SOYBEAN YIELD FORECAST USING DUAL-POLARIMETRIC C-BAND SYNTHETIC APERTURE RADAR
Keywords: Soybean, Yield, Sentinel-1, Polarimetry, Neural Network
Abstract. Crop yield forecast is important, but determining productivity is an on-going challenge for agricultural communities including policy makers. An increase in the number of satellites, improved temporal and spatial resolutions and more open data policies, are leading to wider interest by the agricultural sector in exploiting space-based data at moderate spatial resolutions (10–30 m) and varying wavelengths (optical and microwave) for crop yield monitoring. This study evaluated Sentinel-1 dual-polarimetric data to forecast soybean yields one month before the harvest at field scales over a site in central Argentina. Specifically, polarimetric features were extracted from the Sentinel-1 data using the M-Chi decomposition. An Artificial Neural Network (ANN) model was trained using a time series of the single-bounce and volume scattering parameters derived from M-Chi from stacks of data acquired during the growing season. To estimate soybean yield from the ANN model, an innovative iterative retrieval method was developed. This retrieval approach improved the accuracies of the soybean yield forecast delivering a final coefficient of determination (R2) of 0.81 with a root mean square error (RMSE) of 755.81 kg/ha and mean absolute error (MAE) of 581.65 kg/ha. These accuracies demonstrate a high potential of Synthetic Aperture Radar (SAR) data from Sentinel-1 for soybean yield forecast at field scales.