A global, spatially seamless, daily FY-3B soil moisture dataset based on a spatiotemporal deep learning model

Abstract. The Fengyun-3B (FY-3B) satellite provides an effective platform for monitoring soil moisture (SM) from regional to global scales, and the produced SM data constitute an important component among various SM products. However, the Fengyun-3B (FY-3B) SM product contains a large number of missing data mainly due to orbital gaps, which significantly limit its applicability. To address this issue, a parallel spatiotemporal reconstruction model that integrates local fine-grained features and global semantic representations was proposed, called GSP (multi-scale Gated Convolution-residual Shifted Window Transformer Parallel) model. The GSP model fully utilizes the spatiotemporal information of FY-3B SM, and leverages the complementary modules to enhance spatiotemporal feature representation. Specifically, the multi-scale gated convolutions are applied to focus on the irregular valid pixels and extract multi-scale local spatiotemporal features, while the residual Shifted Window Transformer (Swin transformer) is leveraged to capture global spatiotemporal texture. Based on GSP, a global, spatially seamless, daily FY-3B SM dataset from 12 July 2011 to 19 August 2019 was generated. In the experiments, the model was evaluated with two strategies: 1) real gaps, where in-situ data at the same geographical locations were used as reference data, and 2) simulated gaps, where the original FY-3B SM data were masked by orbital gaps, and originally known data were used as reference. The results based on both strategies indicate that the GSP-reconstructed FY-3B SM data present greater accuracy than three typical reconstruction methods. The ubRMSE based on 17 sparse and five dense in-situ networks is 0.0703 m³/m³ and 0.0631 m³/m³, respectively. This dataset can be downloaded at https://doi.org/10.6084/m9.figshare.30633548 (You et al., 2026).