Enhanced UAV-based SPAD values estimation in tea chrysanthemum: an optimized and interpretable machine learning approach integrating spectral and textural information

Accurate estimation of chlorophyll content, often represented by soil and plant analyzer development (SPAD) values, is essential for monitoring crop health and guiding precision agriculture. Unmanned aerial vehicle (UAV) remote sensing has emerged as a promising tool for rapid, non-destructive SPAD values estimation, yet many existing models suffer from limited generalizability and low interpretability—hindering their practical application across varying conditions and cultivars. To address these challenges, we conducted a two-year field study (2022–2023) on tea chrysanthemums, acquiring UAV-based multispectral imagery alongside ground-truth SPAD measurements. Spectral reflectance was used to calculate vegetation indices (VIs), and texture features from the gray level co-occurrence matrix (GLCM) were used to construct texture indices (TIs). Key features were selected using spearman correlation combined with the least absolute shrinkage and selection operator (LASSO) algorithm. We then developed machine learning (ML) models fusing both VIs and TIs, with hyperparameters optimized via Optuna framework. Model interpretability and feature interactions were examined using shapley additive explanations (SHAP). Results indicated LASSO selected an optimal set of VIs (normalized green-blue difference index (NGBDI), modified simple ratio (mSR), transform chlorophyll absorption in reflectance index / optimized soil adjusted vegetation index (TCARI/OSAVI), chlorophyll normalized vegetation index (NPCI)) and TIs (e.g., re-normalized difference texture index (RDTI(MEA 475 , MEA 560 ))). Models fusing spectral and texture indices demonstrated improved accuracy over single-feature models. The Optuna- VITIs -XGBoost model performed best (R V 2 = 0.847, an improvement of approximately 4.3%; RPD = 2.574, RMSE V/C = 1.409), exhibiting optimal generalization. SHAP analysis revealed NGBDI, mSR, and RDTI(MEA 475 , MEA 560 ) as the most significant contributors to SPAD values estimation. Interactions between normalized difference texture index (NDTI(MEA 475 , MEA 560 )) and VIs enhanced VIs’ impact, improving model interpretability. The model showed stable performance across different years and nitrogen (N) levels, and multi-variety mapping demonstrated a degree of universality. Our findings demonstrate that fusing spectral and texture features improves UAV-based SPAD values estimation accuracy, offering a robust, interpretable, and generalizable approach for high-throughput phenotyping in tea chrysanthemums.