PNNL

Abstract

Accurate crop yield estimations require the combination of physiological indicators such as the Solar-Induced chlorophyll Fluorescence (SIF) and plant structural and biochemical indicators, such as Vegetation Indices. However, integrating these remote sensing signals from multi-source and multi-angle satellite observations (e.g., TROPOMI, MODIS, Sentinel-2 (S2)) to improve crop yield prediction remains challenging. To address this challenge, we propose a physics-guided deep learning framework (M-Net). Unlike traditional models, which often treat angular effects as noise to be corrected, M-Net incorporates the BRDF principles to actively utilize angular signatures as predictive features. This approach enables the mechanistic deconvolution of physiological and structural information from multi-source time-series data. The results show that integrating multi-source data can be highly beneficial. Combining structural VIs from both MODIS and S2 with physiological data increased the model performance to a median R² of 0.69, a substantial improvement over the MODIS single-source configuration (R² = 0.57). Furthermore, the multi-angle model configuration (R² = 0.58) outperforms signal-angle inputs (R² = 0.53). The final M-Net model achieved high accuracy in county-level soybean yield estimation (R²=0.73), outperforming a Random Forest benchmark (R²=0.58) and mitigating saturation at yield extremes. Further analysis identified the reproductive growth period of soybean as the critical predictive window, during which canopy structural VIs from integrated MODIS and S2 data emerged as the most important predictors. This study developed a novel method for robust yield forecasting by treating angular signatures as a signal rather than noise, enabling an innovative disentanglement of biophysical information.