Seamless global daily soil moisture mapping using deep learning based spatiotemporal fusion

Abstract

Soil moisture products with long-term, high spatial continuity, and high accuracy are essential for meteorological management and hydrological monitoring. Microwave remote sensing retrieval and land surface model simulation are the two primary sources of global-scale soil moisture data, but each has inherent limitations, making it difficult to balance accuracy and spatial coverage. In this paper, to tackle this challenge, we propose a deep learning-based spatiotemporal fusion framework to integrate the two data sources and generate a global soil moisture product with high spatial continuity and accuracy. Specifically, we leverage the high accuracy of the Soil Moisture Active and Passive (SMAP) microwave soil moisture data and the spatiotemporal continuity of the Noah assimilation soil moisture data. The proposed model employs a deep residual cycle GAN (DrcGAN) to capture the nonlinear complementary spatiotemporal features between the SMAP and Noah data, generating a seamless global daily product at a 36-km resolution, spanning April 4, 2015, to November 26, 2023, referred to as STSG-SM. Various validation methods, including spatial pattern analysis, time-series comparison, and in-situ validation, are utilized to assess the effectiveness and reliability of the product. In comparison to the selected in-situ measurements, the STSG-SM dataset (original SMAP-P₃₆) exhibits a bias of 0.0230 m³/m³ (0.0243 m³/m³), R of 0.8388 (0.8405), RMSE of 0.0629 m³/m³ (0.0628 m³/m³), and ubRMSE of 0.0585 m³/m³ (0.0579 m³/m³), indicating that the proposed method sustains the high precision of satellite-retrieved soil moisture and demonstrates strong consistency with the in-situ measurements.