Resumen
Soil moisture is one of the critical variables in maintaining the global water cycle balance. Moreover, it plays a vital role in climate change, crop growth, and environmental disaster event monitoring, and it is important to monitor soil moisture continuously. Recently, Global Navigation Satellite System interferometric reflectometry (GNSS-IR) technology has become essential for monitoring soil moisture. However, the sparse distribution of GNSS-IR soil moisture sites has hindered the application of soil moisture products. In this paper, we propose a multi-data fusion soil moisture inversion algorithm based on machine learning. The method uses the Genetic Algorithm Back-Propagation (GA-BP) neural network model, by combining GNSS-IR site data with other surface environmental parameters around the site. In turn, soil moisture is obtained by inversion, and we finally obtain a soil moisture product with a high spatial and temporal resolution of 500 m per day. The multi-surface environmental data include latitude and longitude information, rainfall, air temperature, land cover type, normalized difference vegetation index (NDVI), and four topographic factors (elevation, slope, slope direction, and shading). To maximize the spatial and temporal resolution of the GNSS-IR technique within a machine learning framework, we obtained satisfactory results with a cross-validated R-value of 0.8660 and an ubRMSE of 0.0354. This indicates that the machine learning approach learns the complex nonlinear relationships between soil moisture and the input multi-surface environmental data. The soil moisture products were analyzed compared to the contemporaneous rainfall and National Aeronautics and Space Administration (NASA)?s soil moisture products. The results show that the spatial distribution of the GA-BP inversion soil moisture products is more consistent with rainfall and NASA products, which verifies the feasibility of using this experimental model to generate 500 m per day the GA-BP inversion soil moisture products.