development of on-the-go nitrogen application algorithms for ......2017/03/24  · n wap6 wap7 wap8...

Development of On-the-go Nitrogen Application Algorithms for Cotton Production

Based on Active Reflectance Sensors

Terry Griffin Ph D CCATerry Griffin, Ph.D., CCAAssistant Professor – EconomicsDept of Agricultural Economics & AgribusinessDept. of Agricultural Economics & Agribusiness

Cotton Inc. Crop Management Seminar

Cotton NDVI Studies

• Collaborators from 9 statesFunding from Cotton Inc– Funding from Cotton Inc.

– Applied nitrogen studies

Caveats

• No economic analysis conducted– Statistical analysis of bio-physiological dataStatistical analysis of bio physiological data

Choosing Unit of “Time”

• Heat unit accumulation preferred to ‘time’…– Algorithm development requires ‘categorized’ timeAlgorithm development requires categorized time– Rather than GDD60, using WAP– DAP resulted in too many ‘bins’ for given amount of data

Histogram of wap

0014

00

Histogram of ndvi_dap14

00Histogram of GDD60

700

Freq

uenc

y

600

800

1000

120

Freq

uenc

y

600

800

1000

1200

Freq

uenc

y

300

400

500

600

5 10 15

020

040

0

ndvi_dap

0 50 100 150

020

040

0

GDD60

500 1000 1500 2000 2500

010

020

0

wap

DAP WAPGDD60

Precision Ag: The Tale of Two TechnologiesInformation-intensive & Embodied-knowledge

I f ti i t i E b di d k l dInformation-intensive• Field level data to

Embodied-knowledge• Information purchased

make decisions• Requires additional

in the form of an input• Requires minimal

data and skill• IPM

additional data/skill• Round-up Ready or Btp y

Two Faces of Precision AgricultureTwo Faces of Precision Agriculture

Information intensive Embodied knowledgeInformation-intensive• Yield monitors

Embodied-knowledge• Automated guidance

• Traditional variable rate applications

• On-the-go sensors applying variable rates

• *Data • *Automated

Motivation and Issue

• Analyses have been conducted for single datasetsFailed to model m ltiple climates s stems etc• Failed to model multiple climates, systems, etc.

• Global response estimated from all datasetsObj ti l h li i l i h• Objective: Develop on-the-go N application algorithm

Motivation and Issue• Analyses have been conducted for single datasets• Failed to model multiple climates, systems, etc.• Global response estimated from pooled dataset• Derive on-the-go nitrogen application algorithms

Obj ti l lti t t i t d t t ti t• Objective: analyze multi-state experiment data to estimate response between active sensor reflectance and cotton yield for on-the-go nitrogen management– Develop on-the-go N application algorithm

Data and Methods• Pooled model

– Multiple site-years; 9 states, >=11 PI, >=13 studiesp y– All data (sites and years) in single dataset – Datasets normalized and controlled for heterogeneity– Conditioned by year, location

• Correlation among?– yield, nitrogen rate, NDVI values and timing

• Regression analysis of pooled dataset

Challenges

• Cotton is a perennial• N rate sufficient to cause yield penalty• N rate sufficient to cause yield penalty• Identifying appropriate timing for NDVI

– DAP GDD60– DAP, GDD60

• No economic analysis conducted– Statistical analysis of bio-physiological dataStatistical analysis of bio physiological data

Challenges

• Cotton is a perennial• N rate sufficient to cause yield penalty• N rate sufficient to cause yield penalty• Identifying appropriate timing for NDVI

– DAP GDD60– DAP, GDD60

• Testing functional forms of relationships• Testing indexes for algorithm• Testing indexes for algorithm

– Variety, location, and GDD60 specific– Global algorithm with local intercept/slope shiftersGlobal algorithm with local intercept/slope shifters

• No economic analysis conducted– Statistical analysis of bio-physiological datay p y g

Sensor-Base N Status

• “Yield Potential” approach available for growers not mapping soilsfor growers not mapping soils

• OSU approached “worked” everywhere evaluated (SC TN LAeverywhere evaluated (SC, TN, LA, OK); however:– How early can readings be taken? In y g

some cases response not clear until 4-weeks after early bloomDebate if soil specific reference strips– Debate if soil specific reference strips needed

Visual/NDVI use of Ramp Strips

Th G S k d t ll ti ifi d h t i ll i The Green Seeker data collection verified what was seen visually in the ramp strips in SC. 30% N savings over blanket application with no yield reduction US modified OSU approach – 2007 to 2009

0.8

0.9

150 lbs/acre

0 5

0.6

0.7

DVI

Optimum Nitrogen use is visually seen as around 80 lbs/acre 0.3

0.4

0.5N

0 lbs/acre

around 80 lbs/acre

0.2

0 20 40 60 80 100 120 140 160

N R t (lb / )0 lbs/acre N Rate (lbs/acre)

Other Indexes and Relationships

y = 128 12x 94 249

250

-1

Varco et alMissouri (Scharf et al.)

y = 128.12x - 94.249R2 = 0.59

100

150

200

N R

ate,

lb N

acr

e-

0

50

0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75

Opt

imal

N

Greenseeker Red/NIR Relative to High N P lots

NDVI ~ N by HUA N t f l f d l i l ith d t i l ti

0.90

1.00

300 GDD60

Not as useful for developing algorithms… need categorical time

0 60

0.70

0.80 300 GDD60

600 GDD60

900 GDD60

0.40

0.50

0.60

ND

VI 1200 GDD60

1500 GDD60

0 10

0.20

0.30 1800 GDD60

2100 GDD60

2400 GDD60

-

0.10

0 30 60 90 120 150 180A li d Nit (lb / )

2400 GDD60

2700 GDD60

Applied Nitrogen (lbs/ac)

NDVI response to N application by time of measurement

0 80.91.0

0 9-1 0

0.50.60.70.8

ND

VI

0.9 1.00.8-0.90.7-0.80 6-0 7

120160

0 10.20.30.4N 0.6 0.7

0.5-0.60.4-0.50 3-0 4

0

4080

0.00.1 0.3 0.4

0.2-0.30.1-0.20 0-0 10 0.0 0.1

NDVI and Applied Nitrogen

• Determine the shape of the functional form– LinearLinear– Quadratic– Other?

0.8

– No relationship?

• Can NDVI proxy 0.4

0.6

ndvi

for applied N?0.

2

0 50 100 150

0.0

N

Correlation: N and NDVI by TimePearson Correlation Coefficient applies only to linear relationship

Bolded numbers are significant at 90% level

N WAP6 WAP7 WAP8 WAP9 <=WAP10

Bolded numbers are significant at 90% levelWAP8 and WAP9 become significant at 89 and 88% levels

N 1 0.49 0.73 0.66 0.65 0.83

WAP6 1 0.95 0.98 0.98 0.89

WAP7 1 0.99 0.99 0.99

WAP8 1 1.00 0.97

WAP9 1 0.96

WAP<=10 1

NDVI and Nitrogen by Time

0 80

0.90

1.00

0.60

0.70

0.80

I

NDVI_WAP6NDVI_WAP7NDVI WAP8

0 30

0.40

0.50

ND

VI NDVI_WAP8

NDVI_WAP9NDVI_WAP<=10

R squared values <0 05

0.10

0.20

0.30 R-squared values <0.05

0.6

0.8

-0 50 100 150 200

Applied N (lbs/ac) 0.0

0.2

0.4nd

vi

Applied N (lbs/ac)0 50 100 150

N

Algorithms

Source: Randy Taylor, Oklahoma State University

Histogram of ndviHistogram of NDVI Values

800

600

Algorithm can be “hard bounded”by 0.2 and 0.8

Freq

uenc

y

400

200

0.0 0.2 0.4 0.6 0.8

0

Source: Randy Taylor, Oklahoma State University

ndvi

100

100

100

Yield potential as function of NDVI by WAP70

8090

dat$

rely

ld 6080

dat$

rely

ld 6080

dat$

rely

ld

5060

020

40

020

40

0.3 0.4 0.5 0.6 0.7

dat$ndvi

0.2 0.3 0.4 0.5 0.6 0.7 0.8

dat$ndvi

0.2 0.3 0.4 0.5 0.6 0.7 0.8

dat$ndvi

100

100

100

WAP=5 WAP=6 WAP=7

4060

80

dat$

rely

ld

4060

80

dat$

rely

ld

4060

80

dat$

rely

ld0.2 0.3 0.4 0.5 0.6 0.7 0.8

020

0.2 0.4 0.6 0.8

020

0.0 0.2 0.4 0.6 0.8

020

4dat$ndvi dat$ndvi

dat$ndvi

WAP=8 WAP=9 WAP=10

Yield as Function of NDVI by WAP100

80

90

100

60

70

80

entia

l

YLD WAP5

40

50

Yie

ld P

ote YLD_WAP5

YLD_WAP6YLD_WAP7YLD WAP8

Upper bound for application can be set at NDVI=0.6

10

20

30 YLD_WAP8YLD_WAP9

YLD_WAP5 YLD_WAP6 YLD_WAP7 YLD_WAP8 YLD_WAP9NDVI

0

10

0.2 0.3 0.4 0.5 0.6 0.7 0.8

@maxY 1.2 0.6 0.6 1.5 0.8R2 0.4 0.15 0.22 0.32 0.38

NDVI

Yield Penalty as Function of NDVI: WAP=7

60

40

50

y

30

40

eld

pena

lty

10

20Yi

00.2 0.3 0.4 0.5 0.6

NDVI

Classic Cotton Lint Response to Applied N

• In theory, quadratic without plateau2 5

2.0

2.5

les/

ac)

1 0

1.5

nt y

ield

(bal

0.5

1.0

Cot

ton

lin

0.00 20 40 60 80 100 120 140

Applied N (lbs per acre)Applied N (lbs per acre)

Data is ‘messy’ and ‘all over the place’

Cotton yield by N rate: 7 site-years3000

2000

2500

lbs/

ac)

1500

2000

lint y

ield

(l

500

1000

Cot

ton

l

00 20 40 60 80 100 120 140 160

Applied Nitrogen (lbs N per Ac)Applied Nitrogen (lbs N per Ac)

Yield Potential by N Rate

90

100

60

70

80

tial OK

40

50

60

Yie

ld P

oten AL

TNLA1

Yield max occurs at 110-120 lbs N

10

20

30

Y MSGALA2

-

10

0 50 100 150 200Applied N (lbs/ac)

Yield Penalty by N Rate

30

35

20

25

alty OK

15

20

Yie

ld p

ena

TNLA1MS

5

10 GALA2

-0 50 100 150 200

Applied N (lbs/ac)Applied N (lbs/ac)

Chart Title

0.80 0.90 1.00

tial

0.50 0.60 0.70

Yie

ld P

oten

t

wap6wap7wap8

0.20 0.30 0.40

ND

VI o

r Y wap8wap9wap10yld

-0.10

0 50 100 150 200

yld

Applied N (lbs/ac)

Chart Title

0.80 0.90 1.00

tial

0.50 0.60 0.70

Yie

ld P

oten

t

wap6wap7wap8

0.20 0.30 0.40

ND

VI o

r Y wap8wap9wap10yld

-0.10

50 70 90 110 130 150

yld

Applied N (lbs/ac)

Practical Implications

• Lower bound = 0.2Upper bo nd 0 6• Upper bound = 0.6

• Difficult to predict total on-the-go applied productN d t t k t d t t fi ld– Need to take extra product to field

• Some applicators not capable of abruptly varying rates

Limitations

• Cotton is a perennial grown as summer annual– Cotton is a treeCotton is a tree

• Need more data!– Locations, years, varietiesoc o s, ye s, v e es– DAP/GDD60 of measurements– Variables: canopy height, time of day

• Profit maximization has not been evaluated• Heat unit accumulation metrics

– Use of GDD60 needs to be revisited

Acknowledgements

• Funding from Cotton Inc. • Collaborators in 9 states

State PI

Al b Ki B lk• Collaborators in 9 states Alabama Kip Balkcom

Arizona Ed Barnes

Arizona Pedro Andrade-Sanchez

Arkansas Tom Barber

Georgia George Velidis

Lousiana Brenda Tubana

Mississippi Jac Varco

Mississippi YufengMississippi Yufeng

Oklahoma Randy Taylor

South Carolina Phil Bauer

Tennessee John Wilkerson

Terry GriffinTerry Griffin Assistant Professor - Economics501 671 2182501.671.2182 tgriffin@uaex.edu

Correlation: Yield and GDD60

300 600 900 1200 1500 1800 2100 2400 2700

Pearson Correlation Coefficient applies only to linear relationship

YLD300 GDD60

600 GDD60

900 GDD60

1200 GDD60

1500 GDD60

1800 GDD60

2100 GDD60

2400 GDD60

2700 GDD60

YLD 1 -0.32 -0.25 -0.15 -0.02 0.16 0.37 0.60 0.80 0.92

300 GDD60 1 1 00 0 98 0 95 0 89 0 76 0 57 0 32 0 07300 GDD60 1 1.00 0.98 0.95 0.89 0.76 0.57 0.32 0.07

600 GDD60 1 1.00 0.97 0.92 0.81 0.63 0.39 0.15

900 GDD60 1 0.99 0.95 0.86 0.70 0.48 0.25

1200 GDD60 1 0.99 0.92 0.79 0.59 0.37

1500 GDD60 1 0.98 0.88 0.72 0.53

1800 GDD60 1 0 97 0 86 0 701800 GDD60 1 0.97 0.86 0.70

2100 GDD60 1 0.96 0.86

2400 GDD60 1 0.97

2700 GDD60 1