A TCN-Transformer Parallel model for reconstruction of a global, daily, spatially seamless FY-3B soil moisture dataset

Abstract

Soil moisture (SM) is a critical variable in land-atmosphere interactions. As an important passive microwave remote sensing dataset, the Fengyun-3B (FY-3B) SM has been applied in a variety of scientific studies and applications. However, due to the discontinuous coverage of satellite revisit orbits, the FY-3B SM contains a large range of data gaps, which greatly limit the applicability. To solve this problem, we proposed a one-dimensional deep learning network-based time series reconstruction model called TTP (Temporal Convolutional Network (TCN)-Transformer Parallel) model, which makes full use of the TCN to capture short-term dynamic changes and the Transformer to obtain long-term dependencies, thus, extracting local and global features of the one-dimensional time series simultaneously. Based on the proposed TTP model, a global, daily, spatially seamless FY-3B SM dataset from 12 July 2011 to 19 August 2019 was generated. The performance of TTP was examined using two types of experiments: 1) in-situ data validation (the in-situ data at the same location were served as the reference); 2) original FY-3B SM validation (gaps were simulated by randomly masking the observations, with the original FY-3B SM as the reference). The TTP-based global, daily, spatially seamless dataset presents great consistency with the in-situ data, with an average root mean square error (RMSE) of 0.0900 m³/m³. Additionally, the reconstructed FY-3B SM based on TTP is consistently more accurate than four benchmarks (i.e., the self-supervised learning for interpolation (SSLI), the multivariate time series imputation method (MITST), the ModernTCN, and the harmonic analysis of time series (HANTS)). Moreover, the reconstructed FY-3B SM demonstrates reliable accuracy across various missing data simulated with different proportions and continuous lengths. The TTP can provide methodological support for large-scale remote sensing data reconstruction, and the generated dataset can provide data support for research in fields such as soil science and hydrology. The dataset is publicly available at https://doi.org/10.6084/m9.figshare.27957429.v45.