Estimation of Leaf Area Index Using Radiative Transfer Process-Guided Deep Learning

The leaf area index (LAI) serves as a significant vegetation growth indicator and plays an essential role in vegetation's feedback to climate change. Currently, artificial intelligence (e.g., deep learning) algorithms possess strong capabilities in constructing complex relationships and demonstrate successful integration with remote sensing for LAI inversion. Among these algorithms, the long short-term memory (LSTM) network excels in handling sequence data and features a multi-layer nonlinear structure that effectively captures complex nonlinear relationships between vegetation canopy reflectance and LAI. However, previous researches mainly relied on the strong learning capabilities of LSTM without incorporating essential remote sensing knowledge, which led to the lack of process information guidance in the training stage. Consequently, the performance of the trained model may be significantly limited. In this letter, we proposed a process-guided LSTM (LSTM-PG) deep learning method for LAI estimation by integrating radiative transfer models. The constrained training dataset was generated using the Soil-Leaf-Canopy (SLC) model. We separately utilized the loss function of mean squared error (MSE) and a process-guided loss function to generate LSTM models for LAI predictions from the simulated SLC datasets. Subsequently, we validated the accuracy of the LAI retrieval models using field measurements from the ImagineS project. Our results indicated that the proposed process-guided (PG) method (R² = 0.79, RMSE = 0.87) performed better than the LSTM-MSE estimations (R² = 0.79, RMSE = 0.93). Additionally, statistical analyses across various scenarios demonstrated significant advantages of the proposed method, and the LSTM-PG predictions showed good spatial consistency with the LAI reference maps.