Abstract
Leaf water content (LWC) is a core indicator for precise irrigation decision-making in fruit orchards, which is critical to water use efficiency improvement and stable yield maintenance in arid oasis agricultural systems. In this study, leaf water content (LWC) and synchronously acquired unmanned aerial vehicle (UAV) multispectral imagery were used as data sources. A band combination optimization method was adopted to systematically construct all possible two-band and three-band indices (TBIs). Through correlation analysis with LWC, the variations in the sensitivity of spectral indices in different dimensions (two-band and three-band) to LWC were investigated. On this basis, three machine learning algorithms—Random Forest Regression (RFR), Support Vector Regression (SVR), and Quantile Polynomial Regression (QPR)—were employed to develop LWC estimation models, and the optimal model was selected through comparative analysis. The results showed the following: (1) For the three fruit tree species, the three-band spectral indices (TBI1 and TBI8) exhibited higher correlations with LWC compared with the two-band spectral indices, and the optimal band combinations were mainly located in the red-edge (RE) region. (2) The estimation model based on the TBI1 spectral index combined with the RFR algorithm achieved the best performance, with the validation set coefficient of determination (R2) reaching 0.90–0.94 and the lowest root-mean-square error (RMSE) of 3.18%–3.69%. The above results indicate that the RFR algorithm constructed based on three-band spectral indices is more suitable for addressing nonlinear problems and can effectively improve the estimation accuracy of LWC.