Science motivation

In the era of ‘Big Data’, mining large observational products (satellite measurements, ground-based readings) and massive climate simulations output is key for gaining scientific insights. An important scientific goal is the characterization of extreme weather in current day, and future climate change scenarios. In this work, we consider the problem of finding extreme patterns (such as Tropical Cyclones, Extra-Tropical Cyclones, Atmospheric Rivers) in large climate archives. We present the successful application of Deep Learning, a state-of-the-art machine learning methodology, for finding spatio-temporal patterns. The results from the application of this method can be used for characterizing statistical changes in extreme weather events (both their intensity and frequency) under climate change scenarios.

Methods

We formulate the problem of finding patterns as a classic image classification task. We prepare labeled data (ground truth is obtained from the application of the TECA tool, a catalog of known events from the literature and hand-labeling). We utilize 8 input variables for Tropical Cyclones and 2 variables for Atmospheric Rivers. We construct a Deep Convolutional Neural Network based on the deep learning library-NEON-developed at Nervana System, in conjunction with the Spearmint package for hyperparameter optimization. Our optimal network consists of 4 layers (2 convolutional layer and 2 fully connected layers).

Results

We obtain good classification performance for extreme weather patterns: 99% accuracy for Tropical Cyclones, 90.5% (US Atmospheric Rivers) and 89.5% (European Atmospheric Rivers). The attached figure shows sample weather patterns correctly classified by the Deep Learning architecture.