UAV multispectral data and mixed-pixel segmentation for citrus water deficit diagnosis: Integrating spectral–texture features with machine learning for precision irrigation management

Accurate vegetation water deficit diagnosis is critical for optimizing agricultural water allocation, simulating hydrological processes, and sustaining agroecosystem stability. This study developed a novel citrus water status monitoring strategy by integrating UAV multispectral imagery and Gaussian mixture model based mixed-pixel segmentation algorithm, which can efficiently extract citrus canopy structural information and efficiently eliminate the spectral interference from bare soil. Based on the experimental datasets from 2023 to 2024, four machine learning models, namely Extreme Learning Machine (ELM), Support Vector Machine (SVM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost), were established to estimate citrus water content during four critical growth stages. The predictive performance of these models was compared under two input schemes: fused spectral-texture features, and RGB texture features alone. The results demonstrated that under the spectral-texture feature fusion scheme, the RF model achieved the highest prediction accuracy, followed by XGBoost, ELM, and SVM, with GPI ranges of 0.340∼1.485, 0.212∼0.745, −0.745∼-0.212, −2.660∼-1.515, respectively. For the RGB-only texture feature scheme, the XGBoost model performed best at the young fruit stage, followed by RF and ELM, while SVM exhibited the poorest accuracy, with GPI values of 0.878, 0.321, −0.962, −1.429, respectively. In the other three growth stages, the RF model still achieved superior performance compared with other models, followed by XGBoost, ELM and SVM, with GPI of 0.428∼1.379, 0.330∼1.495, −2.572∼-1.505 and −0.885∼-0.330, respectively. This study enables quantitative canopy-scale assessment of citrus water status, providing a reliable technique for monitoring water deficit and supporting efficient agricultural water management.