Classification and Localization of Extreme Weather Patterns with Deep Learning

Monday, December 12, 2016 - 16:00
Add to Calendar

Extreme weather events pose great potential risk on ecosystem, infrastructure and human health. Analyzing extreme weather in the observed record (satellite, reanalysis products) and characterizing changes in extremes in simulations of future climate regimes is an important task. Thus far, extreme weather events have been typically specified by the community through hand-coded, multi-variate threshold conditions. Such criteria are usually subjective, and often there is no agreement in the community on the specific algorithm that should be used. We propose a completely different approach: machine learning to solve this problem. If human experts can provide spatio-temporal patches of a climate dataset, and associated labels, we can turn to a machine learning system to learn the underlying feature representation. The ‘trained’ ML system can then be applied to novel datasets, thereby automating the pattern detection step. Summary statistics, such as location, intensity and frequency of such events can be easily computed as a post-process.

This talk will touch upon Deep Learning: the most powerful machine learning method at this point in time. We will report compelling results from the successful application of Deep Learning to classify tropical cyclones, atmospheric rivers and weather front events. For all of these events, we observe 90-99% classification accuracy by the Deep Learning system. We will also report on progress in localizing such events: namely drawing a bounding box (of the correct size and scale) around the weather pattern of interest. Both tasks currently utilize multi-layer convolutional networks in conjunction with hyper-parameter optimization. We utilize HPC systems at NERSC to perform the optimization across multiple nodes, and utilize highly-tuned libraries to utilize multiple cores on a single node. We will conclude with thoughts on the frontier of Deep Learning: can we train networks in a semi-supervised, or completely unsupervised manner?

Link for More Information: