Exploring Physics-Informed Neural Networks for Crop Yield Loss Forecasting

In this study, we present a Physics-Informed Machine Learning (ML) approach for high-resolution crop yield loss forecasting. By leveraging simulation data, climate variables, and Sentinel-2 satellite imagery, we model biophysical crop properties. Namely, we estimate the actual evapotranspiration and the crop sensitivity to water scarcity. We employ a sequence-to-sequence approach and introduce a loss term that solves the FAO equation for crop yield response to water. This allows accurate estimates of crop yield loss at a subfield resolution, caused by limiting environmental conditions. Consequently, this enables an improved adaption to extreme environmental constraints. Initial results indicate that we can accurately predict biophysical properties, translating into effective yield forecasting, achieving an $R^2$ value of 0.75 for cereal crops, harvested in Switzerland between 2017 and 2021.