Accurate estimation of soil moisture content (SMC) serves as a crucial foundation for smart irrigation implementation and precision agriculture management. However, traditional methods relying on single data sources or individual machine learning models often suffer from limited generalization ability and insufficient accuracy. To address these challenges, this study utilized Unmanned Aerial Vehicle (UAV) remote sensing to acquire multi-source data (RGB, Multispectral, and Thermal Infrared) across three critical growth stages of winter wheat. We systematically evaluated the perfor-mance of six machine learning algorithms and Stacking ensemble learning strategy for estimating SMC at different soil depths (0-60 cm). The results demonstrated that fusing multi-source data consistently enhanced SMC estimation accuracy compared to single-source data across all growth stages. Temporally, the milk-ripe stage exhibited the strongest correlation with SMC, making it the optimal phenological phase for surface moisture retrieval. During the key filling stage, the XGBoost model combined with fused data (MS + RGB + TIR) achieved the best performance for surface soil (0-20 cm) with an R2 of 0.73 and RRMSE of 0.06. In contrast, the GPR model exhibited poor performance in most cases. Although estimation accuracy decreased with soil depth, the fusion approach maintained acceptable results in deeper layers (0-40 cm and 0-60 cm). Furthermore, the Stacking ensemble strategy effectively overcame the limitations of single models, the performance of combinations of different base models and secondary models varied. Specifically, the ensemble model employing Support Vector Regression (SVR) as the secondary learner yielded the highest overall accuracy (R2 = 0.76, RRMSE = 0.06). These findings provide a theoretical basis and a robust technical reference for optimizing data fusion and model selection in the precision irrigation management of dryland winter wheat fields.