Estimation of rice leaf angle distribution from geometric features in millimetric-resolution UAV RGB imagery acquired at 45° viewing angle

Leaf angle distribution (LAD) is a fundamental determinant of canopy light interception and photosynthetic efficiency, yet its robust quantification remains a major bottleneck. Traditional methods are constrained by either labor-intensive manual measurements or the ill-posed nature of indirect gap probability inversion. Recent advances in deep learning have demonstrated great potential for directly extracting canopy traits from high-resolution imagery. However, whether 2D RGB imagery contains sufficient information to retrieve LAD, a three-dimensional canopy structural trait, is still questionable. We investigated this question by developing a deep learning framework that establishes the relationship between image-derived features and LAD from millimetric-resolution RGB images taken from a UAV. We focused on rice crops representing one of the main cultivated species. A total of 332 rice plants were digitized to construct realistic 3D canopies; we generated 21,600 paired images and GT LAD labels via physically based rendering. In silico experiments demonstrated that deep learning models enable efficient LAD estimation, particularly when taking images inclined at 45° perpendicularly to the rows. Further, our model achieved substantially higher accuracy for average inclination angle (AIA, R 2 = 0.608, RMSE = 1.590°) compared to gap probability-based inversion (R 2 = −0.788, RMSE = 3.397°). Finally, our model showed superior performance from the transplanting to booting stage (R 2 avg > 0.94 for LAD) when applied to actual observations. Application to over 1000 cultivars across two years confirmed high repeatability, establishing a scalable solution for large-scale monitoring. Interpretability analyses, supported by feature visualization and ablation studies, reveal that the model explicitly extracts geometric features specific to leaf inclination, while disentangling them from confounding canopy variables in millimetric-resolution RGB imagery. This feature-level discrimination underpins the model's superior accuracy and ensures robust generalization in diverse field scenarios.