Reconstruction of Spatially Seamless Soil Moisture Data from SMAP Product Using Deep Learning

This paper proposes a deep learning method for spatial data imputation, utilizing a CNN with residual connections to construct a regression model for missing value prediction. The introduction of temporal features from SMAP ensures seamless spatial characteristics for soil moisture. By integrating various spatially seamless, multi-source remote sensing data as auxiliary variables, the method predicts and fills the spatial gaps in SMAP SSM. The result is a long-term, spatially seamless daily soil moisture dataset with a 9km resolution, covering the period from January 1, 2016, to December 31, 2022. The effectiveness of this spatial imputation method is validated by comparing and analyzing it with site-measured data and original SMAP data. This approach provides a solution for generating spatially seamless soil moisture data.