Deep Learning for Extreme Weather Detection

• Activity Number: 106 - AI and Deep Models for Spatial and Spatio-Temporal Data
• Type: Invited
• Date/Time: Monday, August 3, 2020 : 1:00 PM to 2:50 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Mr Prabhat* and Karthik Kashinath and Michael Wehner and William Collins
• Companies: Lawrence Berkeley National Laboratory and Lawrence Berkeley National Lab and Lawrence Berkeley National Lab and Lawrence Berkeley National Lab
• Keywords: Deep Learning; Weather Prediction; Convolution Neural Networks

Abstract

In recent years, Deep Learning (DL) had enabled numerous breakthroughs in the fields of computer vision, speech recognition and control systems. Our group has been working on exploring Deep Learning for the problem of detecting extreme weather (tropical cyclones, atmospheric rivers, weather fronts, extra-tropical cyclones, etc) in climate datasets. Our experiments over the past few years have led us to believe strongly that DL can succeed in detecting extreme weather patterns. We have demonstrated DL architectures for classification, detection and segmentation of tropical cyclones, atmospheric rivers and weather fronts in complex, multi-variate climate datasets.

The key limiting step in application of DL for pattern detection is access to large training datasets. With an eye towards this important requirement, we have launched the ClimateNet project, which aims at acquiring expert-specified, labeled climate datasets. The goal being to train a unified DL network that can be applied to simulation, reanalysis, or observational datasets. We will describe our efforts in acquiring this dataset, and making it available to the rest of the research community.