A physics-informed deep learning method that forecasts crop yield loss by embedding the agronomic equation for crop yield response to water scarcity directly into the training objective. An LSTM-based recurrent network estimates pixel-level crop water use (actual evapotranspiration) and the crop's sensitivity to water stress from Sentinel-2 satellite imagery and climate data, using an enhanced physics-constrained loss so predictions are both accurate (R2 up to 0.77, matching RNN and Transformer baselines) and physically consistent and interpretable.
This work forecasts crop yield loss under water scarcity by combining process-based agronomy with machine learning. It embeds the crop yield response-to-water-scarcity equation directly into a neural…