Estimation of cotton leaf area index from UAV multispectral imagery by integrating vegetation indices and texture features

Objective This study aimed to develop a practical and interpretable unmanned aerial vehicle (UAV) multispectral framework for estimating cotton leaf area index (LAI) in film-mulched, drip-irrigated fields in arid regions. Methods Five-band UAV multispectral imagery and ground LAI measurements were collected at four growth stages in Shihezi, Xinjiang, China, in 2024. Vegetation indices and gray-level co-occurrence matrix texture features were extracted and combined. Repeated Boruta selection was used to identify consistently selected variables. Six regression models were evaluated using vegetation-index, texture, and fused feature sets, and SHapley Additive exPlanations (SHAP) was used to interpret the selected fused-feature BP model. Results Fused features generally outperformed single-source features across the pooled dataset, indicating complementary predictive information from vegetation indices and texture features. Boruta identified 12 core features that were consistently selected across all 15 runs. The fused BP and RFR models showed comparable predictive performance. The fused BP model achieved mean test-set values of R² = 0.86, RMSE = 0.20, and NRMSE = 7.33%, and was selected for subsequent spatial LAI inversion because of its smaller training–test performance gap. SHAP analysis indicated complementary contributions from red-edge and near-infrared vegetation indices and visible-band texture features. Spatial inversion using the fused model illustrated the seasonal pattern of cotton canopy development across the four growth stages within the experimental field. Conclusion Integrating spectral and texture features with feature selection improved the predictive performance and interpretability of cotton LAI estimation. The proposed workflow provides an interpretable approach for field-scale cotton growth monitoring and may serve as a canopy-status information layer for subsequent site-specific diagnosis and precision-management applications in arid production systems.