Transfer Learning for the Canadian Airspace: Leveraging a Globally-Trained UNet Model to Create Enhanced Radar Depictions in Regional Domains

Weather radar plays a vital role in aviation operations but is dependent on land-based networks to provide situational awareness for air traffic controllers and traffic flow managers. To compensate for gaps in radar, prior work at the Massachusetts Institute of Technology Lincoln Laboratory (MIT LL) used deep learning (DL) to provide coverage for large regions of the globe using globally-available datasets, specifically geostationary satellite infrared and visible imagery, lightning detection, and output from numerical weather prediction models. UNet architectures were used to ingest multi-year archives of global datasets and generate weather radar-like images of storm intensity and height, utilizing space-based radar data from the Global Precipitation Mission (GPM) as truth. Currently, MIT LL is working with NAV CANADA to address the need for a DL model to produce synthetic weather radar specialized for Canadian airspace at a higher spatial resolution. Training a specialized DL model for a given region and resolution typically requires extensive datasets, which may not be available in data-sparse regions such as Canada. Here we present a methodology for utilizing transfer-learning to fine-tune and spatially enhance a globally-trained UNet model for the Canadian airspace without the need for extensive archival data. This methodology enables synthetic radar depictions to be generated for domains in which only a limited training set may be available, while leveraging a previously-trained generalized global model. Additionally, the synthetic depictions of vertically integrated liquid (VIL) and echo tops generated for NAV CANADA were compared and bias-corrected to corresponding outputs from the NEXRAD and Canadian radar data.

© 2023 Massachusetts Institute of Technology.