Leaf area index (LAI) is a critical indicator bridging crop dynamic growth and agricultural management implementation. Different irrigation amounts and nitrogen (N) application rates influence crop growth and LAI. Accurate and dynamic LAI monitoring is essential for improving modern agricultural quality and efficiency. While unmanned aerial vehicles (UAVs) equipped with multispectral (MS) or thermal infrared (TIR) sensors can estimate LAI by extracting various features from remote sensing images, reliance on a single data source often limits estimation accuracy. To achieve more precise LAI estimation for cotton, this study acquired UAV-based remote sensing images—incorporating both MS and TIR data—across different growth stages under different irrigation amounts (60 % ET c , 80 % ET c and 100 % ET c , ET c denotes crop evapotranspiration) and N application regimes (0, 200, 300 and 400 kg N ha −1 ) in Xinjiang, China. The multi-feature and multi-dimensional information, including canopy coverage (CC), vegetation indices (VIs), canopy thermal information (CTs) and texture-related information (texture features, TFs; general texture indices, GTIs; three-texture indices, TTIs), was extracted from multi-source images. Five machine learning (ML) models, namely Support Vector Regression (SVR), Light Gradient Boosting Machine (LGBM), Random Forest (RF), Elman Neural Network (Elman), and Transformer, were adopted to estimate the LAI by utilizing the fused MS and TIR data, and the model with the optimal predictive performance was then employed to generate spatial distribution maps of LAI across the study area. The results indicated that texture indices enhanced LAI estimation, with TTIs demonstrating particularly strong potential. Multispectral data outperformed thermal data in standalone LAI estimation, while the integration of MS and TIR features greatly enhanced accuracy. Specifically, the Transformer model with CC + VIs + CTs + GTIs + TTIs as input variables obtained the best estimation accuracy (R 2 = 0.87, RMSE = 0.42, MAE = 0.36 for calibration; R 2 = 0.85, RMSE = 0.45, MAE = 0.37 for validation). The resulting LAI spatial distribution maps effectively characterized cotton growth dynamics under different irrigation and N treatments. These findings provided a reliable technical basis and practical guidance for the precision management of water and N in cotton fields. • Multi-feature and multi-dimensional data were used for cotton LAI estimation. • The novel three-texture indices from texture features estimated LAI efficiently. • Combining multispectral and thermal infrared data enhanced estimation accuracy. • Transformer model was optimal for estimation due to strong generalization ability.
Pei et al. (Thu,) studied this question.