Unsupervised Storm Classification Model Reveals Key Differences in Great Lakes Impacts
Date:
Leveraging machine learning to categorize extratropical cyclones and enhance water level forecasting in the Great Lakes region
Abstract
The Great Lakes basin (GLB) is home to approximately 34 million people in the United States and Canada – about 8% and 32% of their respective populations – who rely on its natural resources for drinking water, commerce, and recreation. In addition, its closed system dynamics and varied physical conditions make the Great Lakes ideal for studying atmosphere-land-hydrosphere interactions. Extratropical cyclones (ETCs), a key component of the region’s weather patterns, impact the GLB through winds, heat transport, precipitation, and evaporation. Understanding these impacts has implications for coastal infrastructure management, planning, and improving predictive hydrologic models.
We used a database of ETCs derived from ERA5, covering storms that impacted the Great Lakes watershed. Using unsupervised classification techniques, we grouped ETCs based on features such as minimum central pressure, maximum storm speed, maximum radius, and two novel metrics: the fraction of time spent in the GLB and a proxy for maturity at GLB entrance relative to the storm’s lifespan.
Using a simple two-class model, we examined the differences in ETC impacts on the GLB. This classification proved robust across methods, including K-means, Gaussian Mixture Models (GMM), and hierarchical clustering. Our model identified two storm types: Type 1 storms, characterized by lower central pressures and higher speeds, predominantly originating from the Atlantic, and Type 2 storms, characterized by higher central pressures and lower speeds, predominantly forming in the west.
We found statistically significant differences in the impacts of these two storm classes on the Great Lakes, particularly in terms of precipitation, with variations across the lakes. After the early 1990s, the stronger Type 1 storms became more frequent annually than the Type 2 storms. We examine possible causes, including changes in the phase of the Pacific Decadal Oscillation (PDO), other shifts in large-scale atmospheric patterns, and background climate change. By categorizing ETCs and quantifying their impacts, our research aims to better understand the connections between large-scale atmospheric patterns, ETCs, and impacts on watersheds, while advancing the use of unsupervised classification for hypothesis generation in the Earth system sciences.
Plain-Language Summary
Extratropical storms, which are powerful weather systems, can impact the Great Lakes region through, for example, heavy rains, strong winds, and flooding. To better understand these storms and how they affect the Great Lakes, we analyzed a large dataset of storms. Using machine learning, we grouped the storms into different categories based on their characteristics, such as pressure, speed, and size.
Our analysis revealed that we can view the system as consisting of two main types of storms: stronger storms with more intense characteristics and weaker storms. We found that stronger storms tend to generally come from more eastern longitudes, while weaker storms generally come from more western longitudes. Interestingly, we noticed an increase in the number of stronger storms since the early 1990s.
By categorizing these storms and understanding their impacts, our research aims to improve predictive water level forecasting and help communities better prepare for these powerful weather events and understand how their impacts might change in the future. We also hope to demonstrate how unsupervised classification, the type of machine learning used here, can help researchers formulate hypotheses about weather and climate that can be further investigated.
Slides
DOI: 10.22541/essoar.173482266.60616090/v1
Additional Information
- Abstract ID and Title: 1578141: Unsupervised Storm Classification Model Reveals Key Differences in Great Lakes Impacts
- Final Paper Number: A44B-08
- Presentation Type: Oral
- Session Number and Title: A44B: AI-Driven Innovations in Earth and Climate Sciences II Oral
- Presentation Length: 17:10 - 17:20 EST
- Session Date and Time: Thursday, 12 December 2024; 16:00 - 17:30 EST
- Location: Convention Center, 154 A-B