Generating high-accuracy and cloud-free surface soil moisture at 1 km resolution by point-surface data fusion over the Southwestern U.S.

Surface soil moisture (SSM) is of great importance in understanding global climate change and studies related to environmental and earth science. However, neither of current SSM products or algorithms can generate SSM with High spatial resolution, High spatio-temporal continuity (cloud-free and daily), and High accuracy simultaneously (i.e., 3H SSM data). Without 3H SSM data, fine-scale environmental and hydrological modeling cannot be easily achieved. To address this issue, we proposed a novel and integrated SSM downscaling framework inspired by deep learning-based point-surface fusion, which was designed to produce 1 km spatially seamless and temporally continuous SSM with high accuracy by fusing remotely sensed, model-based, and ground data. First, SSM auxiliary variables (e.g., land surface temperature, surface reflectance) were gap filled to ensure the spatial continuity. Meanwhile, the extended triple collocation method was adopted to select reliable in-situ stations to address the scale mismatch issue in SSM downscaling. Then, the deep belief model was utilized to downscale the original 9 km SMAP SSM and 0.1º. ERA5-Land SSM to 1 km. The downscaling framework was validated over three ISMN soil moisture networks covering diverse ground conditions in Southwestern US. Three validation strategies were adopted, including in-situ validation, time-series validation, and spatial distribution validation. Results showed that the average Pearson correlation coefficient (PCC), unbiased root mean squared error (ubRMSE), and mean absolute error (MAE) achieved 0.89, 0.034 m3m−3, and 0.032 m3m−3, respectively. The use of point-surface fusion greatly improved the downscaling accuracy, of which the PCC, ubRMSE, and MAE were improved by 3.73, 20.93, and 39.62% compared to surface-surface fusion method, respectively. Comparative analyses have also been carefully conducted to confirm the effectiveness of the framework, in terms of other downscaling algorithms, scale variations, and fusion methods. The proposed method is promising for fine-scale studies and applications in agricultural, hydrological, and environmental domains.