A Radiative Transfer-Driven Deep Learning Framework for Accurate Estimation of Rice Growth Parameters Using Multisource UAV Data

Leaf area index (LAI) and leaf chlorophyll content (LCC) are key indicators for monitoring rice growth dynamics. While UAV-based hyperspectral data is widely used, its high redundancy poses challenges for efficient information extraction. To address this, we propose a two-step generic framework. First, synthetic spectra generated by a field-constrained PROSAIL model are used to train a one-dimensional convolutional neural network with a self-attention mechanism that derives Spectral Composite Variables (SCVs) from redundant hyperspectral data. Then, the SCVs are combined with canopy temperature (from thermal infrared sensors) and crop height (derived from UAV-based LiDAR and RGB imagery) to develop a retrieval model, validated through both within-site and cross-site strategies. Results showed that the SCVs generated exhibited strong correlations with LAI and LCC, averaging 0.83 and 0.85, respectively. Moreover, the proposed framework achieved high retrieval accuracy across all growth stages (e.g., booting, heading, filling), with mean R² values of 0.76 for LAI and 0.71 for LCC. Specifically, both estimations reached peak performance during the heading stage, with an R² of 0.83 and RMSE of 0.47 m²/m² for LAI, and an R² of 0.77 and RMSE of 4.13 μg/cm² for LCC. Cross-site validation confirmed the model’s robustness and transferability, with the best performance consistently observed during the heading stage. Benefiting from this framework, spatial predictions of LAI and LCC at centimeter-level resolution closely aligned with observed patterns, enabling precise monitoring of rice growth. Overall, this study presents a robust and transferable solution for overcoming hyperspectral redundancy and enhancing crop growth estimation accuracy.