Development of a hyperspectral-based inversion model for cotton leaf SPAD measurement

Rapid and accurate chlorophyll monitoring is critical for precision cotton management, yet existing hyperspectral inversion methods often lack interpretability due to insufficient integration of physiological mechanisms. This study established a dual-pathway framework combining data-driven narrowband spectral index screening and mechanism-guided sensitive band extraction to estimate cotton canopy Soil Plant Analysis Development (SPAD) values using hyperspectral data (350–1075 nm). Field experiments across 20 plots with four water-nitrogen treatments collected 120 leaf samples during the flowering-boll stage. Three spectral preprocessing methods (original, first-derivative, continuum-removed) were systematically compared with three index types, namely Difference Spectral Index (DSI), Ratio Spectral Index (RSI), and Normalized Difference Spectral Index (NDSI), to identify optimal features. Correlation analysis revealed SPAD-sensitive regions in the green band and red-edge to near-infrared region, with first-derivative RSI (742, 952) achieving the highest correlation (r = 0.78). Mechanism-driven screening independently identified 742 nm as the core band (r = 0.76, p < 0.001), validating consistency between spectral features and chlorophyll absorption-scattering transitions. Among three machine learning algorithms, namely Partial Least Squares Regression (PLSR), Support Vector Regression (SVR), and Random Forest (RF), were evaluated. Random Forest demonstrated superior performance: the FDS-RSI-RF model achieved R² = 0.79 and RMSE = 2.0 SPAD units (3.4% relative error), while the mechanism-based 742 nm-RF model yielded R² = 0.86 and RMSE = 1.8 SPAD units. SHapley Additive exPlanations (SHAP) analysis confirmed that 742 nm contributed the highest feature importance, with values ranging from -4 to +4. This study validates the synergy between physiological mechanisms and data-driven approaches, providing a robust interpretable framework for non-destructive crop monitoring in precision agriculture.