Downscaling SMAP Soil Moisture to 250 m Using Deep Learning Models and Multi-Source Environmental Variables over the Loess Plateau

Soil moisture (SM) is a key variable in land–atmosphere interactions, but the coarse spatial resolution of existing SM products limits their applications in regional hydrological studies over heterogeneous terrains. This study developed and evaluated a progressive deep-learning (DL) downscaling framework (SE-ResNet) to downscale SMAP L4 SM from 9 km to 250 m over the Loess Plateau, using eight multi-source environmental predictors. The framework was systematically compared with traditional machine learning (ML) models (XGBoost and RF) and its standard counterparts (CNN, ResNet). Results showed that actual evapotranspiration, DEM, and precipitation were identified as the dominant factors controlling SM variability. DL models generally outperformed traditional ML models on the 9 km test dataset, with SE-ResNet achieving the highest accuracy (R = 0.902, RMSE = 0.026 m3/m3). However, in-situ validation showed that 250 m downscaled products did not consistently outperform the original 9 km SMAP SM, with substantial performance differences among stations, but SE-ResNet achieved the lowest Bias among all models. Monthly-scale error analysis further revealed pronounced temporal variations in model prediction uncertainty. Multi-resolution training further demonstrated that the architecture maintains consistently high but gradually decreasing performance when retrained at target resolutions (R = 0.797, 0.785, and 0.772 at 3 km, 1 km, and 250 m, respectively), confirming the presence of scale effects. These findings emphasize the need to consider environmental controls and scale-adaptive strategies when generating high-resolution SM products over complex terrains.