Robust Multisensor Prediction of Drought-Induced Yield Anomalies of Soybeans in Argentina

Abstract

A multisensor method for the prediction of drought-induced agricultural impact is put forth in this letter. The input data considered include MODIS NDVI and land surface temperature (LST), ESA-CCI Soil Moisture, and CHIRPS rain data, which are processed at the department level in a large and sparsely monitored cropland in Argentina. As ground truth, we have used department-scale crop losses estimated by an annual agricultural census. In particular, the period under consideration (2001-2019) includes five severe drought events where soybean production in the area was considerably affected. The proposed method is based on Lasso regression of the corresponding rank values of the satellite data to the relative yield anomalies. Importantly, the proposed methodology is robust to extreme drought events. In addition, an associated early warning classification method results in an overall accuracy no worse than 70% up to one month before the harvest, and 62% two months before the harvest. The proposed methodology offers a valuable method for the prediction of agricultural drought impact and should be especially valuable in sparsely monitored regions of the world.