Prediction of Soil Moisture From Near-Global Cygnss Gnss-Reflectometry Using a Random Forest Machine Learning Model

We developed a random forest model to predict soil moisture at the field scale (∼1 km), combining 6.5 years of CYGNSS Global Navigation Satellite System Reflectometry (GNSS-R) data with observations from 1049 gauges which are part of the International Soil Moisture Network. Several relevant predictor variables were developed, including geophysical parameters related to soil properties, topography, land cover and climatology, and dynamic variables including precipitation from the Global Precipitation Mission and evapotranspiration derived from MODIS temperature. The model achieved good accuracy for predicted soil moisture (R
²
= 0.68; RMSE = 0.064 mm/mm), which was further improved through postprocessing using a linear regression model of the residuals for bias correction (R
²
= 0.82; RMSE = 0.048 mm/mm). Our work demonstrates the potential for field-scale prediction of soil moisture from GNSS-R data, which may be further improved through enhanced co-variates, and is applicable to recent and near-future GNSS-R missions such as Rongowai and HydroGNSS.