Deep Learning for Detecting Extreme Weather and Climate Patterns (Core Science Keynote)

A deeper understanding of climate model simulations can be attained through direct analysis of the phenomena that produce weather. Realistically, this requires procedures that automatically identify such phenomena in an unbiased and comprehensive way. This talk presents a comprehensive assessment of Deep Learning for extracting extreme weather patterns from climate datasets. We demonstrate the application of supervised convolutional architectures for the binary classification task of detecting and tracking tropical cyclones and atmospheric rivers in cropped, centered image patches. Subsequently, we demonstrate the application of similar architectures to detecting the occurrence and categorizing the types of weather fronts at the granularity of a grid-cell. Finally, we develop a unified architecture for simultaneously localizing as well as classifying tropical cyclones, extra-tropical cyclones and atmospheric rivers. We also show the potential benefit of the semi-supervised approach, which provides the possibility of detecting other coherent fluid-flow structures that may not yet have a semantic label attached to them. This talk highlights a number of avenues for future work stemming from pragmatic challenges associated with improving the performance and scaling of Deep Learning methods and hyper-parameter optimization. Extending the methods to 3D space-time grids is a natural next step, although this will require creation of large training datasets produced through community-led labeling campaigns. Finally, improving the interpretability of these methods will be essential to ensure adoption by the broader climate science community.