SYNERGETIC USE OF OPTICAL, MICROWAVE AND THERMAL SATELLITE DATA FOR NON-PARAMETRIC ESTIMATION OF WHEAT GRAIN YIELD
Crop yield maps are very crucial inputs for different practical applications like crop production estimation, pay-out of crop insurance, yield gap analysis etc. Satellite derived vegetation indices across different electromagnetic region has the ability to explain the variation in crop yield and can be used for prediction of yield before harvesting. This study utilised indices derived from multi-temporal Optical, Thermal and Radar data for developing model for Wheat (Triticum aestivum) grain yield using Machine learning approaches i.e., Random Forest Regression (RFR). Time series of Sentinel-2 derived Normalized difference vegetation index (NDVI), Normalized difference water Index (NDWI), Landsat-8 derived GPP using LST-EVI relationship (Temparature-Greeness model) and Sentinel-1 derived cross-polarization backscatter ratio (σVH/σVV) were used as predictor for wheat yield estimation. Actual grain yield measurements at ground were carried out at the end of the season over 178 locations. Seventy five percent of ground yield data were used for training of the model and rest twenty five percent data were used for its validation. All the datasets were grouped into ten fortnightly datasets ranging from November 2017 to March 2018. Through the random forest regression using time-series of NDVI alone, wheat grain yields were estimated with an RMSE of 9.8 Q ha−1. Subsequently by adding the multi-temporal NDWI, GPP and σVH/σVV led to the improvement of RMSE to 8.7, 7.6 and 7.4 Q ha−1 respectively. Variable importance based on the out of box error showed the significance of NDVI, NDWI and GPP during Dec-Jan and σVH/σVV during Feb for wheat grain estimation. It was concluded that the RFR algorithm together with the indices from optical, thermal and microwave satellite data can able to produced significantly accurate estimates of wheat grain yield.