Topological Data Analysis and Machine Learning methods for pattern detection in spatiotemporal climate data
Abstract
Understanding and predicting how extreme weather events change under different global warming scenarios, in particular, how the intensity, frequency, and location statistics of such events may vary due to global warming is one of the most pressing problems in climate science. Furthermore, advances in high-performance computing have enabled producing climate model outputs at ever more prodigious rates---the big data problem. A fundamental requirement for analysis of such a problem is fast, efficient, and accurate methods that can automatically detect, classify, and characterize patterns in climate simulation products. While Deep Learning (DL) is already achieving successes in pattern recognition problems in climate science, a great challenge that remains is to combine feature representation approaches with ML/DL to develop interpretable, fast, and accurate learning methods for detection and characterization of weather patterns usually associated with extreme events. Following a successful application of Topological Data Analysis (TDA) and Machine Learning to classify Atmospheric Rivers in climate data, we propose a new approach based on recent advances in TDA, combined with DL, to detecting Atmospheric Blocking (AB) events, which are often associated with Rossby Wave Breaking, in large climate datasets. Resilient AB events can remain stationary for several days or even weeks and are responsible for many of the severe heat waves and cold snaps in mid-latitudes, including Europe and the United States. In this approach we combine tools from TDA along with DL for detecting and characterizing ABs. Using Persistent Homology from TDA applied to raw images of climate model output we compute a new multiscale representation of geometric features of the associated topological properties in multivariate spatio-temporal climate data. The generated numerical features together with provided labeled data are then used as an input for Deep Convolutional Neural Network models to detect AB patterns (via supervised learning). This approach may be generalizable to a broader set of weather patterns and extreme weather events, with potential to provide meaningful insights across different spatial and temporal resolutions of climate model output.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFMGC43D1360M
- Keywords:
-
- 0555 Neural networks;
- fuzzy logic;
- machine learning;
- COMPUTATIONAL GEOPHYSICS;
- 1626 Global climate models;
- GLOBAL CHANGE;
- 1942 Machine learning;
- INFORMATICS;
- 4313 Extreme events;
- NATURAL HAZARDS