Enhancing SMAP Soil Moisture Retrievals via Superresolution Techniques

Soil moisture is a key state variable that modulates land-atmosphere interactions and its high-resolution global scale estimates are essential for improved weather forecasting, drought prediction, crop management, and the safety of troop mobility. Currently, NASA's Soil Moisture Active/Passive (SMAP) satellite provides a global picture of soil moisture variability at a resolution of 36 km, which is prohibitive for some hydrologic applications. The goal of this research is to enhance the resolution of SMAP passive microwave retrievals by a factor of 2 to 4 using modern superresolution techniques that rely on the knowledge of high-resolution land surface models. In this work, we explore several super-resolution techniques including an empirical dictionary method, a learned dictionary method, and a three-layer convolutional neural network. Using a year of global high-resolution land surface model simulations as training set, we found that we are able to produce high-resolution soil moisture maps that outperform the original low-resolution observations both qualitatively and quantitatively. In particular, on a patch-by-patch basis we are able to produce estimates of high-resolution soil moisture maps that improve on the original low-resolution patches by on average 6% in terms of mean-squared error, and 14% in terms of the structural similarity index.