Transferability and prediction of AMSR2 brightness temperatures over snow-covered land based on AMSR-E brightness temperatures and machine learning algorithms

Tuesday, 16 December 2014
Barton A Forman, University of Maryland, College Park, MD, United States
Recent studies [Forman et al., 2013, IEEE; Forman and Reichle, 2014, IEEE] demonstrated the capability of machine learning (ML) algorithms to predict passive microwave (PMW) brightness temperatures (Tb) over snow-covered land as measured by the Advanced Microwave Sounding Radiometer (AMSR-E). The results presented here investigate the transferability of these techniques using AMSR-E to predict Tb observations as measured by AMSR2. In other words, can historical AMSR-E Tb observations be used to train a ML algorithm in order to predict Tb observations collected by AMSR2 at some point in the future? The NASA Catchment Land Surface Model is first used to characterize snowpack conditions. Next, the ML algorithm is trained on the 9-year record of PMW Tb observations collected by AMSR-E. An additional experiment where the ML algorithm was trained on a split-sample of the 2-year record of AMSR2 PMW Tb observations was conducted for comparison. Results suggest one ML technique – the support vector machine – when trained on AMSR-E observations can sufficiently reproduce AMSR2 Tb in forested and non-forested regions during both the snow accumulation and snow ablation phases of the snow season. These results suggest transferability of machine learning from the AMSR-E sensor to other data records with comparable frequency and polarization characteristics. The eventual goal is to use a ML algorithm as an observation operator within an ensemble-based data assimilation framework where model estimates will be merged with PMW Tb observations in order to improve snow water equivalent (SWE) estimates across regional and continental scales.