Spatiotemporal Reconstruction of FY-3B Soil Moisture Using a Hybrid Attention and Partial Convolution Neural Network

China's FY-3/4 satellite constellation has significantly advanced global environmental monitoring capabilities. However, the inherent temporal resolution limitations of polar-orbiting satellites result in discontinuous spatiotemporal coverage of current soil moisture products, thereby constraining their application in global hydrological modeling. To overcome this challenge, this study introduces a dedicated spatiotemporal reconstruction model that enhances the spatial continuity of FY-3B satellite soil moisture datasets. The proposed model utilizes a dual-channel input architecture integrating continuous observation data with dynamic mask matrixand leverages partial convolution for effective collaborative spatiotemporal feature extraction. Contextual attention and multi-layer transformer encoder are incorporated to generate seamless global daily soil moisture product (2010-2019). Validation indicate notable improvements in accuracy: (1) in-situ validation increased the mean correlation coefficient from 0.582 to 0.671 and reduce the root mean square error from 0.147 m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>/m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> to 0.143 m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>/m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>; (2) temporal consistency analysis confirms that the reconstructed sequence remains highly synchronized with the original soil moisture products; (3) simulated missing region experiments yielded a coorelation of 0.953 with the original products, with a bias as low as -0.004 m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>/m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> and an unbiased root mean square error of 0.006 m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>/m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>. Compared to traditional partial convolution methods, this approach enhances global accuracy by increasing the correlation by 23.8%. Notably, this research marks the first implementation of a hybrid attention-partial convolution deep learning model to generate a seamless global daily soil moisture product derived from Fengyun satellites, effectively addressing previous reconstruction limitations.