Intelligent diagnosis of winter wheat water stress based on UAV multi-modal remote sensing

Water stress severely affects the normal growth and development of winter wheat, resulting in a decrease in both yield and quality. To address the limitations and time lags of traditional water stress monitoring methods, this study develops a winter wheat water stress classification model using Unmanned Aerial Vehicle (UAV)-based multi-source remote sensing data and machine learning algorithms. The experiment was conducted at the Xinxiang Experimental Base of the Chinese Academy of Agricultural Sciences in Henan Province, China, during the winter wheat growing season, involving 120 plots under four irrigation levels. Vegetation indices (VIs) and thermal infrared (TIR) features were extracted by acquiring water stress imagery during three critical growth stages of wheat: heading stage, flowering stage, and filling stage. Unlike many previous UAV-based water stress studies that mainly focus on a single growth stage or a single type of sensor, this study evaluates VI and TIR fusion across multiple key growth stages under controlled irrigation gradients and systematically compares three machine-learning classifiers. The performance of three classifiers, Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM), in water stress identification was evaluated. The results indicated that classification performance varied across single growth stages. Notably, multi-source sensor data fusion consistently achieved the highest classification accuracy across all stages, outperforming single-source data. The SVM model achieved classification accuracy, recall, and F1 score all exceeding 91%, with the flowering stage being particularly outstanding, achieving an accuracy of 97.22%, precision of 97.78%, recall of 97.22%, and F1 score of 97.27%. After merging multiple growth stages, the fusion of multi-source sensor data still showed the best performance. The SVM model had the highest classification accuracy (82.41%), precision (83.71%), recall (82.41%), and F1 score (81.93%). In summary, the SVM classifier fusing VIs and TIR features exhibited excellent performance in classifying wheat water stress. This study effectively achieved non-destructive, accurate, and rapid monitoring of winter wheat water stress, providing technical support for precision agriculture and water management.