote Sensing Platform AggieAir
AggieAir is a UAS designed to carry camera payloads to acquire aerial imagery
for precision agriculture and other types of applications (Figure 2.2). The UAS aircraft is
battery powered and equipped with a payload system (which includes three cameras and
a computer), avionics, two inertial sensors (a GPS module and an inertial measurement
unit), radio controller, and flight control. The aircraft is propelled using an electric,
brushless motor. It does not require a runway and can be flown autonomously or
manually. In autonomous mode, the aircraft follows a preprogrammed flight plan
containing navigation waypoints defined by GPS and altitude. While operational, the
payload computer instructs the three cameras to acquire imagery in the VIS, NIR, and
thermal spectra and records the position and orientation of the aircraft when each image
is taken. Table 2.1 illustrates the UAS specification in more detail.
Figure 2.2. AggieAir airframe layout.
20
Table 2.1
AggieAir Unmanned Aerial System Specifications
Specification Range
Flight duration 45–60 min
Flight altitudes 200–1000 m
Maximum takeoff weight 6.35 kg
Visible/near-infrared resolution 6–25 cm
Thermal resolution 30–150 cm
Wing span 2.5 m
The VIS camera used in AggieAir is a Canon S-95, with a 10-megapixel charge-
coupled device (CCD) sensor and an ISO range of 80 to 3200. The radiometric resolution
of the Canon S-95 is 8-bit color, which means that the digital measurement for a
particular pixel in a given spectral band ranges from 0 to 255. The NIR camera is an
identical Canon S-95, modified by replacing the manufacturer’s optical filter with a
Wratten 87 NIR filter that allows NIR wavelengths of 750 nm. The relative spectral
responses of the VIS-NIR cameras were not provided by the manufacturers but were
obtained using the algorithm provided by Jiang, Liu, Gu, and Susstrunk (2013). The
camera VIS-NIR spectral response is shown in the left portion of Figure 2.3. AggieAir
also carries a small, low-power, microbolometer thermal camera from Infrared Cameras,
Inc. (2014). The relative spectral response of the thermal camera is shown in the right
portion of Figure 2.3.
21
Figure 2.3. Relative spectral response of the VIS-NIR (left) and thermal camera (right). VIS-NIR = visible/near-infrared; ICI = Infrared Cameras, Inc.
Following VIS and NIR image acquisition, a two-step processing phase occurs.
The first step is image mosaicking and orthorectification. This technique, achieved with
EnsoMOSAIC software (MosaicMill, 2009), combines all of the images into one large
mosaic and rectifies it into a ground coordinate system. The software generates hundreds
of tie-points between overlapping images by using photogrammetric principles in
conjunction with image GPS log file data and exterior orientation information from the
on-board cameras to refine the estimate of the position and orientation of individual
images. The resulting image is an orthorectified digital number mosaic. The second step
involves radiometric calibration: the conversion of the digital pixels into a measure of
reflectance. This conversion is based on methods adapted from the research (Crowther,
1992; Miura & Huete, 2009; Neale & Crowther, 1994). The major steps involved in this
0
0.2
0.4
0.6
0.8
1
380 480 580 680 780 880
Wavebands (nm)
VIS-NIR Relative Spectral Response
0
0.2
0.4
0.6
0.8
1
1200 6200 11200 16200 21200
Wavebands (nm)
ICI Thermal Camera Relative Spectral Response
22
methodology are the reference panel calibration and solar zenith angle calculations. This
method converts raw airborne multispectral data by calculating the ratio of linearly
interpolated reference values from the pre- and post-flight reference panel readings, as
discussed in detail by Zaman, Jensen, Clemens, and McKee (2014) and by Clemens
(2012). After completing the two-step process, images are geometrically rectified and
radiometricaly corrected to obtain a four-layer (red, green, blue, NIR) canopy surface
reflectance in a single image (Figure 2.4).
Figure 2.4. Raw natural color images from the unmanned aerial system (left), accurate orthorectified mosaic image from EnsoMOSAIC (center), and radiometric calibration of visible spectrum image (right).
Thermal imagery processing also requires an initial step of mosaicking and
orthorectification similar to the VIS and NIR images. However, the resulting thermal
mosaic is composed of brightness temperature in degrees Celsius (± 0.1 degrees) instead
of digital numbers. Compensating for external disturbance and geometric calibration are
also unique challenges associated with the thermal camera. Jensen (2014) thoroughly
explained the methodology of processing thermal maps adopted by the authors.
23
Data Collection
The collection of the ground and remotely sensed data occurred under similar
weather conditions in a 1- to 2-hour window.
Multispectral image acquisition. Four multispectral mosaics were acquired by
AggieAir during summer 2013. Acquisition dates were planned to coincide with different
development stages and with overflights of Landsat. Images were collected, following the
Landsat image acquisition protocol, close to solar noon (between 12 a.m. and 1 p.m.).
The flight time (beginning to end) ranged from 30–40 min. All four missions were
successfully performed, providing image data covering the earliest, middle, and latest
periods of the oat growth. The spatial resolution is 0.15 m for VIS and NIR images and
0.6m for the TIR images.
Ground data acquisition. Intensive ground truth sampling of plant chlorophyll
was conducted simultaneously with the AggieAir flights at precise GPS locations. The
GPS data were collected using an rtkGPS with < 1 mm precision in a 1 Hz bandwidth
(Trimble® R8, Global Navigation Satellite System, Dayton, Ohio). A SPAD-502
chlorophyll meter (Minolta Corporation, New Jersey) was used for in vivo measurement
of the ratio of light transmittance through the leaf at wavelengths of 650 and 940 nm.
Instrument readings have been shown to correlate well to laboratory measurements of
chlorophyll concentration in several species (Yadawa, 1986). On each sampling
campaign, 40 SPAD measurements were collected on average. The chlorophyll meter
readings were taken midway on the fully expanded top-of-canopy leaves. Each
measurement was characterized by the mean of six replicate measurements. The
24
chlorophyll meter measures an area of 2 x 3 mm with an accuracy of ± 1.0 SPAD unit (at
room temperature). However, the SPAD-502 meter displays the chlorophyll readings in
arbitrary units (SPAD unit) rather than the actual amounts of chlorophyll in mass per leaf
area; thus, further conversions were needed. The SPAD units were transformed to a
chlorophyll concentration index (CCI) unit using Equation 8 and then to chlorophyll in
mass per leaf area using Equation 9 (Parry, Blonquist, & Bugbee, 2014). Equation 9 was
developed for barley crops; however, literature has shown that monocots have a similar
optical/absolute chlorophyll concentration relationship.
67.2*00119.01 SPADCCI (8)
(9)
Linking on-ground measurements to airborne imagery. Ground coordinates of
sampled chlorophyll coincided precisely with the location of the plants in the geo-
rectified imagery. Ground coordinates of the samples were overlaid onto the geo-rectified
imagery, and, using the ArcGIS spatial analyst tool (Extract Multi Values to Points), an
automated process was developed to extract the pixel value representing the center of
each sampled area.
Model Potential Inputs and Performance
Three of the four flights (early growth, mid growth, and early flowering),
excluding the flight 10 days after planting, were used in the dataset to train and test the
model. The dataset contains coincident in situ SPAD measurements used to derive
chlorophyll concentration and remote sensing reflectance measurements. All the data
were collected from inside the center pivot quarter (within field data); the zeros found in
)(146132)µmol.m( 43.0-2 CCIlChlorophyl
25
the data set represent the areas of no vegetation (center pivot wheels trajectory). A
statistical description of the dataset is presented in Table 2.2. Each pair of data consists of
a target, which is the chlorophyll concentration, and a set of 8 potential inputs tabulated
in Table 2.2. The potential inputs are composed of data retrieved form the UAS imagery
(VIS, NIR, TIR); vegetative indices (green model and NDVI) that were reported to be
sensitive in estimating chlorophyll (Gitelson, Vina, Ciganda, Rundquist, & Arkebauer,
2005; Shanahan et al., 2003); and LAI, a well-known and widely used vegetation index
related to crop growth. Table 2.3 shows the indices formulations.
Table 2.2
Statistical Description of the Dataset Used for Plant Chlorophyll Estimation
Input Range M ± SD Range M ± SD
Potential inputs
AggieAir inputs
Blue 0.11–0.36 0.15 ± 0.04 NIR 0.51–0.61 0.57 ± 0.02
Green 0.20–0.49 0.26 ± 0.05 Thermal (oC) 23.11–36.16 29.88 ± 4.13
Red 0.15–0.51 0.22 ± 0.07
Indices inputs
NDVI 0.00–0.78 0.44 ± 0.12 LAI (m2/m2) 0.00–4.23 2.41 ± 0.90
GM 0.17–1.74 1.17 ± 0.34
Target/output
Chlorophyll (μg.cm-2 ) 0.00–61.64 47.01 ± 11.26
Note. GM = green meter; LAI = leaf area index; NDVI = normalized difference vegetation index; NIR = near-infrared.
These potential predictors, exert to a certain degree correlation between each
other. This is because they are derived from the same AggieAir reflectance bands
26
(statistical correlation). While in customary statistics (e.g., linear regression) using these
predictors would raise issues, the Bayesian regression machine applied in this study can
deal with this problem. The kernel or basis function projects these potential inputs into a
higher dimensional space. The way these inputs are projected in the new dimensional
space, as well as the sparse representation of the observations in the final model, help the
model deal with collinearity issues.
Table 2.3
Vegetative Indices Formulation Used for Plant Chlorophyll Estimation
Indices Formula Reference
Green model RNIR / RGreen – 1 Gitelson et al., 2005
NDVI (RNIR – Rred) / (RNIR + RRed) Rouse, Haas, Schell, Deering, & Harlan, 1974
LAIa ln [(NDVI – NDVImax) / (NDVImin – NDVImax)] / –0.54
Duchemin et al., 2006; Smith, Bourgeois, Teillet, Freemantle, & Nadeau, 2008
Note. LAI = leaf area index; NDVI = normalized difference vegetation index. aLAI was calculated empirically and not validated by field measurements.
In preliminary runs different potential inputs were explored. For example, one set
composed of only single bands, and another set composed of the ratio of the single bands.
In addition, the authors tried vegetative indices sensitive to chlorophyll estimations
(transformed chlorophyll absorption reflectance index, modified chlorophyll absorption
ratio index, and modified triangular vegetation index) that were modified to adapt to the
spectral response of AggieAir sensors (e.g., replacing the required red edge by the NIR
band). However, details on these preliminary runs are not reported in this study because
of their low statistical performance.
27
The RVM is a well-established statistical learning algorithm that adopts a full
probabilistic framework. Its key feature is that it can yield a solution function that
depends on only a very small number of training samples (called relevance vectors). In
the RVM framework, the model is built on the few training examples whose associated
hyperparameters do not go to infinity during the training process, leading to a sparse
solution. The implemented RVM is based on the MATLAB code provided via Michael E.
Tipping’s website. The RVM model in this research was first trained and tested using K-
fold cross validation (K = 5); the cross validation technique is utilized to generalize an
independent training data set (Kohavi, 1995). In this procedure, the training set is
partitioned into K disjoint sets. The model is trained, for a chosen kernel, on all the
subsets except for one, which is left for testing. The procedure is repeated for a total of K
trials, each time using a different subset for testing. After the selection of the kernel
function and its width, the whole data set is trained using RVM based regression. The
advantage of this method over a random selection of training samples is that all
observations are used for either training (K times) or evaluation (once).
The model was developed with an input selection process (Guyon & Elisseeff,
2003) in an attempt to explain the data in the simplest way possible. Potential inputs were
examined to see which were most relevant to the target function and thus avoid degrading
the performance of a learning algorithm due to the presence of irrelevant input variables.
In each iteration, the input with the minimum efficiency was eliminated.
The RVM model was tested using six kernel types: Gauss, Laplace, spline,
Cauchy, thin plate spline, and bubble. The performance of the model was evaluated by
28
comparing the RMSE and the Nash-Sutcliffe efficiency (E); these two parameters have
been widely used to evaluate the performance of RVM models. The larger the value of E
and the smaller the value of RMSE, the greater the precision and accuracy of the model
to predict chlorophyll. The RMSE and E are computed as shown in Equations 10 and 11,
respectively:
N
yyRMSE
N
ttt
1
2)ˆ( (10)
N
t t
N
t t
yy
yyE
12
12
)(
)ˆ(1
(11)
where ŷt = predicted chlorophyll concentration; yt = measured chlorophyll concentration;
= mean of the observed chlorophyll concentration; = mean of the estimated
chlorophyll concentration; and N = total number of observations.
Result and Discussion
Each of the six kernel types was tested over a wide range of kernel widths (10-5-
105), and RMSE and E were calculated for all of the resulting models to assess their
predictive capabilities. An embedded loop in the coding model was developed to
represent the backward elimination tool. For each type of kernel and its corresponding
width, the RVM was first run using all of the 8 inputs, consequently generating all of the
needed statistical model performance estimates to assess the model. A set of defined
iterations then eliminated, in order, the input with the minimum efficiency, thus removing
y y
29
the input least relevant to the target function. After numerous computational runs, four
options presented themselves as potential “best model” scenarios (Table 2.4). All four of
these potential best model scenarios had an RMSE < 6 μg.cm-2 and an E > 0.7. In 94% of
all runs conducted across the six kernel types, the thermal band was the last input to be
dropped, suggesting that thermal imagery is an important input, at least in the case of
study area, possessing the most relevant information for estimating chlorophyll
concentration. Thermal data allowed the models to differentiate between the bare soil and
the different level of vegetation per pixel resulting in a more accurate chlorophyll
estimates. A preliminary interpretation could be the fact that oat leaves are very thin, with
minimal heat capacity, and as a result, leaves exposed to full sunlight can warm up
substantially above air temperature. This elevated temperature can help identify greener
leaves and as a result those with higher chlorophyll values. Nevertheless, additional
experiments that explore thermal imagery and its effect on chlorophyll estimations need
to be conducted.
Table 2.4
Potential “Best Scenarios”
Model Kernel type # of inputs Inputs
1 Gaussian 4 LAI, NDVI, thermal, green
2 Gaussian 3 Thermal, green, LAI
3 Laplace 4 NDVI, red, green, thermal
4 Thin plate spline 4 LAI, NDVI, thermal, red Note. LAI = leaf area index; NDVI = normalized difference vegetation index.
30
When plotting the 1-1 plot for the four best scenario candidates, the plots looked
almost identical. Since the statistical performance does not reveal an absolute best model,
visual comparison was made of the chlorophyll estimation maps of the four models, on
one hand, and the NDVI, LAI, and true-color maps, on the other hand. The chlorophyll
estimates for the early growth, mid growth, and early flowering images were developed
considering the unique characteristic of each of the four best models (kernel type, width,
and set of inputs). Models 1 and 2 showed clear overfitting when plotted over the entire
map: In each case, the resulting map was one solid color, with no variation in estimated
chlorophyll between bare soil and fully grown oat plants. Model 3 showed more variation
within the field; nevertheless, visual comparisons with Model 4 indicated that Model 4
was superior. Model 4 showed an RSME of 5.31 μg.cm-2, an E of 0.76, and 9 relevance
vectors. Figure 2.5 illustrates the measured chlorophyll concentration versus estimated
values with a one standard error confidence interval. The three flights are separated by
the yellow line in the graph. Some differences can be observed between the different
dates; the estimates for the first and third flights are more precise than the second flight.
This could be due to the stage of the crop growth or the homogeneity of the vegetation
cover.
Figure 2.6 represents regression diagnostic plots of Model 4 that address model
assumptions like linearity and equality of variances. The 1:1 plot confirms the adequacy
of the model proposed for most of the chlorophyll values lying within the boundaries of
± 1.0 SPAD unit (sensor accuracy), which corresponds to 14 μg.cm-2. The chlorophyll
maps generated from Model 4, along with NDVI and LAI, are presented in Figure 2.7.
31
Figure 2.5. Measured versus predicted chlorophyll concentration in the three flights for Model 4. Vertical yellow lines separate the three flight dates.
Figure 2.6. Model 4: Residual plot of the three flights (left) and one-by-one plot excluding the bare soil–zero chlorophyll points and reflecting the chlorophyll meter accuracy (right).
32
As shown in Figure 2.7, the predicted chlorophyll concentration maps show a
visual good agreement with the LAI and NDVI maps. In the early growth image, the field
exterior had weeds growing in it, which explains the predicted chlorophyll concentration
values. This area was not irrigated during the growing cycle, leaving the weeds to dry and
senescence, thus a near-zero chlorophyll concentration value was assigned by the model
in the following two images. Also, the wheel tracks and the access road that are located
around the center pivot had no vegetation cover, and the model successfully assigned a
near-zero chlorophyll concentration to these features.
Figure 2.7. True-color maps, normalized difference vegetation index (NDVI) maps, leaf area index (LAI) maps, and the estimated chlorophyll concentration (μg.cm-2) map for the three different dates representing early growth, mid growth, and early flowering.
Early growth
Mid growth
Early flowering
33
Another common pattern was the two thick horizontal and vertical lines that
protrude in the images. These were past ditch lines that had been used in flood irrigation
activities prior to the conversion of the field to a center pivot system. The greater water
content in those areas caused the plants growing along those two lines to be very
vigorous. This is reflected in the high chlorophyll concentration values given to the plants
in this area.
Chlorophyll concentration varies widely within the growing season, and therefore
any recommended analytical technique must perform well under unseen data. To explore
the model with unseen data, the May 16 flight was used. Now that the model was
established with a defined set of features (inputs, kernel type, kernel width), data from the
May 16 flight (10 days after planting of the oats) were entered in the model to explore the
model’s performance when subjected to totally unseen data. The predicted chlorophyll
concentration map is shown in Figure 2.8.
Figure 2.8. True-color image of flight zero (left), leaf area index (LAI) map (middle), and estimated chlorophyll concentration map (μg.cm-2, right).
34
Again, the predicted chlorophyll concentration map for the fourth flight showed
good association with the NDVI map. Areas of vigorous growth, bare soil, and low
vegetation were similar in the three maps and represented similar growth patterns. This
test reported an RSME of 8.52 μg.cm-2 and E of 0.71 for this flight. This result showed
that the model successfully performed when given unseen data.
Despite the complexity of the statistical model included in this paper, it is
anticipated that the lucid output (chlorophyll concentration maps) will help agricultural
decision makers quantify field chlorophyll and address its variability and as a result
improve input efficiency, environmental sustainability, and yield. Adoption of precision
agriculture is likely to continue into the foreseeable future. However, studies that explore
high-resolution sensors (< 1 m) with adequate frequent coverage, combined with
techniques capable of extracting information from imagery to provide near real-time
information, will be a determining factor in the adoption rate of precision agriculture.
Conclusion
This paper presented the application of imagery from AggieAir, a remote sensing
platform, combined with machine learning algorithms (RVM) to estimate chlorophyll
concentration as an important biophysical parameter to be used in precision agriculture.
The RVM modeling technique, coupled with cross validation and backward elimination,
was applied to a data set composed of reflectance from high-resolution multispectral
imagery (VIS-NIR), TIR imagery, and vegetative indices, in conjunction with in situ
chlorophyll concentrations derived from SPAD measurements. Six kernel types were
tested over a wide range of kernel widths. Model performance was evaluated by
35
comparing the RMSE and E of various models and later by visual comparison.
Chlorophyll concentration estimation was best achieved with Model 4 (kernel type: thin
plate spline; kernel width: 5.4; selected inputs: LAI, NDVI, thermal, and red band;
RSME: 5.31 μg.cm-2; E: 0.76; and 9 relevance vectors) for the three flights. Of all the
inputs, thermal band was retained last in 94% of the models, proving the significance of
thermal imagery as an input possessing the most relevant information in estimating
chlorophyll concentration.
Converting these chlorophyll estimate maps into actionable information to benefit
the end user now shows promise. Other research that estimates soil moisture, actual
evapotranspiration, and soil nutrient content using the same high-resolution aerial
platforms allows for wider adoption of precision agriculture by future farmers. Although
the results presented in this section are arguably not yet actionable, maps like these could
be used to quantify plant health, predict yield, and indicate where and how much fertilizer
to apply.
AggieAir imagery, combined with appropriate analytic tools, allows spatial
estimation of chlorophyll content. These estimates, made at such fine resolutions in space
and time, can aid farmers in assessing the heterogeneity of their fields and subsequently
implement needed actions accordingly. The high-resolution spatial information generated
from AggieAir imagery could enable far greater precision in the application of nitrogen
fertilizers and water through a center pivot irrigation system.
36
Acknowledgment
This research was supported by the Utah Water Research Laboratory, the Provo,
Utah Office of the U.S. Bureau of Reclamation, and the AggieAir Flying Circus at the
Utah Water Research Laboratory. The authors appreciate the support the Utah State
University AggieAir team that helped in the data collection procedure. Special thanks
goes to the Division of Research Computing in the Office of Research at Utah State
University. Appreciation goes to Mrs. Carri Richards for her timely help in editing the
paper.
References
Bachour, R., Walker, W., Ticlavilca, A., McKee, M., & Maslova, I. (2014). Estimation of spatially distributed evapotranspiration using remote sensing and a relevance vector machine. Journal of Irrigation and Drainage Engineering, 140(8), 04014029. doi:10.1061/(ASCE)IR.1943-4774.0000754
Baret, F., & Buis, S. (2008). Estimating canopy characteristics from remote sensing observations: Review of methods and associated problems. In S. Liang (Ed.), Advances in land remote sensing (pp. 173-201). Dordrecht, The Netherlands: Springer.
Bauer, M. E. (1985). Spectral inputs to crop identification and condition assessment. Proceedings of the IEEE, 73, 1071-1085.
Benedetti, R., & Rossini, P. (1993). On the use of NDVI profiles as a tool for agricultural statistics: The case study of wheat yield estimate and forecast in Emilia Romagna. Remote Sensing of Environment, 45, 311-326.
Berni, J., Zarco-Tejada, P. J., Suárez, L., & Fereres, E. (2009). Thermal and narrowband multispectral remote sensing for vegetation monitoring from an unmanned aerial vehicle. IEEE Transactions on Geoscience and Remote Sensing, 47, 722-738.
Berry, J., & Bjorkman, O. (1980). Photosynthetic response and adaptation to temperature in higher plants. Annual Review of Plant Physiology, 31(1), 491-543.
37
Blackburn, G. A. (1998). Quantifying chlorophylls and caroteniods at leaf and canopy scales: An evaluation of some hyperspectral approaches. Remote Sensing of Environment, 66, 273-285.
Blum, A., Mayer, J., & Gozlan, G. (1982). Infrared thermal sensing of plant canopies as a screening technique for dehydration avoidance in wheat. Field Crops Research, 5, 137-146.
Brisco, B., Brown, R. J., Hirose, T., McNairn, H., & Staenz, K. (1998). Precision agriculture and the role of remote sensing: A review. Canadian Journal of Remote Sensing, 24, 315-327.
Broge, N. H., & Leblanc, E. (2001). Comparing prediction power and stability of broadband and hyperspectral vegetation indices for estimation of green leaf area index and canopy chlorophyll density. Remote Sensing of Environment, 76, 156-172.
Camps-Valls, G., Bruzzone, L., Rojo-Álvarez, J. L., & Melgani, F. (2006). Robust support vector regression for biophysical variable estimation from remotely sensed images. Geoscience and Remote Sensing Letters, IEEE, 3(3), 339-343.
Camps-Valls, G., Gómez-Chova, L., Muñoz-Marí, J., Vila-Francés, J., Amorós-López, J., & Calpe-Maravilla, J. (2006). Retrieval of oceanic chlorophyll concentration with relevance vector machines. Remote Sensing of Environment, 105(1), 23-33.
Carter, G. A., & Spiering, B. A. (2002). Optical properties of intact leaves for estimating chlorophyll concentration. Journal of Environmental Quality, 31, 1424-1432.
Chaerle, L., & Van Der Straeten, D. (2000). Imaging techniques and the early detection of plant stress. Trends in Plant Science, 5(11), 495-501.
Cipollini, P., Corsini, G., Diani, M., & Grasso, R. (2001). Retrieval of sea water optically active parameters from hyperspectral data by means of generalized radial basis function neural networks. IEEE Transactions on Geoscience and Remote Sensing, 39, 1508-1524.
Clemens, S. R. (2012). Procedures for correcting digital camera imagery acquired by the AggieAir remote sensing platform. Thesis, Utah State University, Logan.
Cristianini, N., & Shawe-Taylor, J. (2000). An introduction to support vector machines and other kernel-based learning methods. New York, NY: Cambridge University Press.
Crowther, B. G. (1992). Radiometric calibration of multispectral video imagery. Unpublished doctoral dissertation, Utah State University, Logan.
38
Daberkow, S. G., McBride, W. D., Robert, P. C., Rust, R. H., & Larson, W. E. (2000). Adoption of precision agriculture technologies by U.S. farmers. In P. C. Robert, R. H. Rust, & W. E. Larson (Eds.), Proceedings of the 5th International Conference on Precision Agriculture, Bloomington, Minnesota, 16-19 July, 2000 (pp. 1-12). Bloomington, MN: American Society of Agronomy.
Datt, B. (1999). A new reflectance index for remote sensing of chlorophyll content in higher plants: Tests using eucalyptus leaves. Journal of Plant Physiology, 154(1), 30-36.
Daughtry, C. S. T., Walthall, C. L., Kim, M. S., De Colstoun, E. B., & McMurtrey, J. E., III. (2000). Estimating corn leaf chlorophyll concentration from leaf and canopy reflectance. Remote Sensing of Environment, 74, 229-239.
Dawson, T. P., North, P. R. J., Plummer, S. E., & Curran, P. J. (2003). Forest ecosystem chlorophyll content: Implications for remotely sensed estimates of net primary productivity. International Journal of Remote Sensing, 24, 611-617.
De Martino, M., Mantero, P., Serpico, S. B., Carta, E., Corsini, G., & Grasso, R. (2002). Water quality estimation by neural networks based on remotely sensed data analysis. In Proceedings of the International Workshop on Geo-Spatial Knowledge Processing for Natural Resource Management (pp. 54-58).
Demarez, V., & Gastellu-Etchegorry, J. P. (2000). A modeling approach for studying forest chlorophyll content. Remote Sensing of Environment, 71, 226-238.
Demir, B., & Erturk, S. (2007). Hyperspectral image classification using relevance vector machines. Geoscience and Remote Sensing Letters, IEEE, 4(4), 586-590.
Duchemin, B., Hadria, R., Erraki, S., Boulet, G., Maisongrande, P., Chehbouni, A., . . . Simonneaux, V. (2006). Monitoring wheat phenology and irrigation in Central Morocco: On the use of relationships between evapotranspiration, crops coefficients, leaf area index and remotely-sensed vegetation indices. Agricultural Water Management, 79(1), 1-27.
Evans, J. R. (1989). Photosynthesis and nitrogen relationships in leaves of C3 plants. Oecologia, 78(1), 9-19.
Franke, J., & Menz, G. (2007). Multi-temporal wheat disease detection by multi-spectral remote sensing. Precision Agriculture, 8(3), 161-172.
Genton, M. G. (2002). Classes of kernels for machine learning: A statistics perspective. The Journal of Machine Learning Research, 2, 299-312.
39
Gitelson, A. A., Kaufman, Y. J., & Merzlyak, M. N. (1996). Use of a green channel in remote sensing of global vegetation from EOS-MODIS. Remote Sensing of Environment, 58, 289-298.
Gitelson, A., & Merzlyak, M. N. (1994). Quantitative estimation of chlorophyll-a using reflectance spectra: Experiments with autumn chestnut and maple leaves. Journal of Photochemistry and Photobiology B: Biology, 22, 247-252.
Gitelson, A. A., Viña, A., Ciganda, V., Rundquist, D. C., & Arkebauer, T. J. (2005). Remote estimation of canopy chlorophyll content in crops. Geophysical Research Letters, 32(8). doi:10.1029/2005GL022688
Gitelson, A. A., Viña, A., Verma, S. B., Rundquist, D. C., Arkebauer, T. J., Keydan, G., . . . Suyker, A. E. (2006). Relationship between gross primary production and chlorophyll content in crops: Implications for the synoptic monitoring of vegetation productivity. Journal of Geophysical Research: Atmospheres (1984–2012), 111(D8).
Gómez-Chova, L., Muñoz-Marí, J., Laparra, V., Malo-López, J., & Camps-Valls, G. (2011). A review of kernel methods in remote sensing data analysis. In S. Prasad, L. M. Bruce, & J. Chanussot (Eds.), Optical remote sensing (pp. 171-206). Berlin, Germany: Springer-Verlag Berlin Heidelberg.
González Vilas, L., Spyrakos, E., & Torres Palenzuela, J. M. (2011). Neural network estimation of chlorophyll a from MERIS full resolution data for the coastal waters of Galician rias (NW Spain). Remote Sensing of Environment, 115, 524-535.
Guyon, I., & Elisseeff, A. (2003). An introduction to variable and feature selection. The Journal of Machine Learning Research, 3, 1157-1182.
Haboudane, D., Miller, J. R., Pattey, E., Zarco-Tejada, P. J., & Strachan, I. B. (2004). Hyperspectral vegetation indices and novel algorithms for predicting green LAI of crop canopies: Modeling and validation in the context of precision agriculture. Remote Sensing of Environment, 90, 337-352.
Haboudane, D., Miller, J. R., Tremblay, N., Zarco-Tejada, P. J., & Dextraze, L. (2002). Integrated narrow-band vegetation indices for prediction of crop chlorophyll content for application to precision agriculture. Remote Sensing of Environment, 81, 416-426.
Hassan Esfahani, L., Torres-Rua, A., Jensen A., & McKee, M. (2014a). Assessment of surface soil moisture using high-resolution multi-spectral imagery and artificial neural networks. Remote Sensing, 7, 2627-2646.
Hassan Esfahani, L., Torres-Rua, A., Jensen A., & McKee, M. (2014b). Top soil moisture estimation for precision agriculture using unmmaned aerial vehicle multispectral
40
imagery. In Geoscience and Remote Sensing Symposium (IGARSS), 2014 IEEE International (pp. 3263-3266). doi:10.1109/IGARSS.2014.6947175
Hastie, T., Tibshirani, R., & Friedman, J. (Eds.). (2009). The elements of statistical learning: Data mining, inference, and prediction (2nd ed.). New York, NY: Springer.
Idso, S. B., Jackson, R. D., & Reginato, R. J. (1977). Remote-sensing of crop yields. Science, 196(4285), 19-25.
Infrared Cameras, Incorporated. (2014). Home page. Retrieved from http://www .infraredcamerasinc.com
Jackson, R. D. (1984). Remote sensing of vegetation characteristics for farm management. In 1984 Technical Symposium East (pp. 81-97). Bellingham, WA: International Society for Optics and Photonics.
Jackson, R. D., Idso, S. B., Reginato, R. J., & Pinter, P. J. (1981). Canopy temperature as a crop water stress indicator. Water Resources Research, 17, 1133-1138
Jensen, A. M. (2014). Innovative payloads for small unmanned aerial system-based personal remote sensing and applications. Doctoral dissertation, Utah State University, Logan.
Jiang, J., Liu, D., Gu, J., & Susstrunk, S. (2013). What is the space of spectral sensitivity functions for digital color cameras? In 2013 IEEE Workshop on Applications of Computer Vision (WACV) (pp. 168-179). doi:10.1109/WACV.2013.6475015
Johnson, L. F., Hlavka, C. A., & Peterson, D. L. (1994). Multivariate analysis of AVIRIS data for canopy biochemical estimation along the Oregon transect. Remote Sensing of Environment, 47, 216-230.
Kalluri, S., Gilruth, P., Bergman, R., & Plante, R. (2002). Impacts of NASA’s remote sensing data on policy and decision making at state and local agencies in the United States. In Geoscience and Remote Sensing Symposium, 2002. IGARSS'02. 2002 IEEE International (Vol. 3, pp. 1691-1693). doi:10.1109/IGARSS.2002 .1026223
Kim, M. S. (1994). The use of narrow spectral bands for improving remote sensing estimation of fractionally absorbed photosynthetically active radiation (fAPAR). Master’s thesis, University of Maryland, College Park.
Kim, M. S., Daughtry, C. S. T., Chappelle, E. W., McMurtrey, J. E., III, & Walthall, C. L. (1994). The use of high spectral resolution bands for estimating absorbed photosynthetically active radiation (APAR). In Proceedings of the 6th Symposium
41
on Physical Measurements and Signatures in Remote Sensing, January 17–21, 1994, Val D’Isere, France (pp. 299–306).
Knudby, A., LeDrew, E., & Brenning, A. (2010). Predictive mapping of reef fish species richness, diversity and biomass in Zanzibar using IKONOS imagery and machine-learning techniques. Remote Sensing of Environment, 114, 1230-1241.
Kohavi, R. (1995). A study of cross-validation and bootstrap for accuracy estimation and model selection. In IJCAI’95: Proceedings of the 14th International Joint Conference on Artificial Intelligence (Vol. 2, pp. 1137-1145). San Francisco, CA: Kaufmann.
Lamb, D. W., & Brown, R. B. (2001). PA—Precision agriculture: Remote-sensing and mapping of weeds in crops. Journal of Agricultural Engineering Research, 78(2), 117-125.
Lambert, D., & Lowenberg-De Boer, J. (2000). Precision agriculture profitability review. Retrieved from Purdue University: http://agriculture.purdue.edu/SSMC/Frames /newsoilsX.pdf
le Maire, G., François, C., & Dufrêne, E. (2004). Towards universal broad leaf chlorophyll indices using PROSPECT simulated database and hyperspectral reflectance measurements. Remote Sensing of Environment, 89, 1-28.
Liaghat, S., & Balasundram, S. K. (2010). A review: The role of remote sensing in precision agriculture. American Journal of Agricultural and Biological Sciences, 5(1), 50-55.
MacDonald, R. B., & Hall, F. G. (1980). Global crop forecasting. Science, 208(4445), 670-679.
Mathur, A., & Foody, G. M. (2008). Crop classification by support vector machine with intelligently selected training data for an operational application. International Journal of Remote Sensing, 29, 2227-2240.
Miller, J. R., Hare, E. W., & Wu, J. (1990). Quantitative characterization of the vegetation red edge reflectance 1. An inverted-Gaussian reflectance model. Remote Sensing, 11, 1755-1773.
Miura, T., & Huete, A. R. (2009). Performance of three reflectance calibration methods for airborne hyperspectral spectrometer data. Sensors, 9, 794-813.
Moran, J. A., Mitchell, A. K., Goodmanson, G., & Stockburger, K. A. (2000). Differentiation among effects of nitrogen fertilization treatments on conifer seedlings by foliar reflectance: A comparison of methods. Tree Physiology, 20, 1113-1120.
42
Moran, M. S., Inoue, Y., & Barnes, E. M. (1997). Opportunities and limitations for image-based remote sensing in precision crop management. Remote Sensing of Environment, 61, 319-346.
MosaicMill. (2009). EnsoMOSAIC image processing user’s guide. Version 7.3. Vantaa, Finland: Author.
Moser, G., & Serpico, S. B. (2009). Automatic parameter optimization for support vector regression for land and sea surface temperature estimation from remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 47(3), 909-921.
Murakami, Y., Tsuyama, M., Kobayashi, Y., Kodama, H., & Iba, K. (2000). Trienoic fatty acids and plant tolerance of high temperature. Science, 287(5452), 476-479.
Myneni, R. B., Maggion, S., Iaquinta, J., Privette, J. L., Gobron, N., Pinty, B., . . . Williams, D. L. (1995). Optical remote sensing of vegetation: Modeling, caveats, and algorithms. Remote Sensing of Environment, 51, 169-188.
Neale, C. M., & Crowther, B. G. (1994). An airborne multispectral video/radiometer remote sensing system: Development and calibration. Remote Sensing of Environment, 49, 187-194.
Niinemets, Ü. & Tenhunen, J. D. (1997). A model separating leaf structural and physiological effects on carbon gain along light gradients for the shade‐ tolerant species Acer saccharum. Plant, Cell & Environment, 20, 845-866.
Pal, M., & Mather, P. M. (2005). Support vector machines for classification in remote sensing. International Journal of Remote Sensing, 26, 1007-1011.
Parry, C., Blonquist, J., & Bugbee, B. (2014). In situ measurement of leaf chlorophyll concentration: Analysis of the optical/absolute relationship. Plant, Cell & Environment, 37, 2508-2520.
Peñuelas, J., & Filella, I. (1998). Visible and near-infrared reflectance techniques for diagnosing plant physiological status. Trends in Plant Science, 3(4), 151-156.
Pierce, F. J., & Nowak, P. (1999). Aspects of precision agriculture. Advances in Agronomy, 67, 1-85.
Raison, J. K., Roberts, J. K., & Berry, J. A. (1982). Correlations between the thermal stability of chloroplast (thylakoid) membranes and the composition and fluidity of their polar lipids upon acclimation of the higher plant, Nerium oleander, to growth temperature. Biochimica et Biophysica Acta (BBA)—Biomembranes, 688(1), 218-228.
43
Rouse, J. W., Jr., Haas, R. H., Schell, J. A., Deering, D. W., & Harlan, J. C. (1974). Monitoring the vernal advancements and retrogradation (green wave effect) of natural vegetation. Greenbelt, MD: NASA.
Salvucci, M. E., & Crafts‐ Brandner, S. J. (2004). Inhibition of photosynthesis by heat stress: The activation state of Rubisco as a limiting factor in photosynthesis. Physiologia Plantarum, 120, 179-186.
Schellberg, J., Hill, M. J., Gerhards, R., Rothmund, M., & Braun, M. (2008). Precision agriculture on grassland: Applications, perspectives and constraints. European Journal of Agronomy, 29(2), 59-71.
Seelan, S. K., Laguette, S., Casady, G. M., & Seielstad, G. A. (2003). Remote sensing applications for precision agriculture: A learning community approach. Remote Sensing of Environment, 88, 157-169.
Shanahan, J. F., Holland, K. H., Schepers, J. S., Francis, D. D., Schlemmer, M. R., & Caldwell, R. (2003). Use of a crop canopy reflectance sensor to assess corn leaf chlorophyll content. In Digital imaging and spectral techniques: Applications to precision agriculture and crop physiology (pp. 135-150). Madison, WI: American Society of Agronomy.
Shanahan, J. F., Schepers, J. S., Francis, D. D., Varvel, G. E., Wilhelm, W. W., Tringe, J. M., . . . Major, D. J. (2001). Use of remote-sensing imagery to estimate corn grain yield. Agronomy Journal, 93, 583-589.
Sharkey, T. D. (2005). Effects of moderate heat stress on photosynthesis: Importance of thylakoid reactions, Rubisco deactivation, reactive oxygen species, and thermos tolerance provided by isoprene. Plant, Cell & Environment, 28, 269-277.
Sims, D. A., & Gamon, J. A. (2002). Relationships between leaf pigment content and spectral reflectance across a wide range of species, leaf structures and developmental stages. Remote Sensing of Environment, 81, 337-354.
Smith, A. M., Bourgeois, G., Teillet, P. M., Freemantle, J., & Nadeau, C. (2008). A comparison of NDVI and MTVI2 for estimating LAI using CHRIS imagery: a case study in wheat. Canadian Journal of Remote Sensing, 34, 539-548.
Souza, C. R. (2010). Kernel functions for machine learning applications. Retrieved from http://crsouza.blogspot.com/2010/03/kernel-functions-for-machine-learning.html
Stone, M. L., Solie, J. B., Raun, W. R., Whitney, R. W., Taylor, S. L., & Ringer, J. D. (1996). Use of spectral radiance for correcting in-season fertilizer nitrogen deficiencies in winter wheat. Transactions of the ASAE, 39, 1623-1631.
44
Swinton, S. M., & Lowenberg-DeBoer, J. (1998). Evaluating the profitability of site-specific farming. Journal of Production Agriculture, 11, 439-446.
Thayananthan, A., Navaratnam, R., Stenger, B., Torr, P. H., & Cipolla, R. (2008). Pose estimation and tracking using multivariate regression. Pattern Recognition Letters, 29, 1302-1310.
Ticlavilca, A. M., McKee, M., & Walker, W. R. (2013). Real-time forecasting of short-term irrigation canal demands using a robust multivariate Bayesian learning model. Irrigation Science, 31(2), 151-167.
Tipping, M. E. (2001). Sparse Bayesian learning and the relevance vector machine. The Journal of Machine Learning Research, 1, 211-244.
Tipping, M. E. (2004). Bayesian inference: An introduction to principles and practice in machine learning. In O. Bousquet, U. von Luxburg, & G. Ratsch (Eds.), Advanced lectures on machine learning (pp. 41-62). Berlin, Germany: Springer Berlin Heidelberg.
Tipping, M. E., & Faul, A. C. (2003). Fast marginal likelihood maximisation for sparse Bayesian models. In C. M. Bishop & B. J. Frey (Eds.), Proceedings of the Ninth International Workshop on Artificial Intelligence and Statistics (Vol. 1, No. 3). Society for Artificial Intelligence and Statistics.
Torres, A. F., Walker, W. R., & McKee, M. (2011). Forecasting daily potential evapotranspiration using machine learning and limited climatic data. Agricultural Water Management, 98, 553-562.
Vapnik, V. (2000). The nature of statistical learning theory. Dordrecht, The Netherlands: Springer Science & Business Media.
Verrelst, J., Alonso, L., Camps-Valls, G., Delegido, J., & Moreno, J. (2012). Retrieval of vegetation biophysical parameters using Gaussian process techniques. IEEE Transactions on Geoscience and Remote Sensing, 50, 1832-1843.
Verrelst, J., Schaepman, M. E., Malenovský, Z., & Clevers, J. G. (2010). Effects of woody elements on simulated canopy reflectance: Implications for forest chlorophyll content retrieval. Remote Sensing of Environment, 114, 647-656.
Wallace, L., Lucieer, A., Watson, C., & Turner, D. (2012). Development of a UAV-LiDAR system with application to forest inventory. Remote Sensing, 4, 1519-1543.
Weis, E. (1981). Reversible heat-inactivation of the Calvin cycle: A possible mechanism of the temperature regulation of photosynthesis. Planta, 151(1), 33-39.
45
Wood, C. W., Tracy, P. W., Reeves, D. W., & Edmisten, K. L. (1992). Determination of cotton nitrogen status with a handheld chlorophyll meter. Journal of Plant Nutrition, 15, 1435-1448.
Yadawa, U. L. (1986). A rapid and nondestructive method to determine chlorophyll in intact leaves. HortScience, 21, 1449-1450.
Yoder, B. J., & Pettigrew-Crosby, R. E. (1995). Predicting nitrogen and chlorophyll content and concentrations from reflectance spectra (400–2500 nm) at leaf and canopy scales. Remote Sensing of Environment, 53, 199-211.
Yuan, J., Wang, K., Yu, T., & Fang, M. (2007). Integrating relevance vector machines and genetic algorithms for optimization of seed-separating process. Engineering Applications of Artificial Intelligence, 20, 970-979.
Zaman, B., Jensen, A., Clemens, S. R., & McKee, M. (2014). Retrieval of spectral reflectance of high resolution multispectral imagery acquired with an autonomous unmanned aerial vehicle. Photogrammetric Engineering & Remote Sensing, 80, 1139-1150.
Zarco‐ Tejada, P. J., & Miller, J. R. (1999). Land cover mapping at BOREAS using red edge spectral parameters from CASI imagery. Journal of Geophysical Research: Atmospheres (1984–2012), 104(D22), 27921-27933.
Zarco-Tejada, P. J., Miller, J. R., Morales, A., Berjón, A., & Agüera, J. (2004). Hyperspectral indices and model simulation for chlorophyll estimation in open-canopy tree crops. Remote Sensing of Environment, 90, 463-476.
Zarco-Tejada, P. J., Miller, J. R., Noland, T. L., Mohammed, G. H., & Sampson, P. H. (2001). Scaling-up and model inversion methods with narrowband optical indices for chlorophyll content estimation in closed forest canopies with hyperspectral data. IEEE Transactions on Geoscience and Remote Sensing, 39, 1491-1507.
46
CHAPTER 3
USE OF HIGH RESOLUTION MULTISPECTRAL IMAGERY TO ESTIMATE
PLANT NITROGEN IN OATS (AVENA SATIVA).
Abstract
Remote sensing applications for precision agriculture depend on acquiring
actionable information at high spatial resolution and at a temporal frequency appropriate
for timely responses. This study was conducted to determine if high resolution canopy
reflectance could be used to evaluate leaf Nitrogen (N) status in oats (Avena sativa) for
precision agriculture. An unmanned aerial system platform named AggieAirTM, was used
to acquire high-resolution imagery in the visual, near infrared, and thermal infrared
spectra. This study reports the development of a relevance vector machine (RVM)
coupled with forward input selection process to a dataset composed of reflectance from
high-resolution multispectral imagery (visible, near-infrared, thermal infrared imagery),
and vegetative indices, in conjunction with in situ leaf N sampled and analyzed in the
laboratory, to estimate leaf N content from remotely sensed data at 15 cm resolution. The
results indicate that an RVM with a Gaussian kernel type and kernel width of 0.28,
having Near-infrared, green ratio vegetation index, red/green ratio index, and simple ratio
as the selected set of inputs, can be used to spatially estimate Leaf Nitrogen with a root
mean-squared error (RMSE) of 0.48 mg/100 mg dry tissue (DT), efficiency of 0.76, and 6
relevance vectors.
47
Keywords: Remote Sensing, High Spatial Resolution Imagery, Relevance Vector
Machine, Precision Agriculture, Leaf Nitrogen
Introduction
Monitoring crops and assessing their nutrient status throughout the vegetation
growth is a fundamental practice in precision agriculture. An essential nutrient for plant
growth is nitrogen (N), and when absent a significant decrease in both the photosynthetic
and CO2 assimilation capacity in crops is reported (Prsa, Stampar, Vodnik, & Veberic,
2007; Tracy, Hefner, Wood, & Edmisten, 1992). Farmers must balance the competing
goals of supplying enough N to their crops to reach their production goals while
minimizing the loss of N to the environment (Daughtry et al., 2000). When applied in
excess, N represents a threat to water quality because it can increase the chance of nitrate
contamination of surface and groundwater (Errebhi, Rosen, Gupta, & Birong, 1998; Ju,
Kou, Zhang, & Christie, 2006). Over-application of N fertilizer can also be an economic
loss to a farmer. Precise plant N evaluation has the potential to aid farmers in creating
this balance, by early detection of N stress and matching N supply with crop N
requirement at the correct rate, place, and time.
Traditionally, many methods have been utilized for determining crop N status,
including tissue analysis performed on leaves sampled from the field (Roth, Fox, &
Marshall, 1989). This method is known to be destructive, sample intensive, expensive,
and time consuming, with a significant lag between collecting the samples, performing
the laboratory analysis, and generating the final recommendation. This delay can result in
48
a missed opportunity in providing the crop with the needed nutrient in a timely fashion
and consequently influence the crop productivity. Recently, portable optical instruments
like chlorophyll meters are used to assess N status in crops (Filella, Serrano, Peñuelas,
1995; Han, Hedrickson, & Ni, 2001). These meters are less invasive, nondestructive, and
relatively faster in providing results (Gianquinto, Sambo, & Pimpini, 2003; Schepers,
Francis, Vigil, & Below, 1992; Turner & Jund, 1991). However, they are only practical at
leaf level and limited to evaluating plant status in small areas. This technique measures
the transmittance of radiation through a leaf in two wavelength bands centered near 650
nm and 940 nm (Peterson, Blackmer, Francis, & Schepers, 1993). The meter is actually
measuring leaf chlorophyll, and since the majority of leaf N is contained in chlorophyll
molecules, N content is deducted indirectly from those readings. However, some
researchers challenged the general polynomial equation commonly used to relate N to
measured chlorophyll and proposed an exponential equation that forces a more
appropriate fit (Markwell, Osterman, & Mitchell, 1995). Others commented that at high
N levels, N could be found in the NO3- N form and not only in chlorophyll molecules;
thus, the relationship between leaf chlorophyll and leaf N concentration may be nonlinear
(Wood, Reeves, & Himelrick, 1993). The successful usage of these meters when
assessing N content is also affected by the crop type, growth stages, measurement
positions on leaves, and environmental conditions (Chapman & Barreto, 1997; Ramesh et
al., 2002; Schepers et al., 1992; Turner & Jund, 1991). Both of these methods can
potentially lead to over- or underassessment of N status in the crops and as a result
inaccurate recommendations.
49
The importance and interest in leaf N and its spatial variation suggest a need to
analyze it in a spatially explicit and extensive manner. For instance, a leaf N content map
displaying spatial variations at canopy level could facilitate diagnosis of nutrient
availability and limitation and help to eliminate the need for intensive field sampling or
nutrient addition. Obtaining such data for spatial representation could be achieved using
remote sensing platforms. Remote sensing provides rapid, quantitative information about
crops and, above all, does so nondestructively for diagnosing the spatial variability of
crop field properties. Remotely sensed data infer canopy leaf N status based on leaf
reflectance and transmittance measurements.
An early study by Thomas and Oerther (1972) found that reflectance in the green
and red regions of the electromagnetic spectrum were highly correlated to leaf N content
determined using the Kjeldahl method. Additional research reported by Martin, Shenk,
and Barton (1989) established the relationship between crop N content and canopy
reflectance at the visible (VIS, 400–700 nm) and near-infrared (NIR, 700–1100 nm)
regions. In addition, light reflectance at wavelengths near 550 nm from individual leaves
was found to be a good indicator of N stress in corn crops (Blackmer & Schepers, 1995).
Researchers also developed various combinations of bands knows as vegetative indices
that are linearly related to N concentration, such as the Red-edge Index, R740/R720
(Mokhele & Ahmed, 2010). Canopy reflectance in the red and NIR regions have reported
success at determining crop N content, due to the ability to detect changes associated
with chlorophyll content (Guyot, 1991). Researches have studied leaf N reflectance
response on corn (Blackmer, Schepers, & Varvel, 1994: Lee, Searcy, & Kataoka, 1999),
50
sweet pepper (Thomas & Oerther, 1972), sugarcane (Miphokasap, Honda, Vaiphasa,
Souris, & Nagai, 2012), rice (Stroppiana, Boschetti, Brivio, & Bocchi, 2009; Yi, Huang,
Wang, Wang, & Liu, 2007; Zhu, Zhou, Yao, Tian, & Cao, 2007), and many others.
With airborne and satellite multispectral sensors as the providers of remotely
sensed data, various shortcomings emerged (Merlin et al., 2010; Mulla, 2013). The ability
to deliver information for N management at high spatial resolution as dictated by
precision farming practices was the biggest limitation. In addition, the high cost of
images from aircraft, the infrequency of satellite overpasses, the interference of the
weather conditions with the imagery, and delays between image capture and availability
of usable data were major concerns for farmers. To implement precision N management
successfully, efficient technologies to diagnose crop N status are needed to determine in-
season, site-specific crop N requirements (Li et al., 2010). Unmanned aerial systems
(UAS) have been proposed for precision agriculture applications (Berni, Zarco-Tejada,
Suárez, & Fereres, 2009; Kooistra et al., 2013; Zhang & Kovacs, 2012). UAS can provide
imagery with a higher spatial resolution, allow for more flexible acquisition times, and
can be flown more frequently compared to satellite imagery (Zhang & Kovacs, 2012).
The current research explores the high spatial resolution imagery collected by a
UAS system to estimate leaf N content. The imagery acquires reflectance information
from the VIS, NIR, and thermal infrared (TIR) spectra with a spatial resolution of 15–60
cm. The estimated leaf N content developed by the proposed model was validated against
the ground truthing data collected. The generated high-resolution leaf N map could help
51
farmers in customizing their fertilization practices to suit the variability and needs of the
crop.
Our objective was to use canopy reflectance collected by AggieAir for estimating
the crop N status within fields. Moreover, this research also would identify which inputs
(e.g., individual reflectance, vegetative indices) are most sensitive for detection of crop N
status differences. In conducting the work, we evaluated a relatively inexpensive UAS
application in precision agriculture.
Material and Methods
Study Area
The study area is on a privately owned agricultural land in Scipio, Utah (39°14'N
112°6'W; see Figure 3.1). The field is equipped with a modern center pivot irrigation
sprinkler system to supply water for oat crops. The study reported here was carried out in
the summer of 2013. Four flights were flown over the study site. The flights on May 16,
June 1, June 9, and June 17 reflected the four stages of growth: 10 days after planting,
early growth, midgrowth, and early flowering. The study area was restricted to the
northwest quarter of the center pivot so that samples could be collected within a close
time frame relative to the AggieAir flight.
52
Figure 3.1. The location of the study area in Utah (left), The quarter used in leaf Nitrogen estimation model (right).
Instrumentation: Remote Sensing Platform AggieAir
AggieAir is a UAS designed to carry a payload to acquire aerial imagery in the
VIS, NIR, and TIR spectra. The unmanned aerial vehicle (UAV) navigates over an area
of interest according to a preprogrammed flight plan. While operational, the payload
computer instructs the three cameras to acquire imagery and records the position and
orientation of the aircraft when each image is taken. The UAS aircraft is battery powered
with a wingspan of 2.5 m and weight of 6.35 kg. AggieAir can fly in the air for up to 60
minutes and within elevation of 200–1000m.
The VIS camera used in AggieAir is a Canon S-95, with a 10-megapixel CCD
sensor and an ISO range of 80 to 3200. The NIR camera is an identical Canon S-95,
modified by replacing the manufacturer’s optical filter with a Wratten 87 NIR filter that
allows NIR wavelengths of 750 nm. AggieAir also carries a small, low-power,
microbolometer thermal camera from Infrared Cameras, Inc. (2014). The VIS-NIR
imagery is of a spatial resolution of 0.15 m; the TIR imagery is of 0.6 m spatial
resolution. The relative spectral responses of the VIS-NIR cameras were not provided by
Study site
53
the manufacturers but were obtained using the algorithm provided by Jiang, Liu, Gu, and
Susstrunk (2013). The wavelength range peaks around 420, 500, 600, and 800 nm,
respectively, for blue, green, red, and NIR sensors. Please refer to Chapters 2 and 4 of
this dissertation for more details about the UAV and the sensors.
AggieAir Data Processing
Following image acquisition, the imagery is imported into EnsoMOSAIC
(MosaicMill, 2012). EnsoMOSAIC is a photogrammetry software for aerial image
processing. The software generates hundreds of tie-points between overlapping images
by using photogrammetric principles in conjunction with image GPS log file data and
exterior orientation information from the on-board cameras to refine the estimate of the
position and orientation of individual images. The resulting image is an orthorectified
mosaic in 8-bit digital format. This mosaic is later converted to a mosaic of reflectance
values by applying a modified reflectance mode method (Clemens, 2012; Zaman, Jensen,
Clemens, & McKee, 2014). This radiometric normalization is the ratio of the digital
number from the mosaic to the digital number from a spectralon white reflectance panel
with known reflectance coefficients, multiplied by the reflectance factor, which accounts
for the zenith angle of the sun at the time, date, and location of the photos. The end
product is a four-layered image (blue, green, red, NIR) that is geometrically rectified and
radiometricaly corrected. As for the TIR imagery processing, the collected imagery is
also processed in EnsoMOSAIC, and the resulting thermal mosaic is composed of
brightness temperature in degrees Celsius. These mosaics are later converted to
radiometric temperatures following the procedure of Jensen (2014).
54
Ground Data Acquisition
Intensive ground sampling of plant N was conducted simultaneously with the
AggieAir flights at precise GPS locations. The GPS data were collected using an rtkGPS
with < 1 mm precision in a 1 Hz bandwidth (Trimble R8, Global Navigation Satellite
System, Dayton, Ohio). Leaves from the fully expanded top-of-canopy area were
collected. The samples were placed in paper bags inside an iced cooler and taken for
laboratory analysis. In the laboratory, plant samples were analyzed for leaf N content
(percent). Each plant sample was dried in an oven at 70°C, dry weights measured and
converted into dry biomass (kg/m2), and analyzed for N content.
Relevance Vector Machine (RVM)
The RVM was first introduced in 2000 and comprehensively detailed by Tipping
(2001). The RVM is a sparse Bayesian learning algorithm that relies on probabilistic
Bayesian principles to produce accurate predictions with a good generalization
performance (Tipping, 2001). The original learning algorithm was improved by Tipping
and Faul (2003), providing faster learning.
The RVM is based on a linear model where a prediction, y, is given as a weighted
sum of x basis functions. The model “learns” the dependence between input and output
target with the purpose of making accurate predictions of the target vector for
previously unseen values of as shown in Equation 1:
)(xwy (1)
Where a vector of weight parameters is, TNxxfxxfx )],(),...,,(,1[)( 1 is a design
matrix of vectors of kernel basis functions , and is the error that for
y
x
w
1N f ε
55
algorithmic simplicity is assumed to be zero-mean Gaussian with variance .
Through training of the RVM, the optimal weights are determined. The few
nonzero weights correspond to the so-called relevance vectors that are the sparse core of
the RVM model (Tipping & Faul 2003). The optimal parameters are then used to obtain
the optimal weight matrix with optimal covariance and mean. Model complexity,
overfitting, and computational expenses are controlled by setting weights to zero to
induce sparsity (Bachour, Walker, Ticlavilca, McKee, & Maslova, 2014). The
mathematical formulation, likelihood maximization, and optimization procedure of the
RVM are discussed in detail in Chapter 2. For further reading please refer to Tipping
(2001); Tipping and Faul (2003); and Thayananthan, Navaratnam, Stenger, Torr, and
Cipolla (2008).
Generating the Model Inputs and Targets
Three of the four flights (10 days after planting May 16, early growth June 1, and
midgrowth June 9), were used to generate the model training input data. The dataset
contains coincident in situ leaf N content and remote sensing reflectance measurements.
Each GPSed location where leaves were collected and tested for N content was paired by
its corresponding reflectance value. This was achieved by using the ArcGIS 10.2 ArcMap
software spatial analyst tool (Extract Multi Values to Points). In addition to the
reflectance in the VIS-NIR spectra, additional vegetative indices were generated. The
vegetative indices selected for this study were reported to be sensitive in estimating N
(Gitelson, Viña, Ciganda, Rundquist, & Arkebauer, 2005; Shanahan et al., 2003). Table
3.1 shows the indices formulations. In preliminary runs, different potential inputs were
2
56
explored. However, details on these preliminary runs are not reported in this study
because of their low statistical performance. A statistical description of the dataset is
presented in Table 3.2.
Table 3.1
Vegetative Indices Formulation Used for Leaf Nitrogen Estimation
Index Formula Reference
Normalized difference vegetation index (NDVI)
(RNIR – Rred) / (RNIR +RRed) Rouse, Haas, Schell, Deering, & Harlan, 1974
Leaf area indexa ln [(NDVI – NDVI max) / (NDVI min –NDVI max)] / -0.54
Duchemin et al., 2006; Smith, Bourgeouis, Teillet, Freemantle, & Nadeau, 2008
Green NDVI (RNIR – Rgreen) / (RNIR + Rgreen) Buschmann & Nagel, 1993
Ratio vegetation index NIR/red Jordan, 1969
Green ratio vegetation index (GRVI)
NIR/green Sripada, Heiniger, White, & Meijer, 2006
Red/green ratio index (RGRI)
Red/green Yang, Willis, & Mueller, 2008
Simple ratios (SR) Blue/green Birth & McVey, 1968 Note. NIR = near-infrared. aLAI was calculated empirically and not validated by field measurements.
57
Table 3.2
Statistical Description of the Dataset Used for Leaf Nitrogen Estimation
Index Range Mean ± SD
Potential inputs
AggieAir inputs
Blue 0.11–0.36 0.15 ± 0.04
Green 0.20–0.49 0.26 ± 0.05
Red 0.15–0.51 0.22 ± 0.07
Near-infrared 0.51–0.61 0.57 ± 0.02
Thermal (oC) 23.11–36.16 29.88 ± 4.13
Indices inputs
Green NDVI 0.00–0.46 0.35 ± 0.10
Ratio vegetation index 1.03–3.78 2.70 ± 0.60
Green ratio vegetation index (GRVI) 1.17–2.74 2.16 ± 0.33
Red/green ratio index (RGRI) 0.72–1.13 0.81 ± 0.08
Normalized difference vegetation index (NDVI) 0.00–0.78 0.44 ± 0.12
Simple ratio (SR) 0.47–0.73 0.56 ± 0.03
Leaf area index (m2/m2) 0.00–4.23 2.41 ± 0.90
Target output
Leaf nitrogen (mg/100 mg DT ) 0.00–5.36 3.905 ± 0.91
Model Configuration and Performance
Three of the images collected over the study area were used to build the model
(May 16, June 1, June 9). The remaining imagery (June 17) was later used to test the
model as unseen data scenarios. The RVM code used in this study is based on the
MATLAB code provided via Michael E. Tipping’s website. A key feature of this
statistical learning algorithm is that it can yield a solution function that depends on only a
58
very small number of training samples. These are called relevance vectors and are those
samples from the training set that have nonzero weights. In the RVM regression model,
the weight of each input is governed by a set of hyperparameters that describe the
posterior distribution of these weights. They are estimated iteratively during the machine
learning training step. The model “learns” the dependence between input and output
target with the purpose of making accurate predictions of the target vector, which was the
leaf N content in this study.
The model was developed with forward input selection process, in an attempt to
only include the input variables of most significance (Guyon & Elisseeff, 2003). In each
iteration, inputs with the best statistical performance were added to the set of inputs until
no new predictors could be added. The RVM model was tested using six kernel types:
Gauss, Laplace, spline, Cauchy, thin plate spline, and bubble. The performance of the
model was evaluated by comparing the root mean squared error (RMSE) and the Nash-
Sutcliffe efficiency (E); these two parameters have been widely used to evaluate the
performance of RVM models. The larger the value of E and the smaller the value of
RMSE, the greater the precision and accuracy of the model to predict leaf N. The optimal
values of these parameters were selected by trial and error procedure to obtain the best
RMSE and E values. The RMSE and E are computed as shown in Equations 2 and 3,
respectively:
N
yyRMSE
N
ttt
1
2)ˆ( (2)
59
N
t t
N
t t
yy
yyE
12
12
)(
)ˆ(1 (3)
where, ŷt = predicted nitrogen content ; yt = measured nitrogen content; = mean of the
observed nitrogen content; = mean of the estimated nitrogen content; and N = total
number of observations.
Results and Discussion
Each of the six kernel types was tested over a wide range of kernel widths (10-5–
105), and RMSE and E were calculated for all of the resulting models to assess their
predictive capabilities. An embedded loop in the coding model was developed to
represent the forward selection tool. For each type of kernel and its corresponding width,
the RVM was first run using all of the 8 inputs, consequently generating all of the needed
statistical model performance estimates to assess the model. A set of defined iterations
then added, in order, the input with the second best set of statistics, thus including the
input the most relevant to the target function. After numerous computational runs, the
model with the highest prediction accuracy was selected. The optimal kernel width,
kernel type, selected inputs, and statistical performance are presented in Table 3.3.
Of all runs conducted across the six kernel types, the NIR band followed by the
GRVI band were the two inputs selected first and second. This finding suggested these
two bands in the sensor used possess the most relevant information for estimating leaf N.
Canopy reflectance in the NIR regions have reported success at determining crop N
y
y
60
content, due to the ability to detect changes associated with chlorophyll content and
decreasing cell layers (Guyot, 1991; Thomas & Oerther, 1977).
Table 3.3
Selected Model Characteristics
Characteristic Statistic
Kernel type Gaussian
Kernel width 0.28
Inputs Near-infrared (NIR), green ratio vegetation index (GRVI), red/green ratio index (RGRI), simple ratio (SR)
Root mean square error (RMSE) 0.48 mg/100 mg DT
Nash-Sutcliffe efficiency (E) 0.76
Relevance vectors 6
Reflectance in the NIR spectrum has been reported to be sensitive to N estimate
over potato (Fava et al., 2009). In addition, when N total biomass is measured in the
laboratory by a spectroscopy an NIR band is used for its detection.
As for the GRVI index, a study that was monitoring leaf nitrogen status in rice
with canopy spectral reflectance (Xue et al., 2004) reported that the GRVI index
(NIR/green ratio) was especially linearly related to total leaf N, independent of N level
and growth stage. The study recommended the index to be adopted for nondestructive
monitoring of N status in rice plants. As for the RGRI index, Gamon and Surfus (as cited
in Gitelson, Gritz, & Merzlyak, 2003) stated that this index is particularly influential in
the separation of nonvegetation pixels from vegetation pixels, allowing for easy
discrimination between soil (larger red/green ratio) and vegetation (smaller red/green
61
ratio). To the best knowledge of the authors, the SR index (blue/green) has not been
reported to be influential in any attempts of N estimation from canopy reflectance.
The leaf N estimates for the two images used to build the model was developed
based the unique characteristic (kernel type, width, and set of inputs) of the selected
model. Figure 3.2 illustrates the measured leaf N versus estimated values with a one
standard error confidence interval and the residual plot. It is evident from Figure 3.2 that
the measured and RVM predicted leaf nitrogen values are in agreement, where most of
the values lie within the 95% confidence interval. It is important to mention that zero leaf
nitrogen values representing bare soil pixels were successfully assigned a zero value by
the RVM predicted model.
The lower plot in Figure 3.2 presents the residual error over the value from the
three flights. It is noticeable that the residual errors of the first flight (May 16) are higher
than the other two flights (June 1st and 9th). This could be explained by the drastic
variation in the percentage of crop cover in the first flight compared to the second and
third. In flight one, vegetation was scares and bare soil was more prominent in the
imagery. The predictive RVM model performance is superior in full crop cover compared
to less homogenous covers.
The leaf N estimate maps for May 16, June 1, and June 9 generated from the
model, along with the corresponding true color image, are presented in Figure 3.3.
62
Figure 3.2. Measured versus RVM predicted leaf nitrogen in the two flights used to build the model (above) and residual plot of the selected mode (below).
As shown in Figure 3.3, the predicted leaf N maps show a visual good agreement
with the true-color maps. Two visual comments can be drawn from the imagery in Figure
3.3:
1. The wheel tracks and the access road that are located around the center pivot
had no vegetation cover, and the model successfully assigned a near-zero N
content to these features.
63
2. Two high vegetation horizontal and vertical lines protrude in the images.
Those were past ditch lines that had been used in flood irrigation activities
prior to the conversion of the field to a center pivot system
Figure 3.3. True-color maps (left), the estimated plant leaf nitrogen estimate map (right).
The greater water content in those areas caused the plants growing along those two lines
to be very vigorous. The model estimated higher leaf N content values given to the
64
plants in these areas.
To explore the behavior of the model with previously unseen data, the June 17
flight was used. Now that the model was established with a defined set of features
(inputs, kernel type, kernel width), June 17 data (early flowering stage) were entered in
the model to explore the model’s performance when subjected to totally unseen data. The
predicted leaf N map is shown in Figure 3.4.
Figure 3.4. June 17 image in false color (left), and estimated leaf nitrogen map (right).
Again, the predicted leaf N map for the fourth flight showed good association
with the 4:2:3 map. Areas of vigorous growth, bare soil, and low vegetation were similar
in the maps and represented similar growth patterns. This test reported an RSME of 0.68
mg/100 mg DT and E of 0.61 for this flight. This result showed that the model
successfully performed when given unseen data.
This quarter of the field is irrigated by a modern center pivot sprinkler irrigation
system with a capacity of 610 GPM. A local upstream reservoir feeds this pivot regularly.
65
The pivot is automated to spray water every 15 degrees of rotation, creating six distinct
manageable areas in this quarter as shown in Figure 3.5.
Figure 3.5. The six individually manageable areas as defined in this study.
The estimated leaf N is averaged over the six sectors for the four flown flights.
Figure 3.6 is showing the average change in each sector over the course of the cropping
season. The highest leaf N are seen in the last imagery (June 17th) coinciding with the
early flowering stage in oats. The lowest Leaf N per sector were prominent in the first
flight (May 16th) where sparse vegetation was present in sectors three to six.
The information presented in the bar graph in Figure 3.6 aids the farmer in
quantifying the leaf nitrogen content among the different sectors during the growing
season. This data would be helpful when compared with the fertilization application
schedule to assess its efficiency within the different sectors.
66
Figure 3.6. Average leaf nitrogen per sector during the crop growth.
Averaging estimates over manageable areas allows the decision maker to consider
applying different amounts of fertilizer per sector to better accommodate for the need and
specification of the sector. Formatting the output in a form that the farmer can use is
another beneficial step in this new field. Providing site-specific, practical
recommendations to farmers is the next step in increasing the adoption rate of precision
agriculture.
Conclusion
This paper presented the application of imagery from AggieAir, a remote sensing
platform, combined with machine learning algorithms (RVM) to estimate leaf N content
concentration as an important biophysical parameter to be used in precision agriculture.
The RVM modeling technique, coupled with forward input selection, was applied to a
data set composed of reflectance from high-resolution multispectral imagery (VIS-NIR),
0
1
2
3
4
5
6
1 2 3 4 5 6
Leaf
nit
roge
n m
g/1
00
mg
DT
Sectors
16-May 1-Jun 9-Jun 17-Jun
67
TIR imagery, and vegetative indices, in conjunction with in situ leaf N measurements.
Six kernel types were tested over a wide range of kernel widths. Model performance was
evaluated by comparing the RMSE and E of various models and later by visual
comparison. Leaf N estimation was best achieved with a model of kernel type: Gaussian
kernel width: 0.28; selected inputs: NIR, GRVI, RGRI, SR; RSME: 0.48 mg/100 mg DT;
E: 0.76; and six relevance vectors. Across all of the model scenarios the NIR band
followed by the GRVI band were the two inputs selected, agreeing with the literature.
These two bands possess the most relevant information for estimating leaf N. While the
model developed was trained and built to estimate leaf N in oats, a similar methodology
can be applied to train a predictive model of a different crop type. With the foreseeable
high availability and fast development occurring in the high resolution imagery
platforms, these estimation models could become more common and further embedded in
farmer management tools.
References
Bachour, R., Walker, W., Ticlavilca, A., McKee, M., & Maslova, I. (2014). Estimation of spatially distributed evapotranspiration using remote sensing and a relevance vector machine. Journal of Irrigation and Drainage Engineering, 140(8), 04014029. doi:10.1061/(ASCE)IR.1943-4774.0000754
Berni, J., Zarco-Tejada, P. J., Suárez, L., & Fereres, E. (2009). Thermal and narrowband multispectral remote sensing for vegetation monitoring from an unmanned aerial vehicle. IEEE Transactions on Geoscience and Remote Sensing, 47, 722-738.
Birth, G. S., & McVey, G. R. (1968). Measuring the color of growing turf with a reflectance spectrophotometer. Agronomy Journal, 60(6), 640-643.
Blackmer, T. M., & Schepers, J. S. (1995). Use of chlorophyll meter to monitor nitrogen status and schedule fertigation for corn. Journal of Production Agriculture, 8, 55-60.
68
Blackmer, T. M., Schepers, J. S., & Varvel, G. E. (1994). Light reflectance compared with other nitrogen stress measurements in corn leaves. Agronomy Journal, 86, 934-938.
Buschmann, C., & Nagel, E. (1993). In vivo spectroscopy and internal optics of leaves as basis for remote sensing of vegetation. International Journal of Remote Sensing, 14, 711-722.
Chapman, S. C., & Barreto, H. J. (1997). Using a chlorophyll meter to estimate specific leaf nitrogen of tropical maize during vegetative growth. Agronomy Journal, 89, 557-562.
Clemens, S. R. (2012). Procedures for correcting digital camera imagery acquired by the AggieAir remote sensing platform (Doctoral dissertation). Available from http:// digitalcommons.usu.edu/gradreports/186
Daughtry, C. S. T., Walthall, C. L., Kim, M. S., De Colstoun, E. B., & McMurtrey, J. E., III. (2000). Estimating corn leaf chlorophyll concentration from leaf and canopy reflectance. Remote Sensing of Environment, 74, 229-239.
Duchemin, B., Hadria, R., Erraki, S., Boulet, G., Maisongrande, P., Chehbouni, A., . . . Simonneaux, V. (2006). Monitoring wheat phenology and irrigation in Central Morocco: On the use of relationships between evapotranspiration, crops coefficients, leaf area index and remotely-sensed vegetation indices. Agricultural Water Management, 79(1), 1-27.
Errebhi, M., Rosen, C. J., Gupta, S. C., & Birong, D. E. (1998). Potato yield response and nitrate leaching as influenced by nitrogen management. Agronomy Journal, 90, 10-15.
Fava, F., Colombo, R., Bocchi, S., Meroni, M., Sitzia, M., Fois, N., & Zucca, C. (2009). Identification of hyperspectral vegetation indices for Mediterranean pasture characterization. International Journal of Applied Earth Observation and Geoinformation, 11, 233-243.
Filella, I., Serrano, L., & Peñuelas, J. (1995). Evaluating wheat nitrogen status with canopy reflectance índices and discriminat analysis. Crop Science, 35, 1400-1405.
Gianquinto, G., Sambo, P., & Pimpini, F. (2003). The use of SPAD-502 chlorophyll meter for dynamically optimizing the nitrogen supply in potato crop: First results. Acta Horticulturae, 627, 225-230.
Gitelson, A. A., Gritz, Y., & Merzlyak, M. N. (2003). Relationships between leaf chlorophyll content and spectral reflectance and algorithms for non-destructive chlorophyll assessment in higher plant leaves. Journal of Plant Physiology, 160, 271-282.
69
Gitelson, A. A., Viña, A., Ciganda, V., Rundquist, D. C., & Arkebauer, T. J. (2005). Remote estimation of canopy chlorophyll content in crops. Geophysical Research Letters, 32(8). doi:10.1029/2005GL022688
Guyon, I., & Elisseeff, A. (2003). An introduction to variable and feature selection. The Journal of Machine Learning Research, 3, 1157-1182.
Guyot, G. (1991). Optical properties of vegetation canopies. In M. D. Steven & J. A. Clark (Eds.), Application of remote sensing in agriculture (pp. 19-43). London, England: Butterworths.
Han, S., Hedrickson, L., & Ni, B. (2001). Comparison of satellite remote sensing and aerial photography for ability to detect in-season nitrogen stress in corn. St. Joseph, MI: American Society of Agricultural and Biological Engineers.
Infrared Cameras, Inc. (2014). Home page. Retrieved from http://www .infraredcamerasinc.com
Jensen, A. M. (2014). Innovative payloads for small unmanned aerial system-based personal remote sensing and applications. Doctoral dissertation, Utah State University, Logan.
Jiang, J., Liu, D., Gu, J., & Susstrunk, S. (2013). What is the space of spectral sensitivity functions for digital color cameras? In 2013 IEEE Workshop on Applications of Computer Vision (WACV) (pp. 168-179). doi:10.1109/WACV.2013.6475015
Jordan, C. F. (1969). Derivation of leaf area index from quality of light on the forest floor. Ecology, 50, 663-666.
Ju, X. T., Kou, C. L., Zhang, F. S., & Christie, P. (2006). Nitrogen balance and groundwater nitrate contamination: Comparison among three intensive cropping systems on the North China Plain. Environmental Pollution, 143, 117-125.
Kooistra, L., Suomalainen, J., Iqbal, S., Franke, J., Wenting, P., & Bartholomeus, H. (2013, September). Crop monitoring using a light-weight hyperspectral mapping system for unmanned aerial vehicles : First results for the 2013 season. Paper presented at the Workshop on UAV-based Remote Sensing Methods for Monitoring Vegetation, Cologne, Germany.
Lee, W., Searcy, S. W., & Kataoka, T. (1999, July). Assessing nitrogen stress in corn varieties of varying color. Paper presented at the ASAE Annual International Meeting, Toronto, Ontario, Canada.
Li, F., Miao, Y., Hennig, S. D., Gnyp, M. L., Chen, X., Jia, L., & Bareth, G. (2010). Evaluating hyperspectral vegetation indices for estimating nitrogen concentration of winter wheat at different growth stages. Precision Agriculture, 11, 335-357.
70
Markwell, J., Osterman, J. C., & Mitchell, J. L. (1995). Calibration of the Minolta SPAD-502 leaf chlorophyll meter. Photosynthesis Research, 46, 467-472.
Martin, G. C., Shenk, J. S., & Barton, F. E., II. (1989). Near infrared reflectance spectroscopy (NIRS): Analysis of forage quality (Agriculture Handbook No. 643). Washington, DC: U.S. Department of Agriculture.
Merlin, O., Duchemin, B., Hagolle, O., Jacob, F., Coudert, B., Chehbouni, G., . . . & Kerr, Y. (2010). Disaggregation of MODIS surface temperature over an agricultural area using a time series of Formosat-2 images. Remote Sensing of Environment, 114, 2500-2512. doi:10.1016/j.rse.2010.05.025
Miphokasap, P., Honda, K., Vaiphasa, C., Souris, M., & Nagai, M. (2012). Estimating canopy nitrogen concentration in sugarcane using field imaging spectroscopy. Remote Sensing, 4, 1651-1670.
Mokhele, T. A., & Ahmed, F. B. (2010). Estimation of leaf nitrogen and silicon using hyperspectral remote sensing. Journal of Applied Remote Sensing, 4(1), 043560-043560.
MosaicMill. (2009). EnsoMOSAIC image processing user’s guide. Version 7.3. Vantaa, Finland: Author.
Mulla, D. J. (2013). Twenty five years of remote sensing in precision agriculture: Key advances and remaining knowledge gaps. Biosystems Engineering, 114, 358-371. doi:10.1016/j.biosystemseng.2012.08.009
Peterson, T. A., Blackmer, T. M., Francis, D. D., & Schepers, J. S. (1993). Using a chlorophyll meter to improve N management (Nebguide G93–1171A). Lincoln: University of Nebraska.
Prsa, I., Stampar, F., Vodnik, D., & Veberic, R. (2007). Influence of nitrogen on leaf chlorophyll content and photosynthesis of “Golden Delicious” apple. Acta Agriculturae Scandinavica Section B—Soil and Plant Science, 57, 283-289.
Ramesh, K., Chandrasekaran, B., Balasubramanian, T. N., Bangarusamy, U., Sivasamy, R., & Sankaran, N. (2002). Chlorophyll dynamics in rice (Oryza sativa) before and after flowering based on SPAD (chlorophyll) meter monitoring and its relation with grain yield. Journal of Agronomy and Crop Science, 188, 102-105.
Roth, G. W., Fox, R. H., & Marshall, H. G. (1989). Plant tissue tests for predicting nitrogen fertilizer requirements of winter wheat. Agronomy Journal, 81, 502-507.
Rouse, J. W., Jr., Haas, R. H., Schell, J. A., Deering, D. W., & Harlan, J. C. (1974). Monitoring the vernal advancements and retrogradation (green wave effect) of natural vegetation. Greenbelt, MD: NASA.
71
Schepers, J. S., Francis, D. D., Vigil, M., & Below, F. E. (1992). Comparison of corn leaf nitrogen concentration and chlorophyll meter readings. Communications in Soil Science & Plant Analysis, 23, 2173-2187.
Shanahan, J. F., Holland, K. H., Schepers, J. S., Francis, D. D., Schlemmer, M. R., & Caldwell, R. (2003). Use of a crop canopy reflectance sensor to assess corn leaf chlorophyll content. In Digital imaging and spectral techniques: Applications to precision agriculture and crop physiology (pp. 135-150). Madison, WI: American Society of Agronomy.
Smith, A. M., Bourgeois, G., Teillet, P. M., Freemantle, J., & Nadeau, C. (2008). A comparison of NDVI and MTVI2 for estimating LAI using CHRIS imagery: A case study in wheat. Canadian Journal of Remote Sensing, 34, 539-548.
Sripada, R. P., Heiniger, R. W., White, J. G., & Meijer, A. D. (2006). Aerial color infrared photography for determining early in-season nitrogen requirements in corn. Agronomy Journal, 98, 968-977.
Stroppiana, D., Boschetti, M., Brivio, P. A., & Bocchi, S. (2009). Plant nitrogen concentration in paddy rice from field canopy hyperspectral radiometry. Field Crops Research, 111(1), 119-129.
Thayananthan, A., Navaratnam, R., Stenger, B., Torr, P. H., & Cipolla, R. (2008). Pose estimation and tracking using multivariate regression. Pattern Recognition Letters, 29, 1302-1310.
Thomas, J. R., & Oerther, G. F. (1972). Estimating nitrogen content of sweet pepper leaves by reflectance measurements. Agronomy Journal, 64(1), 11-13.
Thomas, J. R., & Oerther, G. F., Jr. (1977). Estimation of crop conditions and sugar cane yields using aerial photography. Proceedings of the American Society of Sugar Cane Technologies, 6, 93-99.
Tipping, M. E. (2001). Sparse Bayesian learning and the relevance vector machine. The Journal of Machine Learning Research, 1, 211-244.
Tipping, M. E., & Faul, A. C. (2003). Fast marginal likelihood maximisation for sparse Bayesian models. In C. M. Bishop & B. J. Frey (Eds.), Proceedings of the Ninth International Workshop on Artificial Intelligence and Statistics (Vol. 1, No. 3). Society for Artificial Intelligence and Statistics.
Tracy, P. W., Hefner, S. G., Wood, C. W., & Edmisten, K. L. (1992). Theory behind the use of instantaneous leaf chlorophyll measurement for determining mid-season cotton nitrogen recommendations. In 1992 Beltwide Cotton Conference Proceedings (439-443). Memphis, TN: National Cotton Council.
72
Turner, F. T., & Jund, M. F. (1991). Chlorophyll meter to predict nitrogen topdress requirement for semidwarf rice. Agronomy Journal, 83, 926-928.
Wood, C. W., Reeves, D. W., and Himelrick, D. G. (1993), Relationships between chlorophyll meter readings and leaf chlorophyll concentration, N status, and crop yield: A review. Proceedings of the Agronomy Society of New Zealand, 23, 1-9.
Xue, L., Cao, W., Luo, W., Dai, T., & Zhu, Y. (2004). Monitoring leaf nitrogen status in rice with canopy spectral reflectance. Agronomy Journal, 96(1), 135-142.
Yang, Z., Willis, P., & Mueller, R. (2008, November). Impact of band-ratio enhanced AWIFS image to crop classification accuracy. Paper presented at Pecora 17—The Future of Land Imagine: Going Operational, Denver, CO.
Yi, Q. X., Huang, J. F., Wang, F. M., Wang, X. Z., & Liu, Z. Y. (2007). Monitoring rice nitrogen status using hyperspectral reflectance and artificial neural network. Environmental Science & Technology, 41, 6770-6775.
Zaman, B., Jensen, A., Clemens, S. R., & McKee, M. (2014). Retrieval of spectral reflectance of high resolution multispectral imagery acquired with an autonomous unmanned aerial vehicle. Photogrammetric Engineering & Remote Sensing, 80, 1139-1150.
Zhang, C., & Kovacs, J. (2012). The application of small unmanned aerial systems for precision agriculture: A review. Precision Agriculture, 13, 693-712. doi:10.1007 /s11119-012-9274-5
Zhu, Y., Zhou, D., Yao, X., Tian, Y., & Cao, W. (2007). Quantitative relationships of leaf nitrogen status to canopy spectral reflectance in rice. Crop and Pasture Science, 58, 1077-1085.
73
CHAPTER 4
ASSESSMENT ON THE USE OF AGGIEAIR UNMANNED AERIAL SYSTEM
REMOTE SENSING CAPABILITIES TO ESTIMATE CROP
EVAPOTRANSPIRATION AT HIGH SPATIAL RESOLUTION
Abstract
Estimation of spatial distribution of evapotranspiration (ET) based on remotely
sensed imagery has become of critical importance for managing water in irrigated
agriculture at fine spatial scales. Currently, data acquired by conventional satellites
(Landsat, Advanced Space borne Thermal Emission and Reflection Radiometer
[ASTER], etc.) lack the needed spatial resolution to capture within-farm variability to
support crop water-consumption estimates. Newly available remote-sensing technologies,
called unmanned aerial systems (UAS), are proving to be able to deliver time-flexible,
cost-effective, tailored spatial information for decision-making purposes. In this study, a
UAS called AggieAirTM was used to acquire high-resolution imagery in the visible (VIS),
near-infrared (NIR, 0.15 m resolution), and thermal infrared (TIR) spectra (0.6 m
resolution) for estimation of crop water use or actual ET. AggieAir flew over two study
sites in Central Utah and the California Central Valley. The imageries were used as input
to an extensively used surface energy balance model designed for Landsat called
Mapping Evapotranspiration With Internalized Calibration (METRIC™), and results at
AggieAir and Landsat scales were compared. The discussion highlights the difference in
ET estimates and the implications of high-resolution ET map availability.
74
Keywords: Evapotranspiration, Remote Sensing, Unmanned Aerial Systems, High
Resolution Imagery, Precision Agriculture, METRIC
Introduction
Water scarcity is becoming a major concern in agriculture, and the fear that future
water availability might be threatened has been a driving force to improve existing crop
water demand assessments. Failing to meet crop water requirements negatively affects
vegetative growth, yield, and various product-quality attributes. Water requirements are
typically described by the term evapotranspiration (ET). ET is the largest outgoing water
flux from agricultural surfaces and the most difficult to measure directly.
ET is the combination of two processes that occur simultaneously: evaporation
and transpiration, whereby water is lost to the atmosphere from soil and vegetation
respectively. The ET rate is expressed in depth (e.g., mm, inches) per unit time (hour,
day). Estimated ET has been widely used in research related to crop water use (Penman,
1948; Thevs, Peng, Rozi, Zerbe, & Abdusalih, 2015), soil water availability prediction
(Oki & Kanae, 2006), flood and drought forecasting (Bouilloud et al., 2010; Sheffield &
Wood, 2008), and desertification (Zhou, Zhu, & Sun, 2002). Above all, the calculation of
the ET rate has become vital in the planning and management of irrigation practices. In
agriculture, key water-management decisions focus on knowing when to begin irrigation
for the growing season, how often to irrigate, and how much water to apply. In precision
agriculture particularly, these critical decisions are often complicated not only by micro
weather and climate variability, but also by intrafield variations in soil texture, terrain,
and soil fertility. Precision irrigation, a common practice of precision agriculture, has
75
developed in the last few years practices to assess intra- and interfield spatial variability
and to implement site-specific management systems (Arnó, Martínez Casanovas, Ribes
Dasi, & Rosell, 2009; Barnes et al., 2000; Bramley, 2010; Sadler, Evans, Stone, & Camp,
2005). These management practices are designed to supply the crop with the needed
amount of water at the smallest manageable scale to obtain optimum response. However,
ET estimates obtained by various current methodologies are incapable of supporting
precision irrigation decisions due to the limitations discussed below.
Traditionally, ET is measured at the either point or farm scale (e.g., lysimeter–
point level, and eddy covariance system–farm level). These measurements are
problematic because they are average or single values, time consuming, expensive, only
applicable to small homogenous surfaces, and incapable of accounting for the dynamic
nature of fluxes when dealing with large spatial scales (Kaheil, Rosero, Gill, McKee, &
Bastidas, 2008). Therefore, extrapolating ET rates from a single value to a large area
decreases the accuracy of the estimation (Mauser & Schädlich, 1998; Petropoulos,
Carlson, Wooster, & Islam, 2009). For operational applications, many water managers
adopted the methodology of measuring ET published in the Guidelines for Computing
Crop Water Requirements (Allen, Pereira, Raes, & Smith, 1998). This method consists of
estimating crop ET and land-surface fluxes for a crop canopy using a weather station
reference ET and a crop coefficient (Kc), where reference ET is retrieved using the
Penman-Monteith method. The Penman-Monteith equation used in this methodology
calculates the reference ET over grass, assuming optimum soil moisture conditions and a
constant value of surface canopy resistance (Allen et al., 1998). In reality, soil moisture
76
conditions or surface canopy resistance is neither optimal nor constant. Therefore, it is
fundamental to have a spatial distribution of the land-surface fluxes to achieve more
accurate ET estimates.
In 1973 Brown and Rosenberg (as cited in Nouri, Beecham, Kazemi, Hassanli, &
Anderson, 2013) first used TIR remotely sensed temperature for predicting
evapotranspiration; since then, quantitative estimation of ET based on remotely sensed
data has evolved rapidly. Imagery collected from different platforms over various
temporal and spatial scales in conjunction with meteorological data from ground stations
became the most efficient and economic technology in ET estimation (Nouri et al., 2013).
Remote sensing technology collects spatially distributed observations of the land-surface
fluxes as radiances and temperatures. Radiances and temperatures measured are
converted into land-surface characteristics such as albedo, leaf area index (LAI),
vegetation indices, surface emissivity and surface temperature. Better estimating these
inputs individually results in more reliable ET estimates. All remote sensing based ET
estimates make use of the information collected in the TIR, VIS bands (blue, green, and
red [BGR]), and NIR spectrum. The TIR band is the most important variable in ET
estimates because it plays a crucial role in sensible heat flux, ground heat flux, and the
balance of long-wave radiation.
Using remotely sensed satellite imagery (e.g., Landsat, moderate resolution
imaging spectroradiometer [MODIS], geostationary operational environmental satellite
[GOES], ASTER), algorithms have been developed to estimate ET with various inputs
and model-structure complexity. The literature is rich with review papers that categorize,
77
explain, compare, and discuss the advantages and shortcomings of the different remote
sensing based ET estimate methods (Allen, Pereira, Howell, & Jensen, 2011a; Calcagno,
Mendicino, Monacelli, Senatore, & Versace, 2007; Contreras, Jobbagy, Villagra,
Nosetto, & Puiddefabregas, 2011; Courault, Seguin, & Olioso, 2005; Kustas & Norman,
1996; Li et al., 2009; Liou & Kar, 2014; Van Der Tol & Parodi, 2012). Many of the
remotely sensed data techniques are based on the surface energy balance and have been
considered the most accurate methods for estimating ET over spatially varying areas. Of
the energy balance methods, the Surface Energy Balance Algorithms for Land
(Bastiaanssen, Menenti, Feddes, & Holtslag, 1998) and METRIC (Allen, Tasumi, &
Trezza, 2002, 2007) have been extensively used and tested for operational accuracy in the
western United States and other areas in the world (Allen, Tasumi, Morse, et al., 2007;
Allen, Tasumi, & Trezza, 2007; Bastiaanssen et al, 2005). METRIC, developed by Allen
et al. (2002), is an image processing model for mapping ET as a residual of the energy
balance. METRIC uses a self-calibration procedure that involves ground-based hourly
reference ET measurements and the selection of extreme (hot and cold) pixels within
agricultural surfaces (Gowda et al., 2007). Performance of the METRIC model has been
tested by Gowda et al. (2007); Santos, Lorite, Tasumi, Allen, and Ferreres (2008); and
Tasumi, Trezza, Allen, and Wright (2005). These studies showed that ET estimates
derived from METRIC were comparable with data derived from soil moisture budget and
lysimeter.
Despite the advantages of using remote sensing techniques to measure ET, several
disadvantages have been reported (Allen, Pereira, Howell, & Jensen, 2011a, 2011b;
78
Boegh et al., 2009; Chen, Chen, Ju, & Geng, 2005; Courault et al., 2005; Jiang et al.,
2009; McCabe & Wood 2006; Min & Lin, 2006; Mutiga, Su, & Woldai, 2010).
Challenges are common for researchers using various remote sensing platforms like
Landsat (Allen et al., 2002; Moran, Jackson, Raymond, Gay, & Slater, 1989), MODIS
(Zhang & Wegehenkel, 2006), and aircraft-based remote sensing (Chávez, Neale, Hipps,
Prueger, & Kustas, 2005; Gómez, Olioso, Sobrino, & Jacob, 2005; Jacob et al., 2002a;
Neale, Jayanthi, & Wright, 2003). These limitations include (a) dealing with
heterogeneous pixels, which leads to uncertainties of land-surface variables; (b) the
recurrence of satellite passage, which results in missed opportunities for addressing plant
needs in a timely fashion; (c) and the high costs associated with obtaining high-resolution
images, particularly airborne images (Allen et al., 2011b; Boegh et al., 2009; Chen et al.,
2005; Courault et al., 2005; Jiang et al., 2009; McCabe & Wood, 2006; Min & Lin, 2006;
Mutiga et al., 2010; Rana & Katerji, 2000; Stisen, Sandholt, Nørgaard, Fensholt, &
Jensen, 2008; Wu, Cheng, Lo, & Chen, 2010). These limitations are most significant
when the estimates are intended for use in precision agriculture.
A potential remote sensing platform that could overcome some of these
limitations is the unmanned aerial vehicle (UAV) system capable of acquiring high-
resolution imagery in the VIS, NIR and TIR bands. Some of the most attractive features
of UAVs include high spatial resolution and lower operating costs (Berni et al., 2009).
These characteristics make this platform particularly suitable for agricultural application,
where multitemporal flights are required for management applications and high
resolution is crucial for recognizing the inherent spatial variability associated with crops.
79
UAV use is currently increasing due to the advantages mentioned (Baluja et al., 2012;
Chiabrando, Lingua, & Piras, 2013) and because it has been successfully used to research
other agricultural topics (Berni, Zarco-Tejada, Suárez, González-Dugo, & Fereres, 2009;
Demarez, Duthoit, Baret, Weiss, & Dedieu, 2008; Johnson et al., 2003; Pinter et al.,
2003; Zarco-Tejada, Berni, Suárez, & Fereres, 2008). Other studies (Jacob et al., 2002a;
Kustas et al., 2003; Kustas, Jackson, Prueger, MacPherson, & Wolde, 2004; McCabe &
Wood, 2006) evaluated the scale influences on the estimation using remotely sensed data
using conventional satellite platforms (e.g., Landsat 7, MODIS, ASTER, airborn),
Nevertheless, remotely sensed data acquired by UAVs has yet to be explored in ET
estimation applications.
The main objective of this study was to investigate the compatibility of using
high-resolution data (15 cm) acquired by a UAV system called AggieAir in conjunction
with a modified version of METRIC. The developed ET estimate was compared to the
Landsat 8 derived ET. The paper describes the UAV platform as well as the processing
chain from designing the flight plan to orthorectification and lastly radiometric
calibration.
Data Collection
Site Description
The study was conducted in the summer of 2013 over Site S and the summer of
2014 over Site G (see Figure 4.1). Site S is 80 hectares of privately owned agricultural
land in Scipio, Utah (39°14'N 112°6'W at 1628 m). The plot, mainly composed of loamy
clay soil, is equipped with a center pivot sprinkler for irrigating and planted with oats
80
(Avena sativa) and alfalfa (Medicago Sativa). Site G is a 70-hectare commercial vineyard
located in Lodi, California (43°38'N, 116W at 40 m) in a wine-producing region, under
Mediterranean–continental climate. Site G is planted with pinot noir (Vitis vinfiera)
vines, grafted on 110R rootstock, and trained to the Geneva Double Curtain trellis. Vine
spacing was 2 m between rows and 1.5 m in the row. Irrigation was applied during the
growing season through drip irrigation. Standard management practices were applied in
this vineyard. Global radiation, wind speed, air temperature, and humidity were acquired
by a meteorological station located on each site.
Figure 4.1. The location of the study areas in Scipio, Utah, and Lodi, California.
UAV Description
AggieAir is a UAS designed to carry camera payloads to acquire aerial imagery.
The UAS aircraft is battery powered and equipped with a payload system (which includes
three cameras and a computer), avionics, two inertial sensors (a GPS module and an
inertial measurement unit), radio controller, and flight control (see Figure 4.2). The
aircraft is propelled using an electric, brushless motor. It does not require a runway and
Site G Site S
81
can be flown autonomously or manually. In autonomous mode, the aircraft follows a
preprogrammed flight plan containing navigation waypoints defined by GPS and altitude.
While operational, the payload computer instructs the three cameras to acquire imagery
in the VIS, NIR, and TIR spectra and records the position and orientation of the aircraft
when each image is taken. See Table 4.1 for specifications.
Figure 4.2. Details of AggieAir airframe.
Table 4.1
AggieAir Unmanned Aerial System Features Description
Specification Range
Flight duration 45–60 min
Flight altitudes 200–1000 m
Maximum takeoff weight 6.35 kg
Visible/near-infrared spectrum resolution 6–25 cm
Thermal resolution 30–150 cm
Wing span 2.5 m
82
Sensors Description
The VIS camera used in AggieAir is a Canon S-95, with a 10-megapixel CCD
sensor and an ISO range of 80 to 3200. The NIR camera is an identical Canon S-95,
modified by replacing the manufacturer’s optical filter with a Wratten 87 NIR filter that
allows NIR wavelengths of 750 nm. AggieAir also carries a small, low-power,
microbolometer thermal camera from Infrared Cameras, Inc. (2014). The corresponding
relative spectral responses is shown in Figure 4.3.
Figure 4.3. Relative quantum response of the VIS-NIR (left) and thermal camera (right). VIS-NIR = visible/near-infrared; ICI = Infrared Cameras, Inc.
Flight Details
On June 9, 2013 and August 9, 2014 Site S and G, respectively, were flown by
AggieAir. Flights occurred following the Landsat image acquisition protocol, close to
0
0.2
0.4
0.6
0.8
1
380 480 580 680 780 880
Wavebands (nm)
VIS-NIR Relative Spectral Response
0
0.2
0.4
0.6
0.8
1
1200 6200 11200 16200 21200
Wavebands (nm)
ICI Thermal Camera Relative Spectral Response
83
solar noon (between 11 a.m. and 1 p.m.). The flight time (beginning to end) ranged from
30–40 minutes. Sensors used were Canon S95 cameras and an ICI 9000 TIR camera.
Images were collected in the VIS-NIR spectrum at a spatial resolution of 0.15 cm and at
0.60 cm for the TIR. The UAV was flown at an altitude of 450 m acquiring a total of 165
images from Site S and 240 images from Site G.
Processing Chain of AggieAir
The process chain in AggieAir mapping consists of the following steps: mission
planning, image processing and orthorectification, reflectance calibration, and thermal
data calibration. Each is described in detail.
Mission Planning
The two flight missions were planned with AggieAir Mission Complete, a tool
built by the AggieAir team that uses the open source World Wind SDK from NASA
software. The flight routes were designed over the two sites by accounting for three
major parameters: flight area (site dimension), camera specifications (focal length), and
the desired overlap in the imagery. A longitudinal overlap of 70% and a lateral overlap of
60% have been adopted while flying at 450 m altitude. Once all this information and the
flight altitude were introduced in the module, it automatically generated the flight route
and estimated the flight duration according to the total number of images planned. The
route file is exported to a memory card embedded in the UAV via a standard serial link.
84
Image Processing and Orthorectification
Following image acquisition, all images are imported into AggieAir Mission
Complete for further cleaning of the flight lines. In this process all the images collected
outside the area of interest, captured before reaching the desired altitude, or captured in
an oblique manner are deleted. The remaining images are stretched and made ready to be
combined into a single image. The process of stitching the imagery and geometrically
correcting it to align with geospatial coordinates is called orthorectification. A
photogrammetric software named Agisoft PhotoScan was employed for this task. In
Agisoft, the software generates hundreds of tie-points between overlapping images by
using photogrammetric principles in conjunction with image GPS log file data to
automatically align and orthorectify the imagery. To further refine the estimate of the
image position and orientation of individual-image ground control points (GCPs), control
points located within the area of interest were surveyed with an RTK GPS and added to
the image for further accuracy in the geo-referencing procedure. The outputs from
Agisoft are a four-band (BGR, NIR) orthorectified mosaic with digital numbers, a TIR
orthorectified mosaic with brightness temperature, a digital terrain model derived by
creating a mesh from a simplified and filtered version of the point cloud, digital surface
model, a high-density point cloud that makes details easily recognizable, and a detailed
summary report. An example of the summary report is shown in Figure 4.4, which
demonstrates camera positioning and overlapping in addition to other details related to
the software processing algorithm.
85
Figure 4.4. An example of the summary report, the number of overlapped images range from 1 to 9 per frame. The black dots represent the center of the frame collected over the four flight lines.
Radiometric Calibration
Radiometric calibration is a process of converting digital numbers into a measure
of reflectance. The radiometric quality of the images is critical in order to enable the
application of quantitative remote sensing methodologies for a successful estimation of
biophysical parameters from remote sensing imagery (Berni et al., 2009). AggieAir
follows the reflectance mode method conversion, which is based on methods adapted
from the research (Crowther, 1992; Miura & Huete, 2009; Neale & Crowther, 1994) and
discussed in detail by Clemens (2012). In this method the ratio of digital number mosaic
86
to the digital number from a Spectralon white reflectance panel is multiplied by the panel
reflectance factor, as shown in Equation 1. After learning that the new Labsphere is a
perfect Lambertian surface that does not need to account for sun angles or time of day,
weighted averages over the range of each band in the sensor were calculated using
Labsphere’s reflectance factor report. An updated reflectance factor was calculated, and
the product of this method was an orthorectified mosaic in reflectance values.
(1)
where: DNT(t): digital numbers in mosaic at the time t; DNR(t): digital number from
panel at the time t; RR is the reflectance factor of the reference panel with respect to a
Lambertian surface of unit reflectance
Thermal Calibration
TIR data are measured as brightness temperature by the sensor embedded on the
AggieAir payload. After data collecting the imagery, these measurements are corrected
and calibrated to account for various errors. Errors are introduced in the measurement
from the convection across the lens during the course of the flight. To develop the
correction equation, GCPs were collected simultaneously with the flights over the two
sites. The field data collection procedure was designed to acquire three GCP temperature
measurements over various sampling surfaces (e.g., high vegetative crops, inter-row
crops, bare soil, top of the vines canopy, etc.). For every GCP frame (a) TIR imagery is
collected by AggieAir at an elevation of 450 m on average, (b) TIR imagery is collected
by an identical sensor mounted on a moving vehicle from an elevation of 3 m, and (c) an
RR
TT R
tDNtDNR)()(
87
ambient absolute temperature is measured by a Blackbody instrument placed on the GCP.
Figure 4.5 is an example of two measurements collected.
Figure 4.5. An example of a ground control point (GCP), high vegetation, needed for calibration of thermal infrared (TIR) data. A picture of the GCP is captured right, a TIR imagery from a 3 m elevation in the middle, and the Blackbody reading on the left.
Since the Blackbody emissivity is 0.95 and not 1.0 like the TIR sensors,
adjustment to the Blackbody reading was necessary. Equation 2 describes the correction
done on the Blackbody readings.
Tbb_corr = Tbb*(1/BB emissivity) (1/4) (2)
Where: Tbb_corr = corrected temperature of the blackbody; Tbb = reading of the blackbody;
BB emissivity = Blackbody emissivity. The collected data were used to develop a
correction equation used to produce radiometric temperature.
Models: METRIC and METRIC High Resolution (METRIC-HR)
METRIC was selected as the base model to investigate the capability of AggieAir
data in estimating ET because it has been tested with operational applications in Idaho,
California, and Colorado and because of the researcher’s knowledge and expertise in this
model. A brief description of the original METRIC algorithm and the modified version
88
called METRIC-HR is provided below. A more detailed description of METRIC can be
found in the applications manual (Allen, Tasumi, Trezza, & Kjaersgaard, 2010). The
version of METRIC used in this study is METRIC v 2.0.8 (developed March 2012)
METRIC
METRIC algorithm is a satellite image processing model whose approach is
based on the rationale that ET is a change of the state of water using available energy in
the environment for vaporization (Su, McCabe, Wood, Su, & Prueger, 2005). This energy
is partitioned into net incoming radiation, ground heat flux, sensible heat flux, and latent
heat flux. METRIC estimates spatially distributed values of actual ET as the residual of
the energy balance (Allen et al., 2007):
LE =Rn − G − H (3)
where
LE = heat flux density (W m-2);
Rn = incoming radiation flux density (W m-2) and is calculated by solving the
radiation balance as described by (Allen, Tasumi, & Trezza, 2002)
G = soil heat flux density (W m-2) and is estimated as a function of sensible
heat flux (H), net incoming radiation Rn, surface temperature, and the
normalized difference vegetation index (NDVI) and LAI.
H = sensible heat flux density (W m-2) and is estimated as a function of air
density (kg m-3); air specific heat (J kg-1 K-1); temperature difference between
two heights; the aerodynamic resistance to heat transport (s m-1); and
89
temperature gradient between two near-surface air temperatures (K), generally
approximated at 0.1 m and 2 m above the canopy (Bastiaanssen et al., 1998).
The sensible heat flux is considered to be the most difficult term in the energy
balance equation to calculate using remote sensing; thus, computing H might require
multiple iterations. A strong linear relationship exists between dT and the radiometric
surface temperature (Allen, Tasumi, & Trezza, 2007; Basstiaannssen et al., 1998, 2005;
Jacob et al., 2002a). The relationship can be expressed in Equation 4:
𝑑𝑇 = 𝑎 + 𝑏𝑇𝑠 (4)
where a and b are empirically derived parameters based on two extreme conditions
pixels, termed hot and cold pixels. These anchor pixels define the upper and lower
bounds of the sensible heat flux in the study area and thus the gradient. The cold pixels
should represent well-watered and fully vegetated areas of the image, where no water
stress is present such that H is assumed to be minimal and ET maximum or near
maximum. As for the hot pixel, it should be located in a dry and bare agricultural field
where the evaporative flux is almost zero, thus H dominating the turbulent fluxes. Once
surface temperature, Ts, and dT are calculated corresponding to hot and cold conditions,
the linear relationship as indicated in Equation 4 is defined. To effectively select
reasonable cold and hot anchor pixels, the user must be skillful and understand the
principles associated with the energy balance. For further details, refer to (Allen, Tasumi,
Trezza, & Kjaersgaard, 2010).
90
METRIC-HR
Since METRIC is designed to use Landsat 8 imagery as inputs, a few adjustments
were made to accommodate the high-resolution AggieAir input data. These modifications
are described below.
Digital elevation model (DEM) and DEM high resolution (DEM-HR). To
account for the various topography of a site, METRIC incorporated a DEM. This DEM is
used to adjust surface temperatures for lapse effects caused by elevation variation as well
as in estimating solar radiation on slopes. The 30 m resolution DEM used in the original
METRIC is downloadable from the U.S. Geological Survey website. In METRIC-HR a
DEM-HR of 15 cm replaced the original coarse DEM. The DEM-HR was generated with
Agisoft software extension tool. Note that the two versions of the models use DEM term
to refer to the digital terrain model.
National Land Cover Database (NLCD) and NLCD high resolution. Another
basic input file needed to run METRIC is land use maps. The NLCD maps hold land
cover information. Those maps are also used to support the estimation of aerodynamic
roughness and soil heat flux during METRIC processing. In METRIC-HR a high-
resolution NLCD of 15 cm replaced the original coarse NLCD. A supervised
classification was performed on the NDVI maps to identify vegetation, roads, urban
infrastructure, and other features present in the high-resolution imagery. The obtained
classes were used to develop the NLCD high-resolution maps.
Shortwave infrared (SWIR) bands. METRIC requires the usage of the SWIR
band acquired by Landsat 8. SWIR is used to calculate the normalized difference water
91
index primarily to identify water, for it tends to give values less than zero for water and
snow and is therefore a more reliable water indicator than NDVI. However, water and
snow do not exist in either study site, and therefore the absence of SWIR has no effect on
the processed model. To ensure the model ran smoothly a pseudo null band replaced the
SWIR in METRIC-HR. In addition, METRIC incorporated the SWIR bands in the albedo
estimations; this was neglected by METRIC-HR, for albedo was estimated solely over
the available four bands: BGR and NIR.
Thermal band resampling. The thermal band acquired by AggieAir is of 60 cm
spatial resolution. METRIC requires all bands used in the model to have identical
resolution; therefore, resampling the TIR data was necessary. The resampling was
performed using the nearest neighbor resampling method in ESRI’s ArcGIS 10.2 ArcMap
software. Nearest neighbor resampling is a technique of raster data resampling where
each cell in an output raster is computed using the values of the nearest pixel in the input
image. This particular approach was selected as it does not alter the original pixel’s value
(Duggin & Robinove, 1990; Lillesand & Kiefer, 2000).
High-resolution VIS data. Landsat 8 shortwave radiance imagery (BGR) is
replaced by high-resolution shortwave in METRIC-HR. However, AggieAir sensor
acquires the BGR reflectance’s with a consumer-grade sensor whose spectral relative
response is different from Landsat 8, as shown in Figure 4.6.
Comparing the AggieAir and Landsat reflectances in the BGR spectrum revealed
that AggieAir reflectances were relatively higher than Landsat’s with a consistent trend in
both imagery of Site S and Site G. Therefore, a correction to the AggieAir reflectance in
92
the BGR spectrum was developed as described below. This process included two steps:
(a) upscaling AggieAir BGR using Landsat 8 point spread function (PSF) and (b)
developing the individual correction equations.
Figure 4.6. Landsat 8 and AggieAir relative spectral response.
Upscaling AggieAir BGR using Landsat 8 PSF. Several studies noted that
upscaling using the PSF produces reliable upscale data (Goforth, 1998); hence, this
methodology was adopted. The PSF of an imaging system is a measure of the amount of
blurring that occurs due to all of the components that comprise the imaging system
(Wenny et al., 2015). In other words, PSF assumes that the spectral information in a pixel
does not originate solely from within its footprint; a substantial portion comes from
surrounding areas (Forster & Best, 1994; Huang, Townshend, Liang, Kalluri, & DeFries,
2002; Townshend & Tucker, 1981). PSF is computed as the square modulus of the field
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
350 450 550 650 750 850 950 1050 1150
Wavebands (nm)
Landsat 8 & AggieAir Relative Spectral Response
Blue_AggieAir
Green_AggieAir
Red_AggieAir
Landsat_ Blue
Landsat_Green
Landsat_Red
Landsat_NIR
NIR_AggieAir
93
amplitude on the focal plane; the amplitude function, in turn, is built through the
diffraction integral, derived from the wave description of electromagnetic radiation (Gai
& Cancelliere, 2007). Landsat 8 PSF was developed by the team at U.S. Geographical
Survey Earth Resources Observation Systems Data Center and NASA Goddard Space
Flight Center. Using the data they provided, Figure 4.7 represents the PSF of the three
bands (BGR).
Figure 4.7. Landsat 8 point spread function (PSF) along the blue, green, and red (BGR) bands (left); Landsat 2D PSF representation on the right.
A squared matrix of 210 x 210 m slides over the AggieAir imagery every 30 m,
multiplying each pixel by its corresponding weight according to the PSF. The upscaling
equation is described in Equation 5.
(5)
After upscaling the BGR bands to 30 m, a linear equation was identified that
depicts the linear relationship between the 30 m upscale AggieAir and Landsat 8. Table
-0.005
0
0.005
0.01
0.015
0.02
0.025
-100 -50 0 50 100
Res
pons
e
Distance (m)
PSF along BGR bands
RedbandBlueband
𝜌30𝑚 = ∑ 𝑃𝑆𝐹 ∗ 𝜌0.15𝑚
210𝑚
Landsat pixel
94
4.2 presents individual equations used to correct the 30 m upscale AggieAir. Note that the
correction was performed on the VIS (BGR) bands and not on the NIR band. Given the
nature of the long pass filter on the NIR band, no correction was performed or required.
Table 4.2
The Developed Correction Equations for Each Band
Albedo
Surface albedo is defined as the fraction of incident solar energy (diffuse and
direct components) reflected both in all directions above the surface and over the whole
solar spectrum (Jacob & Olioso, 2002; Jacob et al., 2002b; Pinty & Verstraete, 1992).
The in situ albedo is calculated as the ratio of reflected to incoming solar radiation
measurements; however, such data were not collected from either site to calculate albedo.
Estimating albedo from remote sensing imagery has its challenges, bidirectional
reflectance distribution function being one of them. This phenomenon may result in over-
or underestimating of albedo. An overestimate of albedo by 20% occurs when the sun
and sensor angles match, thus resulting in higher canopy reflectance (Liu et al., 2009).
Conversely, when solar angle is substantially different from the sensor view angle, albedo
can be less than hemispherical albedo (Allen et al., 2011b). The bidirectional reflectance
Band Site G Site S
Blue y = 0.62x - 0.06 (r2 = 0.95) y = 0.58x - 0.04 (r2 = 0.99)
Green y = 0.51x - 0.01 (r2 = 0.89) y = 0.47x + 0.01 (r2 = 0.92)
Red y = 0.39x + 0.02 (r2 = 0.90) y = 0.42x + 0.02 (r2 = 0.94)
95
distribution function is corrected for MODIS-based albedo retrievals but not for Landsat
(Salomon, Schaaf, Strahler, Gao, & Jin, 2006). Acknowledging this phenomenon,
METRIC determines the albedo as a weighted average of the shortwave bands based on
the percentage of total at-surface radiation occurring with each band (Tasumi, Allen, &
Trezza, 2008). METRIC weighting coefficients proposed are for low-haze atmospheric
conditions and are optimized for Landsat images. Equation 6 represents the weighting
coefficients associated with the Landsat bands for calculating albedo (Tasumi et al.,
2008).
𝛼𝑠h𝑜𝑟𝑡=0.254𝛼1+0.149𝛼2+0.147𝛼3+0.311𝛼4+0.103𝛼5+0.036𝛼6 (6)
where αshort is the albedo and αn is the at-surface reflectance calculated for each Landsat
shortwave band.
Knowing the importance of the albedo values in selecting the extreme pixels
needed to properly run the model, customized albedo coefficients were developed. A
linear relationship between Landsat albedo and AggieAir reflectances was established as
shown in Equation 7.
NIRredgreenblueshort 6555.0538.0744.0716.0 (7)
Equation 7 was developed using the total number of 30 m pixels in the imagery
from both sites. These pixels included vines, vegetation, dry soil, and wet soil pixels. The
linear regression showed a residual error of root mean square error of 0.007 and an R2 of
0.88. Figure 4.8 shows a 1:1 plot of the albedos from Landsat and AggieAir over the two
sites. Note that the graph has two clusters visually; the lower left cluster of pixels form
Site S, and the upper right cluster are the pixels from Site G.
96
In theory the sum of coefficients should have a value close to 1, and coefficients
should be positive. However, when comparing the derived coefficients with Landsat
Figure 4.8. 1:1 Plot of the albedo from AggieAir and Landsat.
coefficients in the green band, a change of sign was noticed. Nevertheless, a
similar negative response in the green band was also reported in the study by Jacob and
Olioso (2002) that derived albedo coefficient from airborne platforms.
Results and Discussion
Hot and Cold Pixel Selection
Identifying the hot and cold pixels requires expertise and time to select a reliable
representation of the anchor pixels. Various sensitivity analyses reported that selecting
different hot and cold pixels leads to large deviations in final ET estimates (Wang et al.,
2009). The researchers properly select the hot and cold pixels that satisfy the assumptions
97
made in METRIC, such that the linear correlation between the near surface temperature
difference and remotely sensed surface temperature holds true.
METRIC recommends for the cold pixel to be located in a homogenous well-
watered, full-cover crop in the image with an NDVI range of 0.76–0.84 and a surface
albedo range of 0.18–0.24. As for the hot pixel, METRIC recommends that the hot
anchor pixel should be selected in homogenous bare and dry or nearly dry agricultural
soil with little or no vegetation, with an NDVI range no higher than 0.2 and a surface
albedo range of 0.17–0.23. Further details on the selection of the two pixels are described
by Allen, Tasumi, Morese, et al. (2007) and Allen (2008).
All four anchor points were selected from the perimeters of a common
overlapping area. The values of the NDVI, albedo, LAI, surface temperature, and the
coefficients for Equation 4 are reported in Table 4.3. Note that the sensitivity of the
selection of the hot/cold pixels in this study wasn’t performed.
The hot pixel in Site S, planted with alfalfa mainly, was selected from the clear,
bare soil area inside the northern center pivot. As for the cold pixel, more than one
candidate pixel were considered, for the crop was homogenous with full land cover in
various locations in the imagery. The cold pixel was selected from the northern part of the
southern center pivot. Both anchor pixels lay in the METRIC recommended ranges of
NDVI and albedos.
However, Site G, dominated with vineyard canopies, presented different
challenges in the selection process. Challenges were closely linked to the specific
geometrical canopy structure and the radiation balance pattern. The vineyard’s unique
98
characteristics are standing grape canopy, variable shaded areas, wide spacing between
rows, discontinuous soil cover, and various vertical leaf thickness.
Table 4.3
Details on the Selected Hot and Cold Pixels from the AggieAir and Landsat Imagery
Source Albedo NDVI LAI Ts a b
Site S AggieAir
Hot pixel 0.264 0.254 0.149 319.983 0.14270 -39.7412 Cold pixel 0.233 0.869 5.345 308.406
Landsat
Hot pixel 0.177 0.277 0.824 327.481 0.1176 -31.9722
Cold pixel 0.229 0.868 6.000 308.489
Site G AggieAir
Hot pixel 0.320 0.370 0.444 332.395 0.19919 -63.3236 Cold pixel 0.238 0.910 6.000 300.355
Landsat
Hot pixel 0.164 0.174 0.038 322.885 0.25852 -78.2922
Cold pixel 0.226 0.838 5.338 305.076 Note. LAI = leaf area index; NDVI = normalized difference vegetation index.
All these variables create complex dynamics that generate a spatial heterogeneity
in the horizontal distribution of energy at the soil surface; thus, finding a homogenous area
that fulfills the METRIC recommendation was a challenge to the end user. The albedo and
NDVI for the hot pixel exceeded the recommendations, possibly due to the lack of a
prominent bare soil area within the vineyard. The hot pixel was chosen from the northwest
side of the vineyard, which was distinguishable by the leafless vines, part of a separate
yield study.
99
METRIC Results
Landsat 8 cloud-free satellite images were obtained from the U.S. Geological
Survey Earth Explorer site (http://earthexplorer.usgs.gov) for June 9, 2013 and August 9,
2014. The images were processed using the ERDAS Imagine 2014 software. METRIC
was applied on the corresponding Landsat imagery of Site G and Site S to obtain
instantaneous and daily reference ET fraction (ETrf) maps. Maps of reflectance of
shortwave radiation, vegetation indices (NDVI and LAI), surface temperature, net
radiation, and soil heat flux were generated as intermediate products during METRIC
processing. The final outputs from the METRIC energy balance model were images
showing instantaneous ETrf (fraction of alfalfa-based reference ET) at the satellite
overpass time. Instantaneous ETrf produced after running METRIC over Site G and Site
S are shown in Figures 4.9 and 4.10, respectively.
100
Figure 4.9. Site S: Landsat reflectance in false color (left), Mapping Evapotranspiration with Internalized Calibration (METRIC) reference evapotranspiration fraction (EtrF) output (right).
KcHigh : 1.15
Low : 0
00.1
0.20.05
Miles
101
Figure 4.10. Site G: Landsat reflectance bands in true color (left), Mapping Evapotranspiration with Internalized Calibration (METRIC) reference evapotranspiration fraction (ETrf) output (right).
METRIC-HR Results
METRIC-HR was run with the AggieAir derived inputs including corrected
reflectance (BGR, NIR), TIR band, DEM-HR, NLCD high resolution, albedo, and NDVI.
The resulting instantaneous ETrf images are shown in Figures 4.11 and 4.12.
KcHigh : 1.15
Low : 0
00.1
0.20.05
Miles
102
Figure 4.11. Site S: AggieAir reflectance in false color (left), Mapping Evapotranspiration with Internalized Calibration (METRIC) reference evapotranspiration fraction (ETrf) output (center), TIR (Celsius) map (right).
In figure 4.11 it is clear that the “crop coefficient” Kc, also known as “reference
ET fraction” ETrF, follows a similar spatial pattern as its corresponding false color and
TIR imagery. The values of the crop coefficient ranged between 0 and 1.15, with patterns
of ET linked to canopy temperature and cover. Lower values of Kc corresponded to
hotter areas where bare soil is mainly found, particularly in the outer surroundings of the
two center pivots and in the unplanted quarters in the northern center pivot. Higher values
of Kc are prominent in wet areas (one can see the pivot arm in the northern plot). Areas
of cooler temperatures where crops are homogenous and crop cover is dense have higher
Kc values.
00.1
0.20.05
Miles
103
Figure 4.12. Site G: AggieAir reflectance bands in true color (left), Mapping Evapotranspiration with Internalized Calibration (METRIC) reference evapotranspiration fraction (ETrf) output (center) TIR (Celsius) map (right).
Figure 4.12 represents the Kc estimates from site G derived from the AggieAir
imagery. Considering that the vines planted in site G are all of the same type (Pinot noir),
similar growth stage and irrigated in a similar fashion (drip irrigation), there still exist a
variation in crop water demand within the field. This variation could be explained by the
different soil types present in the vineyard.
00.1
0.20.05
Miles
104
Comparison in Site S
Site S is irrigated by two modern, center-pivot sprinkler irrigation systems with a
capacity of 610 GPM each. A local upstream reservoir feed these two pivots regularly.
The pivots are automated to spray water every 15 degrees of rotation, creating 24 distinct
manageable areas in each center pivot as shown in Figure 4.13. The arcs are labeled from
1–48 covering the two center pivots.
Figure 4.13. Site S: Mapping Evapotranspiration With Internalized Calibration high resolution (METRIC-HR) reference evapotranspiration fraction (ETrf) on left, METRIC ETrf on right, both showing the 48 manageable arcs as defined in this study.
9
1
7
43
56
2
8
11
13
19
23
16
20
22 10
15
17
1224
18
14
21
363738 3539 343340
41 32
42 313043
44 2945 28
2746 47 262548
9
1
7
43
56
2
8
11
13
19
23
16
20
22 10
15
17
1224
18
14
21
363738 3539 343340
41 32
42 313043
44 2945 28
2746 47 262548
105
ETrf estimates were averaged over each sector. The obtained measurements from
the northern center pivot are shown in Figure 4.14.
Figure 4.14. Northern center pivot reference evapotranspiration fraction (ETrf) estimates from Landsat and AggieAir inputs.
. METRIC-HR resulted in a higher ETtrf estimated, except for Sectors 20 and 21.
This could be partly explained by the presence of pixels of multiple vegetation growth
with significant differences in cover (2 m and 15 cm alfalfa crops); a variation of surface
roughness is most prominent in these two sectors. However, identical results were
obtained from Sectors 1 and 13. These two sectors are the wettest sectors in the northern
center pivot, where the spraying arm could be seen in the high-resolution imagery.
0.00
0.20
0.40
0.60
0.80
1.00
1.20
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Etrf
Sector
Northern Center Pivot
Landsat AggieAir
106
Sectors 1 and 13 are also surrounded with vegetation outside the center pivot, resulting in
an extended homogenous land cover. The maximum difference in estimates occurs in
sector 7 (20% difference).
A similar analysis was performed on the southern center pivot. The averaged ETrf
over the sectors were plotted from the two tested models, as shown in Figure 4.15.
Figure 4.15. Southern center pivot reference evapotranspiration fraction (ETrf) estimates from Landsat and AggieAir inputs.
In the southern center pivot a more homogenous crop cover was observed. ETrf
data derived from AggieAir imagery estimated higher ETrf across all sectors of the center
0.00
0.20
0.40
0.60
0.80
1.00
1.20
25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48
Etrf
Sector
Southern Center Pivot
Landsat AggieAir
107
pivot. This could be explained by the surrounding dry landscape of the center pivot; these
peripheral mixed pixels lowered the average ET estimates in the sectors. The averaging in
each sector included the field edges where pixel contamination can cause METRIC ET
estimates to deviate from the field average (Allen, Tasumi, & Trezza, 2007).
A comparison between ETrf estimates derived from Landsat and AggieAir data is
presented in Figure 4.16. These values were close and had a 0.90 coefficient of
correlation.
Figure 4.16. Site S: Comparison between reference evapotranspiration fraction (ETrf) estimates derived from Landsat and AggieAir data in the northern and southern pivots
ETrf estimates obtained from the both the high resolution imagery and the
Landsat imagery, showed a high correlation (0.9) in the two center pivots. This implies
0.50
0.60
0.70
0.80
0.90
1.00
1.10
0.50 0.60 0.70 0.80 0.90 1.00 1.10
Land
sat
ETrf
AggieAir ETrF
Landsat vs AggieAir ETrf
Northern Center Pivot Southern Center Pivot Linear (Northern Center Pivot )
108
that both the METRIC- HR and METRIC models showed similar performance
capabilities. The high resolution imagery was superior in the mixed pixel areas (bare soil
and vegetation) specifically in the pixels in the outer contour of the center pivot. The
northern center pivot in this site study had a more heterogeneous surface compared to the
southern pivot. As a result, with respect to the northern pivot, a high resolution ET
estimates would be more beneficiary compared to the southern pivot.
Comparison in Site G
Site G, a well-maintained vineyard, is irrigated by drip system. The irrigation
design divides the vineyards of groups of 30–40 rows on average, each group creating a
block. There are 19 blocks in the vineyard. Each block can be regulated independently
with a ball valve upstream of the pressure regulator. The block is considered as the
smallest manageable area in this study. The distribution of the blocks is visualized in
Figure 4.17 as well as the ETrf estimates of AggieAir and Landsat.
To test the results of the spatially distributed ETrf estimates, all the pixels lying
within the blocks were averaged. The average represented a single water demand that
needs to be met by the water applicators.
The ETrf estimates obtained from both models are shown in Figure 4.18.
Visually, the Landsat ETrf estimate has a more fuzzy appearance. Compared to AggieAir
results, Landsat was underestimating ETrf in 12 of the 19 blocks. The comparison of the
ETrf estimates generated by both METRIC and METRIC-HR showed the largest
differences in Block 19. All the vegetative features in Block 19 are masked by the large
Landsat pixel footprint; visually it is not recognizable as a vegetative area.
109
Figure 4.17. Mapping Evapotranspiration With Internalized Calibration (METRIC) reference evapotranspiration fraction (ETrf) on left, METRIC high resolution ETrf in middle, and 19 blocks of individual manageable areas on right.
The high-resolution AggieAir imagery enabled the METRIC-HR to sense
evaporation occurring from the smallest block (Block 19), where METRIC could not.
19
1
2
3
4
5
6
7
8
9
10
11
12 13 14
15
16
17
18
110
The two estimates showed a correlation coefficient of .73. Figure 4.19 shows the two
estimates plotted against each other.
Figure 4.18. The average reference evapotranspiration fraction (ETrf) estimates from Landsat and AggieAir inputs for each block.
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
1.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
ETrf
Blocks
Sits G blocks
AggieAir
Landsat
111
Figure 4.19. Site G: Comparison between reference evapotranspiration fraction (ETrf) estimates derived from Landsat and AggieAir data.
Conclusion
The objective of this study was to assess the use of AggieAir, an unmanned aerial
system, to estimate crop evapotranspiration at high spatial resolution. A high resolution
METRIC (METRIC-HR) was derived from the well-established METRIC algorithm.
High resolution inputs (RGB, NIR, TIR, DEM, NLCD) and weather data were used to
develop an ET estimates at high resolution (0.15 m). A significant amount of spatial
information was retrieved and detailed ET estimates was established over two sites S and
G. This work demonstrated that it is possible to generate quantitative remote sensing
products by means of a UAV equipped with consumer-grade cameras. However, these
consumer-grade cameras require extensive calibration to relate spectral response to
another scientific sensor (such as Landsat). The high spatial resolution provided make
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.6 0.65 0.7 0.75 0.8 0.85 0.9
Land
sat E
Trf
AggieAir ETrf
Landsat vs AggieAir ETrf
112
this platform particularly suitable for precision agriculture and irrigation scheduling,
where site specific critical management is required. The High resolution ET showed is a
useful tool to monitor crop growth, and crop water demands when managing
heterogeneous surfaces. This model lays the ground for the estimation of ET at high
spatial resolution to be used in precision agriculture. The high resolution spatial
distribution of ET helped evaluating the efficiency of irrigation applications per the
smallest manageable area.
References
Allen, R. G. (2008). REF-ET: Reference evapotranspiration calculation software for FAO and ASCE standardized equations. Kimberly, ID: University of Idaho.
Allen, R. G., Pereira, L. S., Howell, T. A., & Jensen, M. E. (2011a). Evapotranspiration information reporting: I. Factors governing measurement accuracy. Agricultural Water Management, 98, 899-920. doi:10.1016/j.agwat.2010.12.015, 2011a.
Allen, R. G., Pereira, L. S., Howell, T. A., & Jensen, M. E. (2011). Evapotranspiration information reporting: II. Recommended documentation. Agricultural Water Management, 98, 921-929. doi:10.1016/j.agwat.2010.12.016, 2011b
Allen, R. G., Pereira, L. S., Raes, D., & Smith, M. (1998). Crop evapotranspiration—Guidelines for computing crop water requirements (FAO irrigation and drainage paper No. 56). Rome, Italy: Food and Agriculture Organization of the United Nations.
Allen, R. G., Tasumi, M., Morse, A., Trezza, R., Wright, J. L., Bastiaanssen, W., . . . & Robison, C. W. (2007). Satellite-based energy balance for Mapping Evapotranspiration With Internalized Calibration (METRIC)—Applications. Journal of Irrigation and Drainage Engineering, 133, 395-406. http://dx.doi.org /10.1061/(ASCE)0733-9437(2007)133:4(395)
Allen, R. G., Tasumi, M., & Trezza, R. (2002). METRICTM: mapping evapotranspiration at high resolution—Application manual for Landsat satellite imagery. Kimberly, ID: University of Idaho.
113
Allen, R. G., Tasumi, M., & Trezza, R. (2007). Satellite-based energy balance for Mapping Evapotranspiration With Internalized Calibration (METRIC)—Model. Journal of Irrigation and Drainage Engineering, 133, 380-394. doi:10.1061 /(ASCE)0733-9437(2007)133:4(380)
Allen, R. G., Tasumi, M., Trezza, R., Kjaersgaard, J. H. (2010). METRIC: Mapping evapotranspiration at high resolution. Applications manual, V. 2.0.4. Kimberly: University of Idaho.
Arnó, J., Martínez Casanovas, J., Ribes Dasi, M., & Rosell, J. R. (2009). Review. Precision viticulture. Research topics, challenges and opportunities in site-specific vineyard management. Spanish Journal of Agricultural Research, 7, 779-790.
Baluja, J., Diago, M. P., Balda, P., Zorer, R., Meggio, F., Morales, F., & Tardaguila, J. (2012). Assessment of vineyard water status variability by thermal and multispectral imagery using an unmanned aerial vehicle (UAV). Irrigation Science, 30, 511-522.
Barnes, E. M., Pinter, P. J., Jr., Kimball, B. A., Hunsaker, D. J., Wall, G. W., & LaMorte, R. L. (2000). Precision irrigation management using modeling and remote sensing approaches. In National Irrigation Symposium, Proceedings of the Fourth Decennial Symposium, Phoenix, AZ, ASAE (pp. 332-337).
Bastiaanssen, W. G. M., Menenti, M., Feddes, R. A., & Holtslag, A. A. M. (1998). A remote sensing Surface Energy Balance Algorithm for Land (SEBAL). 1. Formulation. Journal of Hydrology, 212-213, 198–212
Bastiaanssen, W. G. M., Noordman, E. J. M., Pelgrum, H., Davids, G., Thoreson, B. P., & Allen, R. G. (2005). SEBAL model with remotely sensed data to improve water-resources management under actual field conditions. Journal of Irrigation and Drainage Engineering, 131, 85-93.
Berni, J. A. J., Zarco-Tejada, P. J., Suárez, L., González-Dugo, V., & Fereres, E. (2009). Remote sensing of vegetation from UAV platforms using lightweight multispectral and thermal imaging sensors. International Archive of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 38(6).
Boegh, E., Poulsen, R. N., Butts, M., Abrahamsen, P., Dellwik, E., Hansen, S., . . . & Søgaard, H. (2009). Remote sensing based evapotranspiration and runoff modeling of agricultural, forest and urban flux sites in Denmark: From field to macro-scale. Journal of Hydrology, 377, 300-316.
Bouilloud, L., Chancibault, K., Vincendon, B., Ducrocq, V., Habets, F., Saulnier, G. M., . . . & Noilhan, J. (2010). Coupling the ISBA land surface model and the TOPMODEL hydrological model for Mediterranean flash-flood forecasting:
114
description, calibration, and validation. Journal of Hydrometeorology, 11, 315-333.
Bramley, R. G. V. (2010). Precision viticulture: Managing vineyard variability for improved quality outcomes. In A. G. Reynolds (Ed.), Understanding and managing wine quality and safety (pp. 445-480). Cambridge, England: Woodhead.
Calcagno, G., Mendicino, G., Monacelli, G., Senatore, A., & Versace, P. (2007). Distributed estimation of actual evapotranspiration through remote sensing techniques. In G. Rossi, T. Vega, & B. Bonaccorso (Eds.), Methods and tools for drought analysis and management (pp. 124-147). Dordrecht, The Netherlands: Springer.
Chávez, J., Neale, C. M., Hipps, L. E., Prueger, J. H., & Kustas, W. P. (2005). Comparing aircraft-based remotely sensed energy balance fluxes with eddy covariance tower data using heat flux source area functions. Journal of Hydrometeorology, 6, 923-940.
Chen, J. M., Chen, X., Ju, W., & Geng, X. (2005). Distributed hydrological model for mapping evapotranspiration using remote sensing inputs. Journal of Hydrology, 305, 15-39.
Chiabrando, F., Lingua, A., & Piras, M. (2013). Direct photogrammetry using UAV: Tests and first results. International Archive of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 41, 81-86. doi:10.5194/isprsarchives -XL-1-W2-81-2013
Clemens, S. R. (2012). Procedures for correcting digital camera imagery acquired by the AggieAir remote sensing platform (Doctoral dissertation). Retrieved from http:// digitalcommons.usu.edu/gradreports/186
Contreras, S., Jobbágy, E. G., Villagra, P. E., Nosetto, M. D., & Puigdefábregas, J. (2011). Remote sensing estimates of supplementary water consumption by arid ecosystems of central Argentina. Journal of Hydrology, 397, 10-22. doi:10.1016/j .jhydrol.2010.11.014, 2011
Courault, D., Seguin, B., & Olioso, A. (2005). Review on estimation of evapotranspiration from remote sensing data: From empirical to numerical modeling approaches. Irrigation and Drainage Systems, 19, 223-249.
Crowther, B. G. (1992). Radiometric calibration of multispectral video imagery. Doctoral dissertation, Utah State University, Logan.
115
Demarez, V., Duthoit, S., Baret, F., Weiss, M., & Dedieu, G. (2008). Estimation of leaf area and clumping indexes of crops with hemispherical photographs. Agricultural and Forest Meteorology, 148, 644-655.
Duggin, M. J., & Robinove, C. J. (1990). Assumptions implicit in remote sensing data acquisition and analysis. Remote Sensing, 11, 1669-1694.
Forster, B. C., & Best, P. (1994). Estimation of SPOT P-mode point spread function and derivation of a deconvolution filter. ISPRS Journal of Photogrammetry and Remote Sensing, 49, 32-42.
Gai, M., & Cancelliere, R. (2007). An efficient point spread function construction method. Monthly Notices of the Royal Astronomical Society, 377, 1337-1342.
Goforth, M. A. (1998). Fusion of differing resolution imagery using multiresolution analysis. In Proceedings of the Fusion International Conference (pp. 419-426).
Gómez, M., Olioso, A., Sobrino, J. A., & Jacob, F. (2005). Retrieval of evapotranspiration over the Alpille/ReSeDA experimental site using airborne POLDER sensor and a thermal camera. Remote Sensing of Environment, 96, 399-408.
Gowda, P. H., Chavez, J. L., Colaizzi, P. D., Evett, S. R., Howell, T. A., Tolk, J. A. (2007). ET mapping for agricultural water management: Present status and challenges. Irrigation Science, 26, 223-237.
Huang, C., Townshend, J. R., Liang, S., Kalluri, S. N., & DeFries, R. S. (2002). Impact of sensor’s point spread function on land cover characterization: Assessment and deconvolution. Remote Sensing of Environment, 80, 203-212.
Infrared Cameras, Inc. (2014). Home page. Retrieved from http://www .infraredcamerasinc.com
Jacob, F., & Olioso, A. (2002). Assessing recent algorithms devoted to albedo estimation using the airborne ReSeDa/PolDER database. In J. A. Sobrino (Ed.), Proceedings of the First International Symposium on Recent Advances in Quantitative Remote Sensing (pp. 191-198). València, Spain: Universitat de València.
Jacob, F., Olioso, A., Gu, X. F., Su, Z., & Seguin, B. (2002a). Mapping surface fluxes using airborne visible, near infrared, thermal infrared remote sensing data and a spatialized surface energy balance model. Agronomie, 22, 669-680.
Jacob, F., Olioso, A., Weiss, M., Baret, F., & Hautecoeur, O. (2002b). Mapping short-wave albedo of agricultural surfaces using airborne PolDER data. Remote Sensing of Environment, 80, 36-46.
116
Jiang, L., Islam, S., Guo, W., Jutla, A. S., Senarath, S. U., Ramsay, B. H., & Eltahir, E. (2009). A satellite-based daily actual evapotranspiration estimation algorithm over South Florida. Global and Planetary Change, 67(1), 62-77.
Johnson, L. F., Herwitz, S., Dunagan, S., Lobitz, B., Sullivan, D., & Slye, R. (2003). Collection of ultra-high spatial and spectral resolution image data over California vineyards with a small UAV. In Proceedings of the 30th International Symposium on Remote Sensing of Environment (Vol. 20, No. 845-849).
Kaheil, Y. H., Rosero, E., Gill, M. K., McKee, M., & Bastidas, L. A. (2008). Downscaling and forecasting of evapotranspiration using a synthetic model of wavelets and support vector machines. IEEE Transactions on Geoscience and Remote Sensing, 46, 2692-2707.
Kustas, W.P., French, A. N., Hatfield, J. L., Jackson, T. J., Moran, M. S., Rango, A., . . . & Schumgge, T. J. (2003). Remote sensing research in hydrometeorology. Photogrammetric Engineering and Remote Sensing, 69, 631-646.
Kustas, W. P., Li, F., Jackson, T. J., Prueger, J. H., MacPherson, J. I., & Wolde, M. (2004). Effects of remote sensing pixel resolution on modeled energy flux variability of croplands in Iowa. Remote Sensing of Environment, 92, 535-547.
Kustas, W. P., & Norman, J. M. (1996). Use of remote sensing for evapotranspiration monitoring over land surfaces. Hydrological Sciences Journal, 41, 495-516. doi:10.1080/02626669609491522
Li, Z.-L., Tang, R., Wan, Z., Bi, Y., Zhou, C., Tang, B., . . . & Zhang, X. (2009). A review of 20 current methodologies for regional evapotranspiration estimation from remotely sensed data. Sensors, 9, 3801-3853.
Lillesand, T. M., & Kiefer, R. W. (2000). Remote sensing and image interpretation. New York, NY: Wiley.
Liou, Y.-A., & Kar, S. K. (2014). Evapotranspiration estimation with remote sensing and various surface energy balance algorithms—A review. Energies, 7, 2821-2849. doi:10.3390/ en7052821
Liu, J., Schaaf, C., Strahler, A., Jiao, Z., Shuai, Y., Zhang, Q., … & Dutton, E. G. (2009). Validation of Moderate Resolution Imaging Spectroradiometer (MODIS) albedo retrieval algorithm: Dependence of albedo on solar zenith angle. Journal of Geophysical Research: Atmospheres (1984–2012), 114(D1).
Mauser, W., & Schädlich, S. (1998). Modelling the spatial distribution of evapotranspiration on different scales using remote sensing data. Journal of Hydrology, 212, 250-267.
117
McCabe, M. F., & Wood, E. F. (2006). Scale influences on the remote estimation of evapotranspiration using multiple satellite sensors. Remote Sensing of Environment, 105, 271-285.
Min, Q., & Lin, B. (2006). Remote sensing of evapotranspiration and carbon uptake at Harvard Forest. Remote Sensing of Environment, 100, 379-387.
Miura, T., & Huete, A. R. (2009). Performance of three reflectance calibration methods for airborne hyperspectral spectrometer data. Sensors, 9, 794-813.
Moran, M. S., Jackson, R. D., Raymond, L. H., Gay, L. W., & Slater, P. N. (1989). Mapping surface energy balance components by combining Landsat Thematic Mapper and ground-based meteorological data. Remote Sensing of Environment, 30, 77-87.
Mutiga, J. K., Su, Z., & Woldai, T. (2010). Using satellite remote sensing to assess evapotranspiration: Case study of the upper Ewaso Ng’iro North Basin, Kenya. International Journal of Applied Earth Observation and Geoinformation, 12, S100-S108. doi:10.1016/j.jag.2009.09.012
Neale, C. M., & Crowther, B. G. (1994). An airborne multispectral video/radiometer remote sensing system: Development and calibration. Remote Sensing of Environment, 49, 187-194.
Neale, C. M., Jayanthi, H., & Wright, J. L. (2003, September). Crop and irrigation water management using high resolution airborne remote sensing. Paper presented at the International Workshop on Use of Remote Sensing of Crop Evapotranspiration for Large Regions, 54th IEC Meeting of the International Commission on Irrigation and Drainage, Montpellier, France.
Nouri, H., Beecham, S., Kazemi, F., Hassanli, A. M., & Anderson, S. (2013). Remote sensing techniques for predicting evapotranspiration from mixed vegetated surfaces. Hydrology and Earth System Sciences Discussions, 10, 3897-3925.
Oki, T., & Kanae, S. (2006). Global hydrological cycles and world water resources. Science, 313(5790), 1068-1072.
Penman, H. L. (1948). Natural evaporation from open water, bare soil and grass. In Proceedings of the Royal Society of London A: Mathematical, Physical and Engineering Sciences (Vol. 193, No. 1032, pp. 120-145). London, England: The Royal Society.
Petropoulos, G., Carlson, T. N., Wooster, M. J. & Islam, S. (2009). A review of Ts/VI remote sensing based methods for the retrieval of land surface energy fluxes and soil surface moisture. Progress in Physical Geography, 33, 224-250.
118
Pinter, J., Hatfield, J., Schepers, J. S., Barnes, E. M., Moran, M. S., Daughtry, C. S. T., & Upchurch, D. R. (2003). Remote sensing for crop management. Photogrammetric Engineering & Remote Sensing, 69, 647-664.
Pinty, B., & Verstraete, M. (1992). On the design and validation of surface bidirectional reflectance and albedo model. Remote Sensing of Environment, 41, 155-167.
Rana, G., & Katerji, N. (2000). Measurement and estimation of actual evapotranspiration in the field under Mediterranean climate: A review. European Journal of Agronomy, 13, 125-153. doi:10.1016/s1161-5 0301(00)00070-8
Sadler, E. J., Evans, R., Stone, K. C., & Camp, C. R. (2005). Opportunities for conservation with precision irrigation. Journal of Soil and Water Conservation, 60, 371-378.
Salomon, J. G., Schaaf, C. B., Strahler, A. H., Gao, F., & Jin, Y. (2006). Validation of the MODIS bidirectional reflectance distribution function and albedo retrievals using combined observations from the aqua and terra platforms. IEEE Transactions on Geoscience and Remote Sensing, 44, 1555-1565.
Santos, C., Lorite, I. J., Tasumi, M., Allen, R. G., & Fereres, E. (2008). Integrating satellite-based evapotranspiration with simulation models for irrigation management at the scheme level. Irrigation Science, 26, 277-288.
Sheffield, J., & Wood, E. F. (2008). Projected changes in drought occurrence under future global warming from multi-model, multi-scenario, IPCC AR4 simulations. Climate Dynamics, 31(1), 79-105.
Stisen, S., Sandholt, I., Nørgaard, A., Fensholt, R., & Jensen, K. H. (2008). Combining the triangle method with thermal inertia to estimate regional evapotranspiration—Applied to MSG-SEVIRI data in the Senegal River basin. Remote Sensing of Environment, 112, 1242-1255. doi:10.1016/j.rse.2007.08.013
Su, H., McCabe, M. F., Wood, E. F., Su, Z., & Prueger, J. H. (2005). Modeling evapotranspiration during SMACEX: Comparing two approaches for local- and regional-scale prediction. Journal of Hydrometerology, 6, 910-922.
Tasumi, M., Allen, R. G., & Trezza, R. (2008). At-surface reflectance and albedo from satellite for operational calculation of land surface energy balance. Journal of Hydrologic Engineering, 13, 51-63.
Tasumi, M., Trezza, R., Allen, R. G., & Wright, J. L. (2005). Operational aspects of satellite-based energy balance models for irrigated crops in the semi-arid U.S. Irrigation and Drainage Systems, 19, 355-376.
119
Thevs, N., Peng, H., Rozi, A., Zerbe, S., & Abdusalih, N. (2015). Water allocation and water consumption of irrigated agriculture and natural vegetation in the Aksu-Tarim river basin, Xinjiang, China. Journal of Arid Environments, 112, 87-97.
Townshend, J. R. G., & Tucker, C. J. (1981). Utility of AVHRR NOAA-6 and -7 for vegetation mapping. In Proceedings of the International Conference on Matching Remote Sensing Technologies and Their Applications (London/Nottingham: Remote Sensing Society (pp. 97-109).
Van der Tol, C., & Parodi, G. N. (2012). Guidelines for remote sensing of evapotranspiration. In A. Irmak (Ed.), Evapotranspiration—Remote sensing and modeling (Ch. 11). Rijeka, Croatia: InTech Open Access. doi:10.5772/18582
Wang, J., Sammis, T. W., Gutschick, V., Gebremichael, M., & Miller, D. R. (2009). Sensitivity analysis of the Surface Energy Balance Algorithm for Land (SEBAL). Transactions of the ASABE, 52, 801-811.
Wenny, B. N., Helder, D., Hong, J., Leigh, L., Thome, K. J., & Reuter, D. (2015). Pre-and post-launch spatial quality of the Landsat 8 thermal infrared sensor. Remote Sensing, 7, 1962-1980.
Wu, C., Cheng, C., Lo, H., & Chen, Y. (2010). Study on estimating the evapotranspiration cover 25 coefficients for stream flow simulation through remote sensing techniques. International Journal of Applied Earth Observation and Geoinformation, 12, 225-232. doi: 10.1016/j.jag.2010.03.001
Zarco-Tejada, P. J., Berni, J. A., Suárez, L., & Fereres, E. (2008, December 17). A new era in remote sensing of crops with unmanned robots. SPIE Newsroom. doi:10 .1117/2.1200812.1438
Zhang, Y., & Wegehenkel, M. (2006). Integration of MODIS data into a simple model for the spatial distributed simulation of soil water content and evapotranspiration. Remote Sensing of Environment, 104, 393-408.
Zhou, X. D., Zhu, Q. J., & Sun, Z. P. (2002). Preliminary study on regionalization desertification climate in China. Journal of Natural Disasters, 11, 125-131.
120
CHAPTER 5
SUMMARY
This dissertation has shown how data from consumer-grade sensors acquired by
an unmanned aerial system (UAS), AggieAir, could be used in precision agriculture
application. The high spatial, spectral, and temporal resolution makes this platform
particularly suitable for precision agriculture and irrigation scheduling, where time-
critical management is required. Chapter 2 presented the retrieval of chlorophyll content
from thermal and multispectral optical imagery. A complex statistical regression model
(Bayesian relevance vector machine) was trained on a dataset of in situ collected leaf
chlorophyll measurements, and the machine learning algorithm intelligently selected the
most appropriate bands and indices for building regressions with the highest prediction
accuracy. Chapter 3, with a similar methodology to Chapter 2, discussed the estimation of
leaf nitrogen content. The predicted estimates were averaged over the smallest
manageable unit in the center pivot, thus providing a preliminary actionable information
for nutrient management. Chapter 4 investigated the possibility of mapping
evapotranspiration using the high spatial resolution data. Appropriate configurations were
applied on the Mapping Evapotranspiration With Internalized Calibration algorithm to
match the input data. Evapotranspiration estimates from AggieAir inputs were compared
to those obtained from Landsat 8 data and showed agreement between the two. The
estimates were averaged over the smallest irrigateable unit in the two study sites to assist
in a more precise efficient irrigation scheduling. All these estimates, made at such fine
resolutions in space and time, can aid farmers in assessing the heterogeneity of their
121
fields and subsequently implement needed actions accordingly. The high-resolution
spatial information generated from AggieAir imagery could enable far greater precision
in the application of various production resources (fertilizers, irrigation water). UAS
facilitate enhanced monitoring with time-critical account of fine-scale variations in plant
health and function.
The final findings presented in this dissertation are in general promising and
encourage the development of future improvements. Plenty could be done to improve the
collection, processing, and quality of the data. Perhaps the most urgent research topics to
address are the automatic georeferencing of imagery, establishing a standardized
procedure for orthorectification, and better understanding of the sensors calibration and
performances. Also, data acquired by the UAS could be explored in different research
within precision agriculture (e.g., crop stress, yield potential, crop disease). In addition,
exploring UAS data with physical-based model approaches rather than with statistical
and engineering approaches will create more robust algorithms that can be generalized
over different kinds of crops. Perhaps closing the link between scientists who generate
information from UAS data and farmers by presenting these findings in an actionable
manner remains the biggest missing link for further adoption of precision agriculture and
a significant research topic that needs to be addressed in future studies.
This work has demonstrated that the technology needed to produce actionable
information is possible but unfortunately, the cost embedded in this technology is far
from being affordable. The cost of flying the payload, collecting enough representable
samples to train the models and processing the data exceeds the limits of the average
122
farmer and as a result restrain the technology to research purposes only. Another big
challenge this technology presents is the big amount of data it produces. The “big data”
itself has advantages and short comes. A prominent advantage of “big data” is the
knowledge that emerges from the data. Such findings can change, modify or better
inform us about the research in question. For example, exploring the TIR data and its
relationship with chlorophyll estimation is something that haven’t been widely explored
in the community and this finding opened the door to such interest. On the other hand,
these “big data” sets are large enough to require supercomputers. Storing, managing and
processing these data is beyond the ability of commonly used computers. Perhaps, the
most fundamental challenge for big data applications is to explore the large volumes of
data and extract useful information or knowledge for future actions.
123
APPENDIX
124
Permission Letters
125
From: Permissions Helpdesk <[email protected]> Date: Mon, Mar 21, 2016 at 1:33 PM Subject: RE: Permission To: Manal Alarab <[email protected]>
Dear Manal, As an Elsevier journal author, you retain the right to Include the article in a thesis or dissertation (provided that this is not to be published commercially) whether in part or in toto, subject to proper acknowledgment; see http://www.elsevier.com/about/company-information/policies/copyright/personal-use for more information. As this is a retained right, no written permission from Elsevier is necessary. As outlined in our permissions licenses, this extends to the posting to your university’s digital repository of the thesis provided that if you include the published journal article (PJA) version, it is embedded in your thesis only and not separately downloadable: 19. Thesis/Dissertation: If your license is for use in a thesis/dissertation your thesis may be submitted to your institution in either print or electronic form. Should your thesis be published commercially, please reapply for permission. These requirements include permission for the Library and Archives of Canada to supply single copies, on demand, of the complete thesis and include permission for Proquest/UMI to supply single copies, on demand, of the complete thesis. Should your thesis be published commercially, please reapply for permission. Theses and dissertations which contain embedded PJAs as part of the formal submission can be posted publicly by the awarding institution with DOI links back to the formal publications on ScienceDirect. Best of luck with your thesis and best regards, Laura Laura Stingelin Permissions Helpdesk Associate Elsevier 1600 John F. Kennedy Boulevard Suite 1800 Philadelphia, PA 19103-2899 T: (215) 239-3867 F: (215) 239-3805 E: [email protected] Questions about obtaining permission: whom to contact? What rights to request? When is permission required? Contact the Permissions Helpdesk at:
126
+1-800-523-4069 x 3808 [email protected] From: Manal Alarab [mailto:[email protected]]
Sent: Monday, March 21, 2016 2:18 AM To: Permissions Helpdesk
Subject: Permission Dear editors, I am requesting permissions to my paper in the “ International Journal of Applied Earth Observation and Geoinformation”, as a Chapter in my dissertation. DOI of my paper: doi:10.1016/j.jag.2015.03.017 Reference: Elarab, Manal, Andres M. Ticlavilca, Alfonso F. Torres-Rua, Inga Maslova, and Mac McKee. "Estimating chlorophyll with thermal and broadband multispectral high resolution imagery from an unmanned aerial system using relevance vector machines for precision agriculture." International Journal of Applied Earth Observation and Geoinformation 43 (2015): 32-42. Manal Elarab, PhD Civil and Environmental Engineering Utah Water Research Laboratory - USU Phone: +1 (435) 754-9299 From: Inga Maslova <[email protected]> Date: Mon, Mar 21, 2016 at 8:02 AM Subject: Re: Permission To: Manal Al Arab <[email protected]>
Dear Manal,
Yes, you have my permission to include the paper in your thesis.
127
Sincerely, Inga Maslova Assistant Professor Department of Mathematics and Statistics American University
On Mar 21, 2016 2:21 AM, "Manal Alarab" <[email protected]> wrote: Dear Inga, I am in the process of preparing my dissertation. And I'm sending this email to request your permission to include in my dissertation a copy of the following paper in which you appear as a coauthor: ”ESTIMATING CHLOROPHYLL FROM THERMAL AND BROADBAND MULTISPECTRAL HIGH RESOLUTION IMAGERY FROM AN UNMANNED AERIAL SYSTEM USING RELEVANCE VECTOR MACHINES FOR PRECISION AGRICULTURE" Best, Manal -- Manal Elarab, PhD Civil and Environmental Engineering Utah Water Research Laboratory - USU Phone: +1 (435) 754-9299 From: Alfonso Torres-Rua <[email protected]> Date: Mon, Mar 21, 2016 at 6:48 AM Subject: Re: Permission To: Manal Alarab <[email protected]> Dear Manal, You have my permission. Alfonso Sent from my iPhone On Mar 21, 2016, at 12:20 AM, Manal Alarab <[email protected]> wrote:
Dear Alfonso,
128
I am in the process of preparing my dissertation. And I'm sending this email to request your permission to include in my dissertation a copy of the following paper in which you appear as a coauthor: ”ESTIMATING CHLOROPHYLL FROM THERMAL AND BROADBAND MULTISPECTRAL HIGH RESOLUTION IMAGERY FROM AN UNMANNED AERIAL SYSTEM USING RELEVANCE VECTOR MACHINES FOR PRECISION AGRICULTURE" Best, Manal -- Manal Elarab, PhD Civil and Environmental Engineering Utah Water Research Laboratory - USU Phone: +1 (435) 754-9299 From: Andres Ticlavilca <[email protected]> Date: Mon, Mar 21, 2016 at 5:52 AM Subject: Re: Permission To: Manal Alarab <[email protected]>
Dear Manal You have my permission. Best regards, Andres On Mon, Mar 21, 2016 at 2:21 AM, Manal Alarab <[email protected]> wrote: Dear Andres, I am in the process of preparing my dissertation. And I'm sending this email to request your permission to include in my dissertation a copy of the following paper in which you appear as a coauthor: ”ESTIMATING CHLOROPHYLL FROM THERMAL AND BROADBAND MULTISPECTRAL HIGH RESOLUTION IMAGERY FROM AN UNMANNED
129
AERIAL SYSTEM USING RELEVANCE VECTOR MACHINES FOR PRECISION AGRICULTURE" Best, Manal -- Manal Elarab, PhD Civil and Environmental Engineering Utah Water Research Laboratory - USU Phone: +1 (435) 754-9299 -- Andres M. Ticlavilca, Ph.D. Research Engineer Utah Water Research Laboratory 8200 Old Main Hill Logan, UT 84322-8200 Phone: 435-757-0851 Email: [email protected] Web: https://sites.google.com/a/aggiemail.usu.edu/ticlavilca/ -- Manal Elarab, PhD Civil and Environmental Engineering Utah Water Research Laboratory - USU Phone: +1 (435) 754-9299
130
CURRICULUM VITAE
Manal Elarab
[email protected]| 435-754-9299 (US Citizen)
Professional Profile An Engineer scientist with extensive background in agricultural research and remote sensing geospatial analysis. Proven ability to generate innovative approaches in precision agriculture combining UAV data and machine learning algorithms. Practical experience working in a team performing research to apply in commercial applications. Excellent initiative, communication, problem solving and management skills. Areas of Expertise Precision Agriculture Computational Modeling Geospatial Data Analysis Statistical Analysis Machine Learning UAV Application Spatial Evapotranspiration Spatial Data Mining Experimental Design
Experience Graduate Research Assistant 2012-2015 Utah Water Research Laboratory, Logan, Utah
Actively involved in development and calibration of UAV imagery—sensor selection, reflectance calibration, thermal calibration, orthorectification, ground data collection, comparative analysis, upscaling/downscaling data techniques
Developed models to estimate leaf chlorophyll and plant tissue nitrogen using high resolution thermal and broadband multi-spectral UAV imagery combined with relevance vectors machine algorithm and validated by ground data
Adapted METRIC model to estimate high resolution crop evapotranspiration using UAV data and evaluated against Landsat METRIC estimates
Graduate Research Assistant 2009-2011 American University of Beirut, Lebanon
Designed and installed irrigation systems Conducted experiments on water saving in agriculture and deficit irrigation
131
Relevant Projects
Private Farm—Oats and Alfalfa, Scipio, Utah (2012 – 2013): Developed models to estimate plant chlorophyll and leaf nitrogen content using UAV high resolution data. Experiment was conducted over two center pivots.
Designed data collection procedure, location and frequency Installed ground sensors and weather stations Collected field data over two growing seasons to develop and train statistical
models Collected data to calibrate thermal and reflectance imagery Coordinated experimental activity with farmers—harvesting, irrigation
scheduling, and fertilization Gallo Commercial Vineyard, Lodi, California (2014-2015): Mapped Evapotranspiration over vineyards using UAV high resolution remote sensing data, funded by ARS-USDA.
Designed and conducted data collection to calibrate thermal and reflectance imagery
Corrected high resolution reflectance data (visual and infrared) to Landsat 8 imagery
Developed Albedo coefficients customized to the UAV Adjusted METRIC model to accommodate the high resolution UAV data to
estimate crop evapotranspiration
Education Doctor of Philosophy in Civil and Environmental Engineering Dec 2015 Utah State University, Logan, Utah
Dissertation: The Application of Unmanned Aerial Vehicle to Precision Agriculture: Chlorophyll, Nitrogen, and Evapotranspiration Estimation
Masters of Science in Irrigation Engineering Dec 2011 American University of Beirut, Lebanon Masters of Science in Plant Production May 2008 Lebanese University, Lebanon Bachelor of Science in Agricultural Engineering May 2007 Lebanese University, Lebanon Skills
Computational: WEAP, Arc Map, Matlab, R, Python, Erdas Imagine Language: Fluent Arabic and English, Intermediate French
132
Interests and Activities
Cooking, Reading, Travelling, Hiking, Languages, Bowling and Mountain Biking Peer Reviewed Publications
Elarab, M., Ticlavilca, A. M., Torres-Rua, A. F., Maslova, I., & McKee, M. (2015). Estimating chlorophyll with thermal and broadband multispectral high resolution imagery from an unmanned aerial system using relevance vector machines for precision agriculture. International Journal of Applied Earth Observation and Geoinformation, 43, 32-42. Torres-Rua, A., Al Arab, M., Hassan-Esfahani, L., Jensen, A., & McKee, M. (2015). Development of unmanned aerial systems for use in precision agriculture: The AggieAir experience. In Technologies for Sustainability (SusTech), 2015 IEEE Conference on (pp. 77-82). IEEE. M. Al-Arab, A. Torres-Rua, A. M. Ticlavilca, A M. Jensen, M. McKee (2013). Use of high-resolution multispectral imagery from and unmanned aerial vehicle in precision agriculture. IEEE International Geoscience and Remote Sensing Symposium, Melbourne, Australia. M. Al-Arab L. Meeks A. Gilchrist (2012). Delaware River basin, Wikipedia www.engr.usu.edu/wiki/index.php/Delaware_River_Basin R. Bachour, M. Al-Arab, M. Nimah., & M. McKee (2011). Research Priorities for Water and Irrigation System Management on the Orontes Benchmark Site, Lebanon. Conferences M. Elarab, A. Torres-Rua, M. McKee; “Use of UAS Remote Sensing Data (AggieAir) to Estimate Crop ET at High Spatial Resolution” American Geophysical Union – Fall Meeting 2015 – Oral Presentation. M. Elarab, McKee, Mac; Torres-Rua, Alfonso FHassan-Esfahani, Leila; Jensen, Austin; “Use of UAS to Support Management in Precision Agriculture: The AggieAir Experience” ASA, CSSA & SSSA International Annual Meeting – Meeting 2015 – Poster Presentation
M. Elarab, A. Torres-Rua, M. McKee (2014). Oral presentation, Mapping ET at High Resolution using AggieAir Airborne Multispectral Imagery. ASABE International Symposium on "Evapotranspiration: Challenges in Measurement and Modeling from Leaf to the Landscape Scale and Beyond" being held in Raleigh, North Carolina, U.S.A. on April 7-10, 2014 M. Elarab, A. Torres-Rua, M. McKee (2014). Oral presentation, Mapping ET at High Resolution using AggieAir Airborne Multispectral Imagery. ASABE International Symposium on "Evapotranspiration: Challenges in Measurement and Modeling from Leaf to the Landscape Scale and Beyond" being held in Raleigh, North Carolina, U.S.A. on April 7-10, 2014
Al-Arab, Manal; Torres-Rua, Alfonso F; Ticlavilca, Andres; Jensen, Austin; McKee, Mac, “Use of high-resolution multispectral imagery from an unmanned aerial vehicle in precision
133
agriculture,” Geoscience and Remote Sensing Symposium (IGARSS), 2013 IEEE International”, 2852-2855, 2013, IEEE. M. Al-Arab, A. M. Ticlavilca, A. Torres-Rua, I. Maslova and M. McKee (2013). Oral presentation, Use of high-resolution multispectral imagery to estimate chlorophyll and plant nitrogen in oats (Avena sativa). Presentation at 2013 Fall Meeting, AGU, San Francisco, Calif. December 9-13, 2013. M.N Nimah, M. Al-Arab 2010 (keynote paper): Solutions for Sustainable Irrigated Agriculture and Food Security 2010 presented at the 1st Kuwait - Food security Expo, Kisr- Kuwait 2010 M.N Nimah, M. Al-Arab (2010) Comparison between Irrigation with Saline Water and regular Deficit Irrigation on Potato Yield presented at MERC Potato Symposium Larnaca, Cyprus, 2010