WCD-20. Impacts of machine learning for selective data thinning in data assimilation

Abstract
Recent increases in the volume and complexity of environmental information through satellite observations means it has become increasingly challenging to extract and utilize meaningful information in real time from the high density information that is being received. Improvements in the usage of satellite data can result in an improved analysis, and consequently forecast, of NWP models. An existing machine learning (ML) algorithm can identify areas with active and rapidly evolving weather. We will show the process for transforming the ML output, which is given by global image segmentation of pixels with active weather or with no active weather, into the selective filters for thinning that are usable by JEDI, the next generation data assimilation software system. We also present preliminary results for a case study showing that a 3D-EnVar data assimilation algorithm is sensitive to selective thinning of satellite data in the ML-identified regions, and results indicating that global forecasts made from the analysis are also sensitive to this selective thinning.