Synchronously considering spatial dependence and temporal persistence into a deep learning based method for improving soil moisture spatial resolution

Accurate high-resolution soil moisture (SM) data are essential for understanding land–atmosphere interactions and supporting hydrological applications. However, soil moisture dynamics are inherently shaped by coupled spatial heterogeneity and temporal memory effects, requiring modeling frameworks that explicitly integrate spatial convolution and long-term temporal dependencies. In this study, we developed a hybrid convolutional long short-term memory (CNN–LSTM) framework to downscale the ESA Climate Change Initiative (CCI) SM dataset over the Huai River Basin, China. Multi-source environmental predictors, including meteorological, vegetation, and topographic factors, were incorporated, and both ESA CCI and in situ observations were used as dual constraints during model training. The results showed that precipitation, NDVI, and slope were the most influential predictors for SM downscaling. Among the tested models, the CNN–LSTM achieved the best performance (R ≈ 0.79, RMSE = 0.04 m3/m3), outperforming single CNN and LSTM models. The downscaled 5 km product captured the spatial and seasonal patterns of SM more accurately than the original ESA CCI, ERA5, and SMAP datasets, with the highest correlation and lowest estimation error. Nevertheless, further downscaling to 1 km introduced additional noise and performance degradation, indicating an optimal resolution around 5 km. Overall, the proposed framework provides a robust and transferable approach for generating high-resolution SM datasets with improved physical consistency, offering valuable potential for basin-scale hydrological modeling and climate studies in data-limited regions.