Enhanced soil moisture estimation with GNSS-R and deep learning techniques

Soil Moisture (SM) is a critical variable for numerous Earth science applications, including agriculture, weather forecasting, and monitoring droughts and floods. Traditional methods for retrieving SM via remote sensing have predominantly used microwave active and passive satellites, such as NASA's Soil Moisture Active Passive (SMAP) mission, which offers a spatial resolution of 36 km and a revisit period of 2-3 days. Recently, Global Navigation Satellite System-Reflectometry (GNSS-R) has emerged as a promising technique for SM estimation, exemplified by missions such as Cyclone GNSS (CYGNSS). This study aims to improve the temporal and spatial resolution of SMAP by employing Convolutional Neural Networks (CNNs) trained on CYGNSS measurements, thus producing daily SM maps at a 9-km resolution. Using CYGNSS data from 2021 to 2023, the study area is divided into 9-km cells and a CNN model is trained for each cell. The primary inputs for these CNNs include the entire 2D Delay-Doppler Maps (DDMs) from CYGNSS, encompassing analog power, effective scattering area and bistatic radar cross-section. Additionally, the models incorporate CYGNSS incident angle, peak reflectivity and various ancillary geophysical parameters of the reflecting surface, such as the Normalized Difference Vegetation Index (NDVI), Vegetation Water Content (VWC), elevation, slope, surface water percentage, and soil clay and silt ratios. Our deep learning models are validated against the enhanced 9-km SMAP product and compared with the University of Luxembourg large-scale near-surface soil moisture product that, which uses a linear regression with SMAP as the reference. Unlike previous studies, we incorporate both ascending and descending SMAP half-orbits and extend our analysis beyond low-vegetation areas. Around 50300 models were trained using 80% of the data for training and the 20% for testing, resulting in each model being trained with 1016 data points on average. The models achieved a median unbiased Root Mean Square Error (ubRMSE) of 0.052 and a correlation coefficient (R) of 0.87 when compared with SMAP.