Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through...

24
Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University of Illinois Urbana-Champaign Craig Wenzel, Brian Gilmore, Larry Hendrickson, John Deere Technology Innovation Center

Transcript of Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through...

Page 1: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

Improving Climate Resilience Through Smart Agriculture

Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker,

University of Illinois Urbana-Champaign

Craig Wenzel, Brian Gilmore, Larry Hendrickson,

John Deere Technology Innovation Center

Page 2: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

Field Readiness Predictions

Input: Rainfall

Output: Qualitative Field Readiness (rated 1-5)

The model is trained on these data to predict:

‘READY’ or ‘NOT READY’

This tractor is a very

expensive waste of

machine resources…

This tractor could become

stuck and/or leave ruts.

Page 3: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

Active Soil Moisture Sensor Poor/Missing Data

EBI_NE

EBI_NE

EBI_CEN

EBI_SE

Jordan'sSampling

Case Study Test Site:

South Farms, Urbana, IL

(Left) The South Farms, Urbana-Champaign, IL

(Right) Soil Sensor Locations

Wetness/Dryness conditions were

recorded by a John Deere intern

during the summer of 2010.

Algorithms were constructed to

predict these conditions remotely

using Illinois Climate Network

(ICN) data along with public

Nexrad radar data.

Page 4: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

Field Readiness Workflow

Page 5: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

Comparison of Algorithm Performance:

Classification Trees, KNN, and Boosted Perceptrons

0.0%

10.0%

20.0%

30.0%

40.0%

50.0%

60.0%

70.0%

80.0%

90.0%

100.0%

ClassificationTree KNN BoostedPerceptron

% O

f D

ays

Cla

ssif

ied

Corr

ectl

y

Agreement of Human and Computer

Page 6: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

KNN Results, Summer 2010

All testing examples for which no rain has fallen for 3+ days have been removed

All errors fall within margin of error (the error band crosses the dotted line)

All errors occur on days for which readiness is “2” or “3”

(Borderline qualitative assessments)

Page 7: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

The Next Step: Understanding the problem on a national scale, using soil

moisture as our proxy for “readiness”

?

?

?

?

?

?

?

Page 8: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

Diagnostic Soil Moisture Equation by Pan et al (2012)

(Overlaid with a machine learning model for error-correction)

Assumes that soil moisture losses due to evapo-transpiration, drainage, etc can be

estimated by an annual sinusoidal function. This requires three parameters…

y = sin(x) y = sin(x) + v {} {v} y = αsin(x) + v {v, α} y = αsin(x-h) + v {v, α, h}

Page 9: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

Diagnostic Soil Moisture Equation:

6 Parameters are required

Once these first three parameters are fit via a genetic algorithm, three final parameters are fit via

a 2nd genetic algorithm based on observed soil moisture values and the chosen loss function.

{v, α, h} {v, α, h, ϴre}

Residual Soil Moisture Effective Porosity Soil Drainage Constant

{v, α, h, ϴre , Фe} {v, α, h, ϴre , Фe , C4}

Page 10: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

Machine Learning Overlay:

Improvement Beyond the Published Literature

In New Mexico, the model tends to perform very well. The

ML algorithm in this place makes a tremendous difference.

ρ = 0.860 ρ = 0.917

Page 11: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

“So…you can model soil moisture at these

locations where you have soil moisture

sensors…but what about everywhere else?”

(We need a national system for hydrologic classification)

Page 12: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

Hydrologic Classification of Watersheds: 4 Features

Seasonality (is rainfall consistent year-round or variable?)

Aridity (how does total annual precipitation compare with annual potential evap.?)

Maximum Precipitation Timing (when is rainfall the highest?)

Maximum Runoff Timing (when is streamflow the highest?)

Page 13: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

Level One (2 splits)

­

Level Two (4 splits)

­

The first split divides the country via seasonality, dividing

the less seasonal east from the more seasonal west.

Exceptions are found in the Rockies and in Florida.

The second split

divides the eastern half

of the country via

rainfall timing and the

Western half via aridity.

Recursive Splitting: Building a “Classification Tree”

Splitting continues until clusters of

similar hydrologic behavior emerge.

Page 14: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

Classes - 428 MOPEX Catchments

IACJ

IAF

IAQ

IHD

IHM

ISCB

ISCJ

ISQJ

ITC

ITF

LBMH

LBMS

LJ

LPC

LPM

LPQ

LWC

XACJ

XADB

XHD

XSC

XSMB

XTM

XVM

­

The Finished Tree, All 24 Classes

LWC

LJ

LPC

ITC

ISQJ

LPM

Only 6 classes describe over 77% of

the 428 catchments in the database.

Page 15: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

What About Non-Stationarity?

428 MOPEX Catchments:

Divided into Pre & Post-1975 data

Page 16: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

Second Split (Before & After)

High Seasonality & Arid (Red), High Seasonality & Humid (Yellow), Low Seasonality

& Early Precip (Dark Blue), Low Seasonality & Late Precip (Dark Blue)

Before 1980 After 1980

-130 -120 -110 -100 -90 -80 -7025

30

35

40

45

50

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

-130 -120 -110 -100 -90 -80 -7025

30

35

40

45

50

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Page 17: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

The Midwest, Before & After 1980:

Later Peak Runoff

Higher Low-Flows

Page 18: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

The Pacific Northwest, Before & After 1980:

Diminished

Snowpacks

Diminished

Snowpacks

Unchanged

Winter Runoff

Page 19: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

The Rocky Mountains, Before & After 1980:

Smaller

Snowpacks

Smaller

Snowpacks

Steadier Rainfall

Steadier Rainfall

Page 20: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

The Southeast, Before & After 1980:

Systemic Drying Systemic Drying

Page 21: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

“You can model soil moisture where you have soil moisture sensors…

You can tell me if two locations are similar…

You can select more recent years to define classification similarity…

Does this mean you can calibrate at one location and apply the

parameters at another ‘similar’ site?

(Yes!)

Page 22: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

The red samples consist of (x,y) pairs where x & y are

from unrelated classes on the classification tree.

The yellow samples consist of (x,y) pairs where x & y

are not from the same class, but from classes that are

different only by one “split” (feature) on the tree.

The green samples consist of (x,y) pairs where x & y are

from the same class.

Page 23: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

Making Predictions in Champaign…without

any prior knowledge of soil moisture

The Complete Process (A Summary)

Calibration

(Pan et al, 2012)

Classes - 428 MOPEX Catchments

IACJ

IAF

IAQ

IHD

IHM

ISCB

ISCJ

ISQJ

ITC

ITF

LBMH

LBMS

LJ

LPC

LPM

LPQ

LWC

XACJ

XADB

XHD

XSC

XSMB

XTM

XVM

­(Coopersmith et al, 2012)

Generalization

Via Classification

Cross-

Validation

Topography Nexrad Data,

From Champaign

Calibration of parameters

using a SCAN site of the

same class (Arkansas)

Machine

Learning

Future

Page 24: Improving Climate Resilience Through Smart Agriculture · Improving Climate Resilience Through Smart Agriculture Evan Coopersmith, Murugesu Sivapalan, Barbara Minsker, University

Potential Applications (Disaggregation)

SMAP data

(kilometer scale)

Precipitation radar data

(kilometer scale)

LiDAR data

(meter scale)