In-season estimation of sunflower and wheat yields using optical multi-temporal satellite images

Rémy Fieuzal - Vincent Bustillo - David Collado - Gérard Dedieu

1

THEIA WORKSHOP – 13-14 June 2018

2

Accurate managements need to combine sustainability of resources and sufficient level of production to meet the food needs…

Introduction

• General context

• Global issuesClimate change (increase of mean temperature,

modification of precipitation patterns) Effects on agriculture?

Population growth (9,3 milliards in 2050 ?) Increase of food needs…

• Satellite missions at high spatial and temporal resolutionsOn going microwave missions: TerraSAR-X, Tandem-X, Radarsat-2, COSMO-SkyMed, Sentinel-1a/b, Alos-2…

On going optical missions: Landsat, Sentinel-2, Venµs, Pléiades…

Coming soon : TerraSAR-X2, Radarsat Constellation, Tandem-L…

• Crop • Approach • Method

3

Introduction

• General context • Crop • Approach

• Distribution of the world production in 2010

5 countries account for 58% of the total

productionUkraine, Russia, China, Argentina and France

• Method

• Sunflower worldwide – From 1961 to 2016 (FAOSTAT)

Trend: +0.4 Million of Ha per year Trend: +0.6 Million of Tonnes per year

ProductionArea harvested

4

Introduction

• General context • Crop • Approach

• Distribution of the world production in 2010

5 countries account for 52% of the total

productionChina, India, USA, Russia

and France

• Method

• Wheat worldwide – From 1961 to 2016 (FAOSTAT)

Maximum: 239 Million of Ha Trend: +8.7 Million of Tonnes per year

ProductionArea harvested

5

OPTICAL

Objective Estimation of the sunflower and wheat yields all along theagricultural season (updating estimates after each satellite acquisition)

Frequency

Wavelength

Visible

Infrared

0,38×10-6 m

0,78×10-6 m

1×10-4 m

1 m0,3×109 Hz

3×1012 Hz

3,8×1014 Hz

7,9×1014 Hz

Microwaves

Introduction

• Crop • Approach• General context • Method

SOWING +30

SOWING +60

SOWING +90

6

Introduction

• Crop• General context • Method• Approach

OPTICAL

SOWING +30

SOWING +60

SOWING +90

Yield forecast

in real-time

Random Forest - Cross Validation

1 to n images acquired during the crop cycle

Time

HarvestSowing

Successive Landsat-8 and Sentinel-2 images

Yields collected at the intra-field scale

Yield forecast throughout the agricultural season

7

Introduction

• Study area

Experiment

High spatial and temporal dynamics of the surface states

• Meteorological conditions are steered by a temperate climate

• Surface dedicated to agriculture:56,8% seasonal crops32,1% grasslands7,9% forests2,4% urban areas0,8% lakes

• Satellite Data • Ground Measurements

8

• The approach consists in using multi-temporal optical acquisitions

Introduction Experiment

• Study area • Satellite Data • Ground Measurements

Wheat

Sunflower

Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Year 2016 Year 2017

6 images Landsat-8

4 images Sentinel-2

11 images Sentinel-2

7 images Landsat-8

7 images Sentinel-2

12 images Sentinel-2

Use of 6 reflectances: blue, green, red, NIR, SWIR-1/2 NDVI derived from red and NIR reflectances

9

Introduction

• Measurements of wheat yieldsDescriptive statistics by field (μ,σ)Agricultural seasons 2016 et 2017 (10 et 12 fields)

• Satellite Data • Ground Measurements• Study area

Experiment

Wheat

Mean yield:- 2016 58.0 q.ha-1 (CV 14 to 21%)- 2017 50.4 q.ha-1 (CV 16 to 22%)

Year 2016 Year 2017

10

Introduction

• Measurements of sunflower yieldsDescriptive statistics by field (μ,σ)Agricultural seasons 2016 et 2017 (12 et 10 fields)

• Satellite Data • Ground Measurements• Study area

Experiment

Mean yield:- 2016 25.1 q.ha-1 (CV 18 to 36%)- 2017 21.5 q.ha-1 (CV 18 to 31%)

Sunflower

Year 2016 Year 2017

Introduction Experiment Results

• Estimation of wheat yields

11

• Test of different ratio of data for Cal/ValUsing all the images during the agricultural season

Wheat

• Estimation of sunflower yields

NDVI

6 Bands

NDVI

6 Bands

Year 2016 Year 2017

Statistics for the 50-50% ratio:- 2016 R²=0.63/0.63, RMSE=7.0/7.2 q.ha-1 for NDVI or 6 bands, - 2017 R²=0.46/0.50, RMSE=7.9/7.6 q.ha-1 for NDVI or 6 bands

Introduction Experiment Results

12

• Forecast of yield throughout the agricultural seasonUsing an increasing number of successive images

Wheat

• Estimation of wheat yields • Estimation of sunflower yields

NDVI

6 Bands

NDVI

6 Bands

Year 2016 Year 2017

Statistical performances increase with the number of images… Best performances when using 6 bands in early stages in 2017

Introduction Experiment Results

13

Sunflower• Test of different ratio of data for Cal/ValUsing all the images during the agricultural season

• Estimation of wheat yields • Estimation of sunflower yields

NDVI

6 Bands

NDVI

6 Bands

Year 2016 Year 2017

Statistics for the 50-50% ratio:- 2016 R²=0.59/0.64, RMSE=4.6/4.5 q.ha-1 for NDVI or 6 bands, - 2017 R²=0.66/0.67, RMSE=3.3/3.3 q.ha-1 for NDVI or 6 bands

Introduction Experiment Results

14

Sunflower

• Estimation of wheat yields • Estimation of sunflower yields

• Forecast of yield throughout the agricultural seasonUsing an increasing number of successive images

NDVI

6 Bands

NDVI

6 Bands

Year 2016 Year 2017

Statistic performances saturate from flowering Early accurate estimates: start of July…

Introduction Experiment Results

15

Sunflower

• Estimation of wheat yields • Estimation of sunflower yields

• Forecast of yield throughout the agricultural seasonUsing an increasing number of successive images

Statistic performances saturate from flowering Early accurate estimates: start of July…

1 image2 images3 images4 images5 images6 images7 images8 images9 images10 imagesNDVI

6 Bands

Year 2016

Introduction Experiment Results

16

Sunflower

• Estimation of wheat yields • Estimation of sunflower yields

• Forecast of yield throughout the agricultural seasonUsing an increasing number of successive images

Statistic performances saturate from flowering Early accurate estimates: start of July…

NDVI

6 Bands

Year 2016

Yields at flowering Yields at harvest

6 images 10 images

Observed yields

R² : 0.59RMSE : 4.7 q.ha-1

R² : 0.64RMSE : 4.5 q.ha-1

17

Introduction Experiment Results Conclusion

• The statistical approach based on multi-temporal optical images allows the estimation of crop yields with acceptable performances at a decametric spatial scale. This approach is in the framework of the on-going generation of satellite mission and

must be extended adding other satellite data…

• The proposed approach provide a useful tool for the monitoring of two crops with specific agricultural seasons (winter and summer). The approach must be extended to other crops…

• Interesting early accurate estimation of yield are observed for sunflower, whatever the considered year. The approach must be confirmed analyzing several other agricultural seasons…

• Those promising results are consistent with previous studies.

For more details…Estimation of corn yield using multi-temporal optical and radar satellite data and artificial neural networksR. Fieuzal, C. Marais Sicre and F. Baup - International Journal of Applied Earth Observation and Geoinformation – 2017Forecast of wheat yield throughout the agricultural season using optical and radar satellite imagesR. Fieuzal and F. Baup - International Journal of Applied Earth Observation and Geoinformation - 2017

3 months before harvest

Just before harvest

R² : 0.69RMSE : 7.0 q.ha-1

R² : 0.77RMSE : 6.6 q.ha-1

Other studies…Introduction Experiment Results Conclusion

• Yield estimates at the field scale…Approach applied to corn and wheat

• Combination of optical and C-band data

Best performances and satellite configurations throughout the agricultural season of corn

• Crop modeling

18

Other studies…Introduction Experiment Results Conclusion

• Yield estimates at the field scale…Approach applied to soybean and sunflower

• Combination of optical and C-band data • Crop modeling

For more details…Assimilation of LAI and Dry Biomass data from optical and SAR images into an agro-meteorological model to estimate soybean yield - J. Betbeder, R. Fieuzal and F. Baup - IEEE Jour. of Sel. Top. in App. Earth Obs. and Remote Sensing – 2016Estimation of sunflower yield using a simplified agro-meteorological model controlled by multi-spectral satellite data (optical or radar) - R. Fieuzal, C. Marais-Sicre and F. Baup - IEEE Jour. of Sel. Top. in App. Earth Obs. and Remote Sensing - 2017

RADAR

Biomass

Leaf Area Index

Soil moisture

Vegetation indices

OPTICAL

Transpiration

Evaporation

19

Backscattering coefficients

Sunflower

Crop model & optical images

Crop model & optical + radar images

Crop model & C-band images

Estimation of crop yieldsR²OPT+SAR= 0.87

R²OPT= 0.65, R²SAR= 0.57

Soybean

Thank you for your attention

20

THEIA WORKSHOP – 13-14 June 2018