Research Article Comparison of Sampling Designs for ...

Research ArticleComparison of Sampling Designs for EstimatingDeforestation from Landsat TM and MODIS ImageryA Case Study in Mato Grosso Brazil

Shanyou Zhu1 Hailong Zhang1 Ronggao Liu2 Yun Cao1 and Guixin Zhang1

1 School of Remote Sensing Nanjing University of Information Science amp Technology Nanjing 210044 China2 Institute of Geographic Sciences and Natural Resources Research CAS Beijing 100101 China

Correspondence should be addressed to Shanyou Zhu zsyzgx163com

Received 22 July 2014 Accepted 14 August 2014 Published 1 September 2014

Academic Editor Junsheng Yu

Copyright copy 2014 Shanyou Zhu et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Sampling designs are commonly used to estimate deforestation over large areas but comparisons between different samplingstrategies are required Using PRODES deforestation data as a reference deforestation in the state of Mato Grosso in Brazil from2005 to 2006 is evaluated using Landsat imagery and a nearly synchronous MODIS dataset The MODIS-derived deforestationis used to assist in sampling and extrapolation Three sampling designs are compared according to the estimated deforestationof the entire study area based on simple extrapolation and linear regression models The results show that stratified sampling forstrata construction and sample allocation using theMODIS-derived deforestation hotspots providedmore precise estimations thansimple random and systematic sampling Moreover the relationship between the MODIS-derived and TM-derived deforestationprovides a precise estimate of the total deforestation area as well as the distribution of deforestation in each block

1 Introduction

Human-induced and natural forest disturbances change for-est systems by influencing their composition structure andfunctional processes [1] Deforestation is the conversion offorested areas to nonforest land uses such as arable landurban areas logged areas or wasteland [2] and is importantfor forest resource management biodiversity conservationclimate change the global carbon cycle and sustainabilitymanagement [3ndash6] Research on the accurate monitoringof deforestation and its influence is a topic of considerableinterest in the context of global warming

Remotely sensed data with coarse spatial resolutionsuch as the Advanced Very High Resolution Radiome-ter (AVHRR) and Moderate-resolution Imaging Spectrora-diometer (MODIS) are commonly used over large areassuch as national continental climate zone or global scales[6ndash8] Because most deforestation occurs at subpixel scales[9 10] these data are inadequate for directly and preciselyestimating deforestation [11] High spatial resolution data

such as Landsat data allow for more accurate quantifica-tion of deforestation areas [12ndash14] However the infrequentrepeat coverage frequent cloud cover and data costs oftenpreclude the use of wall-to-wall mapping approaches withLandsat data for monitoring long-time deforestation overlarge regions [6 15 16] Most research adopts sample-basedmethodologies to estimate deforestation with higher spatialresolution imagery [6 10 13 17] Based on the estimatedresults from the sampling regions the deforestation area oreven the distribution of deforestation of the entire study areacan be extrapolated

Sample-based methods that use a probability samplingdesign provide a quantitative measure of the precision ofthe uncertainty that is attributable to sampling and constructconfidence bounds for the area of deforestationTheprecisiondepends on the number of samples and their locations Sam-pling is a cost- and time-efficient alternative if the objectiveis to estimate the area of deforestation rather than to mapdeforestation Sample-based methodologies can be classifiedas random sampling stratified sampling and systematic

Hindawi Publishing Corporatione Scientific World JournalVolume 2014 Article ID 919456 10 pageshttpdxdoiorg1011552014919456

2 The Scientific World Journal

Mato Grosso

Brazil

0 60 120 240 360 480

(Miles)

N

60∘09984000998400998400W 55∘09984000998400998400W 50∘09984000998400998400W

10∘09984000998400998400S10∘09984000998400998400S

15∘09984000998400998400S15∘09984000998400998400S

60∘09984000998400998400W 55∘09984000998400998400W 50∘09984000998400998400W

Figure 1 The study area of Mato Grosso (MT) in Brazil

sampling The random sampling method randomly selectscomplete images or several small blocks within the areaof interest and then analyzes the deforestation [6] Thistechnique has been applied by lots of applications [15ndash19]Based on parameters such as biome precipitation elevationdominant forest types land cover types disturbance degreetopography and soil types the stratified sampling methoddivides the study area into several strata and then selects thesame number of samples or allocates more samples into thestrata with greater expected levels of deforestation [20 21]The stratified sampling strategy is the common used methodin areas such as humid tropical forests [10 22ndash24] Systematicsampling selects samples at a defined spatial interval and iseasily performed compared to the two sampling methodsdescribed above Systematic sampling designs have beenadopted to monitoring deforestation over large areas bynumerous researchers [13 14]

Deforestation is a complex phenomenon of forest coverchange that is usually unevenly distributed and often con-verges within a smaller region [18 25 26] Although somesampling method has been recommended for operationalassessments of global and regional deforestation rates inthe tropics [27] or has been implemented in several studies[10 20 21 24] the sampling strategies have both advantagesand disadvantages sample-based methods for this specificapplication has only been the subject of limited evaluation[6] Comparisons between different sampling designs arerequired to further illustrate the adaptability of the sam-pling method to various spatial and temporal scales Thispaper compares and discusses the precisions of deforestationestimated by different sampling strategies using the state ofMato Grosso in Brazil as an example Landsat imagery and

a MODIS dataset from nearly the same period are used toidentify areas of deforestation and select sample blockswithinthe forest cover regions

2 Study Area and Data

21 Study Area The study area of Mato Grosso (MT) issituated in the midwestern part of Brazil and is one of thestates that have been designated as part of the Legal Amazon(Figure 1) MTrsquos total land area is 903386 km2 of which upto 678 is part of the southern basin of the Amazon River[28] The primary vegetation is characterized by latitudinalvariations from forests in the north to mixed zones in themiddle to Cerrado in the south [29] Deforestation in theBrazilian Amazon has been monitored for more than twodecades using a variety of satellite sensors Deforestationdata products have identified periods of increasing (2001ndash2004) and decreasing (2005ndash2007) deforestation rates inMato Grosso [30]

22 Data and Preprocessing

221 Landsat TM Based on the characteristics of the studyarea the high spatial resolution (30m times 30m) multispectralsatellite sensor Landsat Thematic Mapper (TM) is selectedto assess the distribution of deforestation This deforestationdataset is used as a reference for evaluating different samplingstrategies Forty-eight Landsat TM scenes from paths 223 to231 and rows 65 to 72were obtained during dry periods (timesof vegetation growth) from June 2005 to August 2006 Allof the TM image data were preprocessed with radiometriccalibration atmospheric correction based on the 6S radiant

The Scientific World Journal 3

Landsat imagery MODIS dataset

Data preprocessing Data preprocessing

Deforestation detectionbased on DI and NDVI

Deforestation detectionbased on MGDI

PRODESdeforestation

PRODESdeforestation

Random sampling(1) based on Landsat TM

(2) combined TM andMODIS-derived deforestation

Stratified sampling(1) proportionally distributed(2) based on Neyman optimal

allocation

Systematic sampling

Simple extrapolation Extrapolation based on the regression model

Comparison of different sampling designs

Deforestation from TM Deforestation from TM

(1) with an interval of 1∘

(2) with an intervals of 05∘


Figure 2 Research methodology flowchart

transfer model geometric correction and registration andimage mosaicking

222 MODIS The MODIS imagery dataset is composed ofthe 8-day composite 1 km daytime land surface tempera-ture (LST) dataset MYD11A2 the 16-day composite 500menhanced vegetation index (EVI) dataset MYD13A1 andthe yearly 500m land cover type dataset MCD12Q1 from2005 The LST data and EVI data were imaged from June10 2002 to August 15 2006 in which the data obtainedfrom June 2005 to August 2006 (the period of beingevaluated for deforestation) corresponded to the same periodas the TM data and the other data were needed as thereference in the deforestation detection algorithm The EVIdata were used because saturation levels are avoided whereasthe normalized difference vegetation index (NDVI) tends toapproach saturation levels in high biomass regions whichhas important consequences for change detection [31] TheMODIS land cover is used to mask all nonforest pixelsleaving all forest type pixels classified as evergreen needleleafforests evergreen broadleaf forests deciduous needleleafforests deciduous broadleaf forests and mixed forests Allof the deforestation detection processes are focused on theforest cover pixels The preprocessing steps for the MODISdata include resampling the LST data to a spatial resolutionof 500m mosaicking projection conversion quality controlto remove pixels with cloud contamination or low quality andcreating subsets using the vector data of MT

223 PRODES The Brazilian National Institute for SpaceResearch (INPE) PRODES has provided annual wall-to-wall

deforestation maps of the Brazilian Legal Amazon [32] since2000 These maps are used to evaluate the precision ofthe deforestation detected from the Landsat and MODISimagery PRODES employed a linear spectral mixing algo-rithm to generate vegetation soil and shade fraction imagesThe soil and shade fraction images were classified usingimage segmentation followed by unsupervised classificationimage editing andmosaicking PRODES pixels that had beeninterpreted as deforestation by the PRODES analysts betweenthe study periods were marked as deforestation and pixelsinterpreted as other PRODES classes were labeled as no-change

3 Methodology

A flowchart of the research methodology is shown inFigure 2

31 Deforestation Detection from Landsat TM Data Thedisturbance index (DI) is a simple and effective means oftracking vegetation disturbance across a variety of forestecosystems and is especially useful for identifying completeforest canopy removal [33 34] The DI is based on theTasseled-Cap data space The DI algorithm assumes that acombination of the greenness and the wetness can highlightthe spectral response characteristics of the vegetation whilethe brightness can express the characteristics of nonvegeta-tion areasThe brightness values of disturbed areas are higherthan those of nondisturbed forest areas while the greennessand wetness values are lower [33]


At a basic level the DI records the normalized spectraldistance of any given pixel from a nominal ldquomature forestrdquoclass to a ldquobare soilrdquo class The DI is calculated using theTasseled-Cap indices for Landsat TMETM+ [34]

DI = 1198611015840 minus (1198661015840 +1198821015840) (1)

where 1198611015840 1198661015840 and1198821015840 represent the Tasseled-Cap brightnessgreenness and wetness indices respectively normalized by adense forest class for each Landsat scene such as

1198611015840=

119861 minus 119906119861

120590119861

(2)

where 120583119861is the mean of the Tasseled-Cap brightness index

and 120590119861is the standard deviation of brightness within the

dense forest class for a particular sceneBecause the DI values are based on the statistics of the

forest reflectance from individual scenes the DI metric isrelatively insensitive to the variability in solar geometry orthe bidirectional reflectance distribution function (BRDF)between scenes and lessens the effect of vegetation phenolog-ical variability between image dates [34]

The original application by Healey et al [33] relied onthe absolute value of the DI from individual scenes to assessthe extent of disturbance Masek et al [34] used the decadalchange in DI value (ΔDI) as a more robust metric to identifydisturbance and recovery In this study we combined theabsolute value of DI

2005of the initial stage the change in DI

value and the change in NDVI to detect deforestation fromthe TM imagery The ΔDI and ΔNDVI were calculated asthe temporal change DI

2006-DI2005

and NDVI2006

-NDVI2005

respectively DI values greater than 10 have a high probabilityof being disturbed or nonforest Large positive values ofΔDI or ΔNDVI correspond to likely disturbance events andΔDI times ΔNDVI can increase the pixel difference betweendeforestation and nondeforestation Thresholds were appliedto the DI

2005and ΔDI times ΔNDVI values to identify potential

deforestation (DI2005lt thresh1 andΔDItimesΔNDVI gt thresh2)

32 Deforestation Detection from MODIS Data TheMODISglobal disturbance index (MGDI) algorithm [35 36] wasadopted for deforestation mapping from the LST and EVIdata The MGDI algorithm has been successfully tested overthe Western US and North America

Deforestation caused by instantaneous disturbances suchas wildfire results in an immediate departure of the LSTEVIratio from the range of natural variability [36] which can bedetected by MGDIInst as defined by

MGDIInst =(LSTmaxEVImax post)

119910

(LSTmaxEVImax post)119910minus1

(3)

where MGDIInst is the instantaneous MGDI value 119910 is theperiod (from June 2005 to August 2006) being evaluated fordeforestation (119910 minus 1) is the period from June 2002 to May2005 LSTmax is the maximum 8-day composite LST for thecomputation period and EVImax post is the maximum 16-day

EVI that occurred after the maximum LST during the sameperiod

Noninstantaneous disturbance events such as hurricanesand insect epidemics cause departures from the range ofnatural variability in the year following the disturbance [36]The noninstantaneous variant of the MGDI algorithm isgiven as follows

MGDINon-Inst =(LSTmaxEVImax)119910

(LSTmaxEVImax)119910minus1

(4)

whereMGDINon-Inst is the noninstantaneousMGDI value andEVImax is the maximum 16-day composite EVI during thecomputation period

When calculating MGDIInst and MGDINon-Inst for thenumerator in (3) and (4) LSTmax EVImax post and EVImaxduring the deforestation evaluated period 119910were determinedfor each pixel and then the LSTEVI ratio could be calculatedAs for the denominator the period from June 2002 toMay 2005 was averagely divided into three parts and eachsubperiod included 12 months (from June to May of the nextyear) the LSTmax EVImax post and EVImax for each subperiodwere extracted to compute the LSTEVI ratio and then themean LSTEVI ratio for all times (119910 minus 1) was determined foreach pixel

The difference between the instantaneous disturbancesand the noninstantaneous disturbances is controlled by therate of disturbance event The MGDI is a dimensionlessvalue and a given pixelrsquos MGDI value will tend toward unityin the absence of disturbance When a major disturbanceevent such as wildfire occurs LST will increase and EVIwill decrease instantaneously resulting in a MGDI valuethat is much larger than the multiyear mean MGDIInstand MGDINon-Inst are combined to be used for continuouswall-to-wall deforestation detection in the research Given asuitable threshold the deforestation pixels can be determinedbased on the MGDI map

33 Sampling Strategies Design Tucker and Townshend [17]randomly sampled complete Landsat images and determinedthat using whole scenes will result in smaller standard errorsand save little or nothing in acquisition costs HoweverDuveiller et al [13] suggested that the sampling efficiency canbe increased significantly by using small image extracts assampling units and having them systematically (rather thanrandomly) distributed over the forest domain Even if thestandard deviation of a small region is greater than that ofa large region more samples could compensate for the largerstandard deviation which means that more sampling unitswith small areas might achieve a higher precision [37ndash39]

Based on these analyses we sampled small 15 km times 15 kmblocks and the entire study area was divided into 4081blocksThe area and proportion of deforestation in each blockcorresponding to the Landsat TM and MODIS results couldthen be calculated Three sampling strategies were designedto select sample sites to estimate the deforestation areas inMTduring the period from 2005 to 2006The difference betweenour sampling designs and those in previously published


studies is that the selected sample blocks must include someforest pixels based on the MODIS MCD12Q1 data

331 Random Sampling Two random sampling methodswere used First we simply randomly selected samples withinthe forest cover region However because the deforestationregions usually distribute densely sampling sites and thevariation in densities could influence the precision of theresults when using simple random sampling Thus we thenselected samples within the regions where the proportion ofdeforestation from the MODIS dataset was greater than athreshold Various deforestation proportion thresholds wereused to further analyze the differences in the estimationresults

332 Stratified Sampling The stratification was based on theMODIS-derived deforestation The resulting low mediumhigh and very high deforestation strata were defined asMODIS-derived deforestation proportions of 0-1 gt1ndash5gt5ndash8 and gt10 per block respectively The sample siteswere allocated to the four strata using two different methodsFirst the samples are proportionally distributed And secondwe selected samples based onNeyman optimal allocation [6]The optimal allocation was determined using per stratumvariances of the MODIS-derived percentage of deforestationfor all blocks within each stratum

333 Systematic Sampling The systematic sampling designis based on the number of samples and fixed intervals wereused to obtain sample blocks Furthermore if there wereno forest pixels for a certain selected block we sampled thenearest block to the right or down as a substitute

34 Deforestation Extrapolation and Precision EvaluationTwo methods were used to compare the extrapolation pre-cision of the deforestation result over the entire study areaThe first simple extrapolation method was based on

119910 =

sum119899

119894=1119909119894

sum119899

119894=1119878119894

times 119860 (5)

where 119910 is the estimated area of deforestation in the studyarea 119909

119894and 119878119894are the deforestation area and the forest area

respectively for the 119894th sample block and 119860 is the total forestarea of the study area

Equation (5) can only be used to obtain the total defor-estation area but the distribution of deforestation cannot bespatially determined To estimate the distribution of defor-estation within each block a second extrapolation methodwas adopted according to the relationship between the TMand MODIS deforestations derived from the sample blocksas in Hansen et al [10] A simple linear regression model (6)was used to estimate the Landsat-scale deforestation area ineach blockThe total deforestation in the study area is the sumof the deforestation in all of the blocks

119863TM = 119886 times 119863MODIS + 119887 (6)

where 119863TM and 119863MODIS are the corresponding deforestationareas in each block and 119886 and 119887 are the coefficients determinedfrom the sample blocks

Compared to the deforestation derived from the Landsatimagery the relative error (120576) of the extrapolated deforesta-tion results for the entire study area was evaluated with

120576 =

1003816100381610038161003816119910 minus 119910

0

1003816100381610038161003816

1199100

times 100 (7)

where 119910 is the estimated deforestation area of the study areaand 119910

0is the TM-derived deforestation area

4 Results and Discussions

41 Deforestation Derived from TM Imagery and MODISData Based on the PRODES deforestation data as the ref-erence by visually interpreting and comparing the TM colorcomposite images pre- and postdeforestation the thresholdsused for detecting the deforestation from TM imagery weredetermined after several experiment tests The pixels werepreliminarily classified as deforested if their ΔDI times ΔNDVIvalues were greater than 50 and the DI

2005values were less

than 200 Furthermore each pixel was flagged as deforestedusing a median filtering method with a 3 times 3 pixel window toeliminate isolated pixels

In the TM deforestation results all of the nondeforesta-tion and no-data pixels were masked out in white and thedeforestation pixels were displayed in green The PRODESdeforestation data were overlaid on the TM deforestationresults to determine the detection precision In general thedeforestation detected based on the TM imagery is spatiallyconsistent with that determined from the PRODES dataFigure 3 compares the deforestation detected from the TMimagery (path 226 row 69) and the PRODES deforestationvector data The deforestation detected from the Landsat TMdata was consistent with that from the PRODES data suchas at site 1 Some locations such as site 2 were only detectedfrom the TM data while site 3 was only identified from thePRODES data

The MGDI algorithm described above was applied to theMODIS data for the years 2002ndash2006 to detect deforestationduring the period from 2005 to 2006 The thresholds fordetecting deforestation from the MGDI map were deter-mined using the deforestationmap derived from the TMdataas the reference We selected six deforestation areas wherethe disturbance information could be identified in MODISdata and tested different thresholds from40 to 80with theinterval 5 increases from themulti-sub-periodsmean valueBased on the calculation of total deforestation area across theselected area at each threshold any pixels with MGDI valuesgreater than 70 of the multiyear mean for instantaneousdisturbance or greater than 50 of the multiyear mean fornoninstantaneous disturbance were flagged as deforestationThe PRODES deforestation data were also overlaid on theMODIS deforestation results Figure 4 illustrates the preci-sion of the MODIS deforestation results compared to thePRODES results The large deforestation sites were success-fully detected based on the MGDI algorithm


1

1

11

1

1

2

2

3

Figure 3 A comparison of the deforestation detected from the TM imagery (green areas) and the PRODES deforestation data (red polygons)

Figure 4 A comparison of the deforestation detected from theMODIS data (red areas) and the PRODES deforestation data (greenpolygons)

The differences in the deforestation distributions gen-erated from the TM imagery MODIS data and PRODESdata were primarily caused by two factors First the TM andMODIS imagery were not collected at exactly the same timeas the PRODES data Some of the orbits were more thanonemonth apart and new deforestationmight have occurredduring the intervening period that could not be detectedin both this study and the PRODES results Second thePRODES deforestation data represent only the loss of intactforest area (old growth forest) so the target parameter doesnot include areas of forest cover change due to degradationregeneration afforestation regrowth or clearing of regrowthHowever several sites in which deforestation occurred before2005 as well as between 2005 and 2006 such as site 2 inFigure 3 were detected in this study

Because of the nearly identical image times of theMODISand TM data the deforestation results derived from theMODIS dataset and the TM imagery were compared by cal-culating the deforestation area within 99 15 k times 15 km blocks(Figure 5) There is a highly linear correlation between theMODIS-derived and the TM-derived deforestations thoughthe area derived from the TM data is slightly larger than thatof MODIS

0

5

10

15

20

25

30

0 5 10 15 20 25 30MODIS-derived deforestation area (km2)

TM-d

eriv

ed d

efor

esta

tion

(km

2)

y = 09363x + 02126

R2 = 06966n = 99

Figure 5 Plot of MODIS-derived versus TM-derived deforestationper sample block

42 Comparison of Sampling Designs Based on Simple Extrap-olation We analyzed the influence of the number of sam-pling blocks on the estimation of deforestation using (5)The analysis began with three randomly selected blocks andincreased to the total number of blocks in the study (4081)and each block could only be used once The variation ofthe estimation error with the number of sampled blocks ascalculated with (7) is shown in Figure 6

Figure 6 shows that the sample locations and the samplingdensity have a large effect on the estimation results If thefirst sampling blocks are located at a site where a very smallor a very large proportion of deforestation has occurred therelative error will fluctuate until it reaches a stable valueMoreover the estimation error varies little above a certainnumber of sampling blocks To further illustrate the influenceof the sampling density we compared various samplingdesigns using 100 350 and 500 sampling blocks

421 Error Analysis of Random Sampling Estimation Therandom sampling technique was designed to randomly selectsample blocks across the study area and in the blocks where


Table 1 Deforestation error 120576 () estimated based on random sampling

Sampling blocks Random sampling regionStudy area gt01lowast gt05lowast gt1lowast

100 2943 728 935 3117350 1352 476 2412 2933500 440 813 2684 3410lowastThe proportion means the MODIS-derived deforestation proportion in each block

0

20

40

60

80

100

1 409 817 1225 1633 2041 2449 2857 3265 3673 4081

Rela

tive e

rror

()

Sample blocks

Figure 6 The variation of deforestation estimation error with thenumber of sample blocks

MODIS-derived deforestation proportion is greater than01 05 and 1 Table 1 shows the relative estimationerrors extrapolated from the sample blocks corresponding tothe designed random sampling strategies

The relative error decreases with an increasing numberof sampling blocks when the samples are selected acrossthe entire study area The relative error when selecting 500blocks is 440 Based on the MODIS-derived proportion ofdeforestation the estimation precision improves dramaticallyif more than 01 of the blocks within the regions aresampled Because the study period is approximately one yearthe proportion of deforestation of each block might not belarge so the small areas of deforestation detected with theTM imagery most likely cannot be monitored using MODISimagery this changes the variance of the selected samplesand results in larger errors when using a larger deforestationproportion for the sample blocks The random samplingresults indicate that the deforestation can be determined withhigh precision if a suitable MODIS-derived deforestationproportion is used to assist in the sampling and extrapolationfrom the TM data

422 Error Analysis of Stratified Sampling The relative errorsof the extrapolation results from two sample allocationmethods proportionally distributed allocation and Neymanoptimal allocation compared to the TM-derived deforesta-tion map are given in Table 2

The extrapolation results of the stratified sampling showthat the precision of the Neyman optimal allocation methodis higher than that of the proportional allocation methodThe precisions of the two methods are similar when largenumbers of samples such as 350 and 500 blocks are used

Table 2 Deforestation error 120576 () estimated from two methods ofstratified sampling

Sampling blocks Proportionalallocation

Neyman optimalallocation

100 1430 834350 414 374500 404 362

Furthermore the precisions of the two stratified samplingmethods changed slightly when the number of samplesincreased from 350 to 500 which indicates that approxi-mately 350 samples are sufficient to obtain a reliable estimatewith stratified sampling in this study area

423 Error Analysis of Systematic Sampling Estimation Tocompare random sampling and stratified sampling withsystematic sampling using the same number of samples weadopted three systematic sampling intervals in which thesample blocks are located at the intersections of the 10∘05∘ and 042∘ lines of latitude and longitude These intervalscorrespond to 83 355 and 507 sample blocks respectivelyThe relative errors are shown in Table 3 The precision withthe 05∘ sampling interval is the highest of the three system-atic sampling designs However because fewer samples arelocated in strata with a high deforestation probability (gt5)the precision decreases when using 507 samplesThese resultsindicate that both the sample density and the sample locationsinfluence the precision of the deforestation extrapolation

424 Comparison of Sampling Strategies The estimationerrors shown in Tables 1 2 and 3 indicate that stratifiedsampling with approximately 350 sample blocks especiallyfor the Neyman optimal allocation method provides thehighest precision of estimates of the total deforestation areain the study area Random sampling requires 500 samplesto achieve the same level of precision Furthermore theestimation precision is not stable with a change in the sampledensity and with few sample blocks Systematic samplingcan obtain more reliable deforestation results than randomsampling which also depends on the sample locationsSampling blocks with forest cover can provide higher levelsof precision for the entire study area Based on this analysisstratified sampling is recommended as the best method tocombine information from both the low and high spatialresolutions because the low resolution signal allowed for theefficient targeting of deforestation hotspots


Table 3 Deforestation error 120576 () estimated from systematic sampling

Strata One degree interval 05 degree interval 042 degree intervalBlocks in strata () Relative error () Blocks in strata () Relative error () Blocks in strata () Relative error ()

0-1 68 (82)

1502

284 (80)

556

405 (80)

7211ndash5 10 (12) 55 (155) 86 (17)5ndash8 2 (24) 10 (28) 12 (23)gt8 3 (36) 6 (17) 4 (07)Total 83 (100) 355 (100) 507 (100)The number in the parentheses is the proportion of the sample blocks in each stratum

Table 4 Comparison of the relative error 120576 () extrapolated from two methods

Sampling blocks Random sampling Stratified proportional allocation Stratified Neyman optimal allocation Systematic samplingE1lowast E2lowast E1 E2 E1 E2 E1 E2

100 (83) 2943 1651 1430 932 834 415 1502 1214350 (355) 1352 1272 414 612 374 336 556 606500 (507) 440 363 404 516 362 282 721 1012The number in the parentheses is the number of sampling blocks used for the systematic samplinglowastE1 means the simple extrapolation from (5) and E2 is the regression extrapolation from (6) based on the relationship between the MODIS-derived and TM-derived deforestation

43 Comparison of Extrapolation Methods To obtain thespatial distribution of the deforestation sites a linear regres-sion model between the MODIS-derived and TM-deriveddeforestation (6) was constructed using the selected sampleblocks The model was then applied to the MODIS-deriveddeforestation map to obtain the corresponding deforestationarea in each block The deforestation results were evaluatedwith the TM-derived results and the relative errors arecompared with the results of the simple extrapolation inTable 4

The relative errors shown in Table 4 lead to severalconclusions First the regularity of the precision extrapolatedfrom (6) is similar to that obtained with (5) The two extrap-olation methods both provide the highest precision withthe stratified Neyman optimal allocation sampling methodand the lowest precision with random sampling Second theprecision extrapolated from the regression model is higherthan that from (5) The regression model (6) fully utilizes thedeforestation data of nonsampled blocks detected from theMODIS dataset but they are not considered in (5) Whileonly the total deforestation area is retrieved by the simpleextrapolation method the regression extrapolation methodcan provide the detailed deforestation proportion in eachblock with higher precision However the results estimatedfrom the regression model depend on accurately detecteddeforestation in both the MODIS and the TM datasets

5 Conclusions

Using the PRODES deforestation data as a reference defor-estation areas are detected from nearly synchronous TMimagery and MODIS datasets Several sampling designsemploying TM-derived and MODIS-derived deforestationwere compared to estimate the deforestation across thestudy area from 2005 to 2006 In general the sampling

approaches merit consideration as timely and cost-effectivecomponents for monitoring deforestation over large areasThe complete coverage TM-derived deforestation providesa unique opportunity to assess different sampling designsbecause it allows comparisons that are based on wall-to-wall estimators and are not estimated from single samples Astratified sampling method that included strata constructionand sample allocation provided more precise estimates thanboth simple random sampling and systematic samplingMoreover regressions between theMODIS-derived and TM-derived deforestation results provide precise estimates of boththe total deforestation area and the deforestation distributionin each block

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This work was conducted under the auspices of the Chinese973 Project (no 2010CB950701) the Natural Science Founda-tion of China (nos 41001289 and 41201369)TheMODIS datawere downloaded from httpreverbechonasagovreverband the PRODES deforestation vector data were downloadedfrom httpwwwobtinpebrprodes

References

[1] V H Dale L A Joyce S McNulty et al ldquoClimate change andforest disturbancesrdquo BioScience vol 51 no 9 pp 723ndash734 2001

[2] G Tejaswi Manual on Deforestation Degradation and Frag-mentation Using Remote Sensing and GIS MAR-SFM WorkingPaper Rome Italy 2007


[3] R A Houghton ldquoThe annual net flux of carbon to the atmo-sphere from changes in land use 1850ndash1990rdquo Tellus B vol 51no 2 pp 298ndash313 1999

[4] R Avissar and D Werth ldquoGlobal hydroclimatological tele-connections resulting from tropical deforestationrdquo Journal ofHydrometeorology vol 6 no 2 pp 134ndash145 2005

[5] LM Curran and SN Trigg ldquoSustainability science from spacequantifying forest disturbance and land-use dynamics in theAmazonrdquo Proceedings of the National Academy of Sciences of theUnited States of America vol 103 no 34 pp 12663ndash12664 2006

[6] M Broich S V Stehman M C Hansen P Potapov and Y EShimabukuro ldquoA comparison of sampling designs for estimat-ing deforestation from Landsat imagery a case study of theBrazilian Legal Amazonrdquo Remote Sensing of Environment vol113 no 11 pp 2448ndash2454 2009

[7] M C Hansen R S DeFries J R G Townshend R SohlbergC Dimiceli and M Carroll ldquoTowards an operational MODIScontinuous field of percent tree cover algorithm examples usingAVHRR andMODIS datardquo Remote Sensing of Environment vol83 no 1-2 pp 303ndash319 2002

[8] M C Hansen R S Defries J R G Townshend M CarrollC Dimiceli and R A Sohlberg ldquoGlobal percent tree cover at aspatial resolution of 500meters first results of theMODIS vege-tation continuous fields algorithmrdquoEarth Interactions vol 7 no10 pp 1ndash15 2003

[9] S Jin and S A Sader ldquoMODIS time-series imagery for forestdisturbance detection and quantification of patch size effectsrdquoRemote Sensing of Environment vol 99 no 4 pp 462ndash4702005

[10] M C Hansen S V Stehman P V Potapov et al ldquoHumidtropical forest clearing from 2000 to 2005 quantified by usingmultitemporal and multiresolution remotely sensed datardquo Pro-ceedings of the National Academy of Sciences of the United Statesof America vol 105 no 27 pp 9439ndash9444 2008

[11] D C Morton R S DeFries Y E Shimabukuro et al ldquoRapidassessment of annual deforestation in the brazilian amazonusing MODIS datardquo Earth Interactions vol 9 no 1 2005

[12] R L Czaplewski ldquoCan a sample of Landsat sensor scenesreliably estimate the global extent of tropical deforestationrdquoInternational Journal of Remote Sensing vol 24 no 6 pp 1409ndash1412 2003

[13] G Duveiller P Defourny B Desclee and P Mayaux ldquoDefor-estation in Central Africa estimates at regional national andlandscape levels by advanced processing of systematically-distributed Landsat extractsrdquo Remote Sensing of Environmentvol 112 no 5 pp 1969ndash1981 2008

[14] C Huang S Kim K Song et al ldquoAssessment of Paraguayrsquosforest cover change using Landsat observationsrdquo Global andPlanetary Change vol 67 no 1-2 pp 1ndash12 2009

[15] FAO ldquoTropical resources assessment 1990rdquo FAO forestry paperRome Italy 1993

[16] FAO ldquo Forest resources assessment 1990 survey of tropicalforest cover and study of change processesrdquo FAO Forestry Paper130 FAO Rome Italy 1996

[17] C J Tucker and J R G Townshend ldquoStrategies for monitoringtropical deforestation using satellite datardquo International Journalof Remote Sensing vol 21 no 6-7 pp 1461ndash1471 2000

[18] D Skole and C Tucker ldquoTropical deforestation and habitatfragmentation in the amazon satellite data from 1978 to 1988rdquoScience vol 260 no 5116 pp 1905ndash1910 1993

[19] S N Goward J G Masek W Cohen et al ldquoForest disturbanceand North American carbon fluxrdquo Earth Observing SystemTransactions vol 89 no 11 pp 105ndash106 2008

[20] FAO Global Forest Resources Assessment 2000 Main ReportFAO Forestry Paper Rome Italy 2001

[21] F Achard H D Eva H-J Stibig et al ldquoDetermination of defor-estation rates of the worldrsquos humid tropical forestsrdquo Science vol297 no 5583 pp 999ndash1002 2002

[22] G A Sanchez-Azofeifa ldquoSampling global deforestation data-bases the role of persistencerdquoMitigation and Adaptation Strate-gies for Global Change vol 2 no 2-3 pp 177ndash189 1997

[23] F J Gallego ldquoStratified sampling of satellite images with asystematic grid of pointsrdquo ISPRS Journal of Photogrammetry andRemote Sensing vol 59 no 6 pp 369ndash376 2005

[24] M Baumann M Ozdogan T Kuemmerle K J WendlandE Esipova and V C Radeloff ldquoUsing the Landsat record todetect forest-cover changes during and after the collapse of theSovietUnion in the temperate zone of EuropeanRussiardquoRemoteSensing of Environment vol 124 pp 174ndash184 2012

[25] D L Skole W H Chomentowski W A Salas and A DNobre ldquoPhysical and human dimensions of deforestation inAmazoniardquo BioScience vol 44 no 5 pp 314ndash322 1994

[26] T A Stone P Schlesinger R A Houghton and G MWoodwell ldquoA map of the vegetation of South America basedon satellite imageryrdquo Photogrammetric Engineering and RemoteSensing vol 60 no 5 pp 541ndash551 1994

[27] P Mayaux P Holmgren F Achard H Eva H J Stibig and ABranthomme ldquoTropical forest cover change in the 1990s andoptions for futuremonitoringrdquo Philosophical Transactions of theRoyal Society B vol 360 pp 373ndash384 2005

[28] Agencia Nacional de Aguas The Evolution of Water ResourcesManagement in Brazil ANA Setor Policial Sul Brazil 2002

[29] P M Fearnside ldquoDeforestation control in Mato Grosso a newmodel for slowing the loss of Brazils Amazon Forestrdquo Ambiovol 32 no 5 pp 343ndash345 2003

[30] D C Morton M H Sales C M Souza Jr and B GriscomldquoHistoric emissions from deforestation and forest degradationin Mato Grosso Brazil 1) source data uncertaintiesrdquo CarbonBalance and Management vol 6 article 18 2011

[31] A Huete K Didan T Miura E P Rodriguez X Gao andL G Ferreira ldquoOverview of the radiometric and biophysicalperformance of theMODIS vegetation indicesrdquo Remote Sensingof Environment vol 83 no 1-2 pp 195ndash213 2002

[32] Monitoramento da Floresta Amazonica Brasileira por SateliteProjeto PRODES 2008

[33] S P Healey W B Cohen Z Q Yang and O N KrankinaldquoComparison of tasseled cap-based Landsat data structuresfor use in forest disturbance detectionrdquo Remote Sensing ofEnvironment vol 97 no 3 pp 301ndash310 2005

[34] J G Masek C Huang R Wolfe et al ldquoNorth American forestdisturbance mapped from a decadal Landsat recordrdquo RemoteSensing of Environment vol 112 no 6 pp 2914ndash2926 2008

[35] D J Mildrexler M Zhao F A Heinsch and S W Running ldquoAnew satellite-based methodology for continental-scale distur-bance detectionrdquo Ecological Applications vol 17 no 1 pp 235ndash250 2007

[36] D JMildrexlerM Zhao and SW Running ldquoTesting aMODISglobal disturbance index across North Americardquo Remote Sens-ing of Environment vol 113 no 10 pp 2103ndash2117 2009

[37] R Dunn and A R Harrison ldquoTwo-dimensional systematicsampling of land userdquo Journal of Applied Statistics vol 42 no 4pp 585ndash601 1993


[38] E Tomppo R L Czaplewski and K Makisara ldquoThe role ofremote sensing in global forest assessmentrdquo Forest ResourcesAssessment Working Paper Rome 61 2002

[39] S V Stehman T L Sohl and T R Loveland ldquoAn evaluation ofsampling strategies to improve precision of estimates of grosschange in land use and land coverrdquo International Journal ofRemote Sensing vol 26 no 22 pp 4941ndash4957 2005

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Active and Passive Electronic Components

Control Scienceand Engineering

Journal of



RotatingMachinery


Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics


Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation



DistributedSensor Networks



Mato Grosso

Brazil

0 60 120 240 360 480

(Miles)

N

60∘09984000998400998400W 55∘09984000998400998400W 50∘09984000998400998400W

10∘09984000998400998400S10∘09984000998400998400S

15∘09984000998400998400S15∘09984000998400998400S

60∘09984000998400998400W 55∘09984000998400998400W 50∘09984000998400998400W

Figure 1 The study area of Mato Grosso (MT) in Brazil

sampling The random sampling method randomly selectscomplete images or several small blocks within the areaof interest and then analyzes the deforestation [6] Thistechnique has been applied by lots of applications [15ndash19]Based on parameters such as biome precipitation elevationdominant forest types land cover types disturbance degreetopography and soil types the stratified sampling methoddivides the study area into several strata and then selects thesame number of samples or allocates more samples into thestrata with greater expected levels of deforestation [20 21]The stratified sampling strategy is the common used methodin areas such as humid tropical forests [10 22ndash24] Systematicsampling selects samples at a defined spatial interval and iseasily performed compared to the two sampling methodsdescribed above Systematic sampling designs have beenadopted to monitoring deforestation over large areas bynumerous researchers [13 14]

Deforestation is a complex phenomenon of forest coverchange that is usually unevenly distributed and often con-verges within a smaller region [18 25 26] Although somesampling method has been recommended for operationalassessments of global and regional deforestation rates inthe tropics [27] or has been implemented in several studies[10 20 21 24] the sampling strategies have both advantagesand disadvantages sample-based methods for this specificapplication has only been the subject of limited evaluation[6] Comparisons between different sampling designs arerequired to further illustrate the adaptability of the sam-pling method to various spatial and temporal scales Thispaper compares and discusses the precisions of deforestationestimated by different sampling strategies using the state ofMato Grosso in Brazil as an example Landsat imagery and

a MODIS dataset from nearly the same period are used toidentify areas of deforestation and select sample blockswithinthe forest cover regions

2 Study Area and Data

21 Study Area The study area of Mato Grosso (MT) issituated in the midwestern part of Brazil and is one of thestates that have been designated as part of the Legal Amazon(Figure 1) MTrsquos total land area is 903386 km2 of which upto 678 is part of the southern basin of the Amazon River[28] The primary vegetation is characterized by latitudinalvariations from forests in the north to mixed zones in themiddle to Cerrado in the south [29] Deforestation in theBrazilian Amazon has been monitored for more than twodecades using a variety of satellite sensors Deforestationdata products have identified periods of increasing (2001ndash2004) and decreasing (2005ndash2007) deforestation rates inMato Grosso [30]

22 Data and Preprocessing

221 Landsat TM Based on the characteristics of the studyarea the high spatial resolution (30m times 30m) multispectralsatellite sensor Landsat Thematic Mapper (TM) is selectedto assess the distribution of deforestation This deforestationdataset is used as a reference for evaluating different samplingstrategies Forty-eight Landsat TM scenes from paths 223 to231 and rows 65 to 72were obtained during dry periods (timesof vegetation growth) from June 2005 to August 2006 Allof the TM image data were preprocessed with radiometriccalibration atmospheric correction based on the 6S radiant






PRODESdeforestation

PRODESdeforestation




allocation

Systematic sampling












3 Methodology





DI = 1198611015840 minus (1198661015840 +1198821015840) (1)


1198611015840=

119861 minus 119906119861

120590119861

(2)








2006-DI2005

and NDVI2006

-NDVI2005







119910


(3)






(4)












119910 =

sum119899

119894=1119909119894

sum119899

119894=1119878119894

times 119860 (5)





119863TM = 119886 times 119863MODIS + 119887 (6)



120576 =

1003816100381610038161003816119910 minus 119910

0

1003816100381610038161003816

1199100

times 100 (7)










1

1

11

1

1

2

2

3





0

5

10

15

20

25

30


TM-d

eriv

ed d

efor

esta

tion

(km

2)

y = 09363x + 02126

R2 = 06966n = 99









0

20

40

60

80

100

1 409 817 1225 1633 2041 2449 2857 3265 3673 4081

Rela

tive e

rror

()

Sample blocks









100 1430 834350 414 374500 404 362







0-1 68 (82)

1502

284 (80)

556

405 (80)







5 Conclusions





Acknowledgments


References












































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation














PRODESdeforestation

PRODESdeforestation




allocation

Systematic sampling












3 Methodology





DI = 1198611015840 minus (1198661015840 +1198821015840) (1)


1198611015840=

119861 minus 119906119861

120590119861

(2)








2006-DI2005

and NDVI2006

-NDVI2005







119910


(3)






(4)












119910 =

sum119899

119894=1119909119894

sum119899

119894=1119878119894

times 119860 (5)





119863TM = 119886 times 119863MODIS + 119887 (6)



120576 =

1003816100381610038161003816119910 minus 119910

0

1003816100381610038161003816

1199100

times 100 (7)










1

1

11

1

1

2

2

3





0

5

10

15

20

25

30


TM-d

eriv

ed d

efor

esta

tion

(km

2)

y = 09363x + 02126

R2 = 06966n = 99









0

20

40

60

80

100

1 409 817 1225 1633 2041 2449 2857 3265 3673 4081

Rela

tive e

rror

()

Sample blocks









100 1430 834350 414 374500 404 362







0-1 68 (82)

1502

284 (80)

556

405 (80)







5 Conclusions





Acknowledgments


References












































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











DI = 1198611015840 minus (1198661015840 +1198821015840) (1)


1198611015840=

119861 minus 119906119861

120590119861

(2)








2006-DI2005

and NDVI2006

-NDVI2005







119910


(3)






(4)












119910 =

sum119899

119894=1119909119894

sum119899

119894=1119878119894

times 119860 (5)





119863TM = 119886 times 119863MODIS + 119887 (6)



120576 =

1003816100381610038161003816119910 minus 119910

0

1003816100381610038161003816

1199100

times 100 (7)










1

1

11

1

1

2

2

3





0

5

10

15

20

25

30


TM-d

eriv

ed d

efor

esta

tion

(km

2)

y = 09363x + 02126

R2 = 06966n = 99









0

20

40

60

80

100

1 409 817 1225 1633 2041 2449 2857 3265 3673 4081

Rela

tive e

rror

()

Sample blocks









100 1430 834350 414 374500 404 362







0-1 68 (82)

1502

284 (80)

556

405 (80)







5 Conclusions





Acknowledgments


References












































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation















119910 =

sum119899

119894=1119909119894

sum119899

119894=1119878119894

times 119860 (5)





119863TM = 119886 times 119863MODIS + 119887 (6)



120576 =

1003816100381610038161003816119910 minus 119910

0

1003816100381610038161003816

1199100

times 100 (7)










1

1

11

1

1

2

2

3





0

5

10

15

20

25

30


TM-d

eriv

ed d

efor

esta

tion

(km

2)

y = 09363x + 02126

R2 = 06966n = 99









0

20

40

60

80

100

1 409 817 1225 1633 2041 2449 2857 3265 3673 4081

Rela

tive e

rror

()

Sample blocks









100 1430 834350 414 374500 404 362







0-1 68 (82)

1502

284 (80)

556

405 (80)







5 Conclusions





Acknowledgments


References












































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










1

1

11

1

1

2

2

3





0

5

10

15

20

25

30


TM-d

eriv

ed d

efor

esta

tion

(km

2)

y = 09363x + 02126

R2 = 06966n = 99









0

20

40

60

80

100

1 409 817 1225 1633 2041 2449 2857 3265 3673 4081

Rela

tive e

rror

()

Sample blocks









100 1430 834350 414 374500 404 362







0-1 68 (82)

1502

284 (80)

556

405 (80)







5 Conclusions





Acknowledgments


References












































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation













0

20

40

60

80

100

1 409 817 1225 1633 2041 2449 2857 3265 3673 4081

Rela

tive e

rror

()

Sample blocks









100 1430 834350 414 374500 404 362







0-1 68 (82)

1502

284 (80)

556

405 (80)







5 Conclusions





Acknowledgments


References












































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation












0-1 68 (82)

1502

284 (80)

556

405 (80)







5 Conclusions





Acknowledgments


References












































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation


















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation









Research Article Comparison of Sampling Designs for ...

Documents

Transcript of Research Article Comparison of Sampling Designs for ...