Quantifying Uncertainty in Machine Learning Based Soil Moisture Retrieval From GNSS-R Measurements

Spaceborne platforms like the NASA SMAP and the ESA SMOS offer global-scale coarse-resolution observations that can be used to estimate surface soil moisture. However, they acquire observations at a moderate revisit frequency, reducing the impact on applications that require higher temporal resolutions. Global navigation satellite system reflectometry (GNSS-R) is a prime example of a signal-of-opportunity (SoOP) which has been shown to be highly sensitive to variations in soil moisture. In this work, we consider measurements from the NASA Cyclone GNSS (CYGNSS) mission which consists of eight low-orbit observatories, that are equipped with two nadirlooking antennas to receive reflected GPS signals. To extract soil moisture content, the traditional approach involves inverting parameterized forward models which account for aspects like dielectric properties of the soil, topography, and incidence angle information among others. Despite their potential, these models typically involve simplified assumptions and demonstrate increased sensitivity to the values of different parameters. To address this challenge, a new line of research tries to address this challenge by utilizing machine learning models and treating the problem as that of supervised regression.