Daily Soil Moisture Retrieval by Fusing CYGNSS and Multi-Source Auxiliary Data Using Machine Learning Methods

Abstract

The Cyclone Global Navigation Satellite System (CYGNSS), a publicly accessible spaceborne Global Navigation Satellite System Reflectometry (GNSS-R) data, provides a new alternative opportunity for large-scale soil moisture (SM) retrieval, but with interference from complex environmental conditions (i.e., vegetation cover and ground roughness). This study aims to develop a high-accuracy model for CYGNSS SM retrieval. The normalized surface reflectivity calculated by CYGNSS is fused with variables that are highly related to the SM obtained from optical/microwave remote sensing to solve the problem of the influence of complicated environmental conditions. The Gradient Boost Regression Tree (GBRT) model aided by land-type data is then used to construct a multi-variables SM retrieval model with six different land types of multiple models. The methodology is tested in southeastern China, and the results correlate very well with the existing satellite remote sensing products and in situ SM data (R = 0.765, ubRMSE = 0.054 m3m−3 vs. SMAP; R = 0.653, ubRMSE = 0.057 m3 m−3 vs. ERA5 SM; R = 0.691, ubRMSE = 0.057 m3m−3 vs. in situ SM). This study makes contributions from two aspects: (1) improves the accuracy of the CYGNSS retrieval of SM based on fusion with other auxiliary data; (2) constructs the SM retrieval model with multi-layer multiple models, which is suitable for different land properties.