Abstract
Accurate soybean yield estimation requires the effective fusion of solar-induced chlorophyll fluorescence (SIF), which is a key indicator of photosynthetic function, and structural information from reflectance-based vegetation indices. However, both SIF and vegetation indices are subject to Sun–sensor geometry, and robust multisource integration that exploits canopy directional signatures remains challenging. To address this issue, we propose M-Net, a physics-guided deep learning framework that explicitly incorporates canopy angular anisotropy for yield prediction. Instead of treating angular effects as noise to be normalized, M-Net incorporates bidirectional reflectance distribution function (BRDF)-derived angular features into an attention-based architecture to encode 3-dimensional canopy angular anisotropy as predictive information. Sensitivity analysis demonstrated that the accuracy of yield estimation decreased as Sun–sensor geometries deviated from the nadir, a trend particularly pronounced for physiological SIF signals susceptible to angular distortion. In comparative experiments, standard recurrent baselines (e.g., gated recurrent unit and long short-term memory) gained little from physics-reconstructed angular anisotropy features. In contrast, M-Net effectively used these anisotropic signatures and converted canopy directional anisotropy into predictive information. Validated across 613 US soybean-growing counties (2019 to 2023), M-Net achieved a high accuracy ( R 2 = 0.69). Notably, yield estimation using coarse-resolution Moderate Resolution Imaging Spectroradiometer data with BRDF-derived multiangular canopy anisotropy features ( R 2 = 0.64) outperformed that of high-resolution Sentinel-2 data with fixed viewing angles ( R 2 = 0.56), suggesting that effectively using angular information may be more important than spatial resolution alone. The results demonstrate that physics-guided deep learning transforms angular anisotropy from uncertainty into a valuable predictive signal, providing a practical approach for multisource crop monitoring.
Showing the abstract — retrieve the full paper via the Exa API.