Machine Learning-Based Soil Moisture Retrieval Using CYGNSS and Interpolated SMAP Data

Abstract

Soil moisture is a crucial component of the global terrestrial ecosystem water vapor cycle, and higher spatial-temporal soil moisture product has significant impacts on research in agriculture, hydrology, and ecology. Spaceborne Global Navigation Satellite System Reflectometry (GNSS-R) is a new remote sensing technology which can be used to retrieve soil moisture. Cyclone Global Navigation Satellite System (CYGNSS), launched by NASA in 2016 to observe ocean surface hurricanes, has also collected a substantial amount of data of GNSS signals reflected over the land. This paper focuses on soil moisture retrieval using CYGNSS and SMAP dataset. In this study, three machine learning methods (GSVM, BP and RF) are used for spatial interpolation of SMAP soil moisture data. Result shows RF interpolation is the best with the RMSE of 0.044 cm3/cm3 and the CC of 0.959 compared with SMAP truly SM. Then soil moisture data after RF interpolation and conventional linear interpolation are respectively used as response variables to train the retrieval model by ELM, BP and RF machine learning methods respectively. Results show that, among the six regression models, the RF-RF regression model performs the best with the RMSE of 0.012 cm3/cm3 and the CC of 0.986. Using this model, we get a global soil moisture product and have higher temporal and spatial resolution.