Soil moisture is a key variable influencing crop growth and irrigation management in arid agricultural systems. However, accurately monitoring soil moisture across scales remains challenging due to limitations in single-source remote sensing. This study integrates UAV multispectral imagery and Sentinel-2 data using two upscaling strategies: pixel-aggregation resampling and an enhanced TsHARP method incorporating a nonlinear gradient boosting tree (GBT) trend surface. Soil moisture inversion models were developed for multiple soil depths and growth stages using support vector machines, random forests, and XGBoost. The results show that TsHARP substantially outperformed resampling, increasing test-set R-2 by more than 0.20 in most cases. The XGBoost-TsHARP configuration achieved the highest accuracy, with R-2 reaching 0.671 for the 0- to 20-cm layer during winter wheat heading stage, while the 0- to 40-cm layer provided superior temporal stability and root-zone relevance. These findings demonstrate that transferring fine-scale spatial gradients from UAV data significantly enhances satellite-based soil moisture retrieval. The TsHARP-XGBoost approach offers an effective framework for operational soil moisture monitoring in arid and semi-arid irrigated regions.