Impacts of Machine Learning for Targeted Observation Thinning in Satellite Data Assimilation

The volume and complexity of environmental information received via satellite observations has increased rapidly in recent years. This dramatic increase in the density of available data results in many challenges for the numerical weather prediction (NWP) community to extract and utilize meaningful information in real time from these measurements. Knowing which data are most meaningful, and how to use it effectively, is itself a challenging task. Improving analyses and forecasts in regions of active and rapidly evolving weather, such as extreme weather events like tropical cyclones or atmospheric rivers, could mean significant reduction in the damages caused by such severe weather events. Utilizing more satellite data in these types of regions when performing data assimilation can result in improved analyses, thus forecasts. Once trained, machine learning (ML) algorithms can identify these types of regions in real time. We took a previously trained ML U-Net for tropical cyclone detection and tested it in satellite data thinning experiments. We will demonstrate the workflow for using ML to create cyclone region observation filters for selective satellite data thinning, where a greater density of satellite observations is used in the ML-identified cyclone regions, and show results for case studies that demonstrate the sensitivity of the resulting analysis to this selective thinning. We will also show forecasts demonstrating improvements in the cyclone forecast results for several case studies. The data assimilation experiments are performed using the Joint Effort for Data Assimilation Integration (JEDI) system, a new object oriented data assimilation software system.