The research on LAI estimation of multiple rice varieties based on multi-modal data fusion and machine learning algorithms

Accurate estimation of Leaf Area Index (LAI) in multi-variety rice using optical remote sensing remains challenging due to spectral saturation under dense canopy conditions and inter-varietal physiological differences. To address this, we developed a multimodal data fusion framework integrating RGB and multispectral imagery acquired by unmanned aerial vehicles (UAVs), combined with features derived from Digital Surface Models (DSM), vegetation indices (VIs), texture, and depth representations. Using field data collected across 60 rice varieties, four machine learning models were evaluated for LAI estimation. Our results demonstrate that multimodal fusion substantially outperforms conventional VI-based approaches. Among them, the Random Forest Regression (RFR) model achieved optimal performance ( R ²=0.76, RMSE=0.57), representing a 26–58% improvement in R ² over baseline models. SHAP-based feature importance analysis identified DSM feature, height-stratified vegetation indices, and depth features as key contributors to model accuracy. This study establishes that incorporating canopy structural information and deep features mitigates saturation effects and enhances generalizability across varieties. The proposed approach offers a robust and efficient solution for high-throughput LAI estimation, supporting applications in precision agriculture and rice breeding programs.