Soybean Yield Prediction at Field Level Using Sentinel-1 Data

Timely information of crop yield is important for farmers, insurance companies and policy makers. In many studies, optical satellite data were used for crop yield estimation and prediction. However, cloud cover impedes using optical data during the cloudy weather condition. In this study, it is demonstrated that Synthetic Aperture Radar (SAR) has high potential for soybean yield prediction at field level. Sentinel-1 Cross-Polarization (VH) intensity were used for prediction of soybean crop yield at different crop growth stages. Ground data was collected for 57 soybean fields over central Argentina. 29 Sentinel-1 Ground Range Detected (GRD) products were preprocessed. Artificial Neural Network (ANN) algorithm was trained using time-series of VH intensities for prediction of soybean yield. Three scenarios were considered for the ANN training and so crop yield prediction. The first scenario was to use time -series of VH intensities from crop planting date to two months before the harvest. This model provided us predicted soybean yield two months before the harvest date. The second scenario was to use time-series of VH intensities from planting date to one month before the harvest date and this model was used for yield prediction one month before the harvest date. And the third scenario was using all the Senetinel-1 VH intensities from the planting date to the harvest date and this model gave us yield estimation at the harvest date. Correlation coefficient (R), root mean square error (RMSE) and mean absolute error (MAE) are 0.74, 1.012 t/ha, 0.852 t/ha for the first scenario (two months before the harvest), 0.85, 0.769 t/ha, 0.651 t/ha for the second scenario (one month before the harvest) and 0.87, 0.701 t/ha and 0.567 t/ha for the third scenario (at the harvest date), respectively. It is observed that the accuracies of soybean yield prediction were improved as we moved closer to the harvest date and so having Sentinel-1 acquisitions for a wider portion of crop growth season. These improvements are significant between first and second scenarios however the second and third scenarios provided very close accuracies. Therefore, it is demonstrated that accurate soybean yield prediction one month before the harvest is feasible using time-series of Sentinel-1 data.