Soil moisture content (SMC) is vital for agriculture and water management, but accurate monitoring remains challenging. Current methods rely on standalone physical models (e.g., Water Cloud Model) or machine learning (ML), but physical models lack accuracy in complex environments, while ML lacks physical constraints. We hypothesize that hybrid models can improve SMC estimation. Using 90 soil samples from arid oasis ecosystems, we compared four approaches on Google Earth Engine: (1) physical model (M1), (2) ML models (M2: RF/GBRT), and hybrid models (M3/M4) integrating both. M3 used physical model outputs as ML inputs, while M4 added the physical-in situ difference as features. Results showed standalone models had limited accuracy (R 2: 0.129 for M1; 0.239-0.419 for M2), whereas hybrid models improved performance significantly. The best model (M4-GBRT) achieved the highest accuracy (R 2 = 0.554, RMSE = 0.035), significantly outperforming standalone models. The superiority of the M4 scheme is attributed to its residual learning mechanism, which incorporates the difference between physical model estimates and in situ measurements as a feature. This allows the algorithm to correct systematic biases inherent in the simplified physical model and capture complex nonlinear interactions in heterogeneous environments. This approach effectively overcoming the accuracy bottlenecks of physical models and the lack of constraints in ML models. Ultimately, this study validates a robust strategy suitable for regional-scale soil moisture mapping in arid oases constrained by limited field data, providing a scientific basis for precision water management and food security.