Toward an Enhanced Surface Classification Within MiRS Using a Deep Neural Network Approach

The Microwave Integrated Retrieval System (MiRS) is an operational microwave-only satellite retrieval system, developed at NOAA, that currently processes passive microwave measurements from 10 different low earth orbit satellites. It implements a one-dimensional variational approach to simulate radiances and to determine the atmospheric state which best fits the observed radiometric measurements, subject to geophysical a priori constraints and assumptions.

The extent and degree of surface snow and ice cover can have a great impact on MiRS’ retrievals over the cryosphere, as low-frequency microwave channels are highly sensitive to surface conditions. In addition, some of the a priori constraints within the MIRS variational retrieval are dependent on the surface preclassification. In this presentation, we discuss an extension of MiRS’ treatment of the Earth’s surface from categorical (i.e. water, land, ice, etc.) to a continuum of states. This improves MiRS’ ability to retrieve cluttered and complex scenes, which may represent situations like partial snow cover or mixed sea ice/ocean. The extent to which an observation is in a particular state is determined from a deep neural net-based retrieval, implemented using TensorFlow. The neural net is trained using data from Suomi NPP’s ATMS and VIIRS instruments. S-NPP launched in October 2011, orbits at an inclination of 98.7°, and has yielded a lengthy dataset with excellent polar coverage. The neural net correlates ATMS’ microwave channels with the well-established VIIRS snow cover algorithm, which uses optical channels. Further algorithmic improvements show potential for operational implementation for sea ice detection in cloudy conditions. Additionally, we compare retrieval results with similar NOAA-20 observations and discuss the possibility of a synergistic operational product.