Uncertainty quantification of a radiative transfer model and a machine learning technique for use as observation operators in the assimilation of microwave observations into a land surface model to improve soil moisture and terrestrial snow

Abstract

Soil moisture and terrestrial snow mass are two important hydrological states needed to accurately quantify terrestrial water storage and streamflow. Soil moisture and terrestrial snow mass can be measured using ground-based instrument networks, estimated using advanced land surface models, and retrieved via satellite imagery. However, each method has its own inherent sources of error and uncertainty. This leads to the application of data assimilation to obtain optimal estimates of soil moisture and snow mass. Before conducting data assimilation (DA) experiments, this dissertation explored the use of two different observation operators within a DA framework: a L-band radiative transfer model (RTM) for soil moisture and support vector machine (SVM) regression for soil terrestrial snow mass. First, L-band brightness temperature (Tb) estimated from the RTM after being calibrated against multi-angular SMOS Tb's showed good performance in both ascending and descending overpasses across North America except in regions with sub-grid scale lakes and dense forest. Detailed analysis of RTM-derived L-band Tb

in terms of soil hydraulic parameters and vegetation types suggests the need for further improvement of RTM-derived Tb in regions with relatively large porosity, large wilting point, or grassland type vegetation. Secondly, a SVM regression technique was developed with explicit consideration of the first-order physics of photon scattering as a function of different training target sets, training window lengths, and delineation of snow wetness over snow-covered terrain. The overall results revealed that prediction accuracy of the SVM was strongly linked with the first-order physics of electromagnetic responses of different snow conditions. After careful evaluation of the observation operators, C-band backscatter observations over Western Colorado collected by Sentinel-1 were merged into an advanced land surface model using a SVM and a one-dimensional ensemble Kalman filter. In general, updated snow mass estimates using the Sentinel-1 DA framework showed modest improvements in comparison to ground-based measurements of snow water equivalent (SWE) and snow depth. These results motivate further application of the outlined assimilation schemes over larger regions in order to improve the characterization of the terrestrial hydrological cycle.