Evaluation of spectral features and regression methods for leaf area index estimation

Leaf Area Index (LAI) is critical for vegetation structure analysis and precision agriculture. Remote sensing technology enables rapid and non-destructive LAI estimation, yet the selection of optimal spectral features and regression models remains challenging. In this study, the optimal spectral features for LAI estimation and the robustness of regression methods across crop types, phenological periods and sensors were analyzed. Firstly, multiple categories of spectral features were compared. Then, six commonly used regression methods were evaluated for LAI estimation across diverse datasets. Lastly, robust spectral-regression combinations were identified. Hyperspectral data were processed to extract spectral features. Results demonstrated that regression methods performance varied significantly, with Partial Least Squares Regression (PLSR) and Gaussian Process Regression (GPR) showing superior stability (mean R2 > 0.68 across crops); sensor differences impacted spectral features more substantially than phenological stages; PCA_FDS_R (correlation-selected principal components of first-derivative spectra) was the most stable spectral feature; and the PCA_FDS_R + PLSR combination achieved optimal accuracy (R2 up to 0.75). This study concludes that PCA_FDS_R with PLSR regression method provides a transferable solution for multi-scenario LAI estimation, offering critical guidance for hyperspectral agricultural monitoring.