A Physics-Informed Spatiotemporal Deep Learning Algorithm for High-Resolution Leaf Area Index Retrieval

High-resolution leaf area index (LAI) retrieval is crucial for ecological and agricultural applications, yet it remains challenging due to the enhanced spatial heterogeneity and limited temporal observations of high-resolution imagery. Although combining physics-based models with deep learning is a promising approach, existing approaches are limited by the restricted application scope of physical models and the lack of deep learning networks that can capture both spatial and temporal information. This study develops an image-to-image algorithm, termed UM-RCA (Unified BRDF Model-informed ResNet-ConvLSTM-Attention), that integrates a physical model with a new spatiotemporal deep learning network to generate 30 m LAI products using Landsat imagery. The method first applies a unified bidirectional reflectance distribution function model (UN-MODEL), suitable for both continuous and discrete canopies, to generate a high-fidelity training dataset. A ResNet ConvLSTM-Attention (RCA) network is then trained on the dataset to learn the relationships between Landsat reflectance and LAI patch time series. The three components of the RCA network combine spatial, temporal, and spectral information through an “Extract–Aggregate–Refine” structure. Results show that the UM-RCA algorithm achieves high accuracy on LAI field measurements across different biome types (overall MAE = 0.510, RMSE = 0.685) and produces more spatiotemporally coherent LAI maps than three mainstream algorithms or products: the Random Forest model, the Sentinel Application Platform (SNAP) biophysical processor, and the High-resolution Global LAnd Surface Satellite (Hi-GLASS) LAI product. The UM-RCA algorithm offers significant practical value for precision agriculture and ecological monitoring.