Detecting and understanding extreme temperature events and heatwaves using machine-learning

Abstract

The intensification of extreme temperature events, particularly heatwaves, is among the most severe impacts of human-induced climate change. These events have profound implications for public health, agriculture, infrastructure, and natural systems. Accurate detection and understanding of such extremes are essential for adaptation and mitigation strategies. However, traditional methods for detecting extreme temperature events often rely on assumptions of unimodal distributions, which may inadequately capture the complex and evolving nature of extreme heat events in a warming world. This dissertation addresses these limitations by developing and applying unsupervised machine learning approaches to detect and understand extreme temperature events, with a focus on both statistical frequency and associated atmospheric dynamics.

In the first part of the dissertation, I present a novel approach for detecting extreme temperature events using Gaussian Mixture Model (GMM), applied to global, daily near-surface maximum temperature data from both European Centre for Medium-Range Weather Forecast (ECMWF) reanalysis v5 (ERA5) and Coupled Model Intercomparison Project Phase 6 (CMIP6) simulations. By modelling temperature distributions as multimodal rather than unimodal, GMM provides a better fit for daily maximum temperature data, particularly in mid-latitude regions where cold and warm seasons create distinct modes in the temperature distribution. This method captures the differences in geographical regions and shows that the frequency of extreme events will be even higher than reported in previous studies. Globally, a 10-year extreme temperature event relative to 1985–2014 conditions will occur 13.6 times more often in the future under a Global Warming Levels (GWL) of 3.0C. The frequency increase can be even higher in tropical regions, such that 10-year extreme temperature events will occur almost twice a week. Additionally, the hot temperatures are increasing faster than cold temperatures in low latitudes, while the cold temperatures are increasing faster than the hot temperatures in high latitudes under different GWL. The smallest changes in temperature distribution can be found in tropical regions, where the annual temperature range is small.

The second part of the dissertation shifts from univariate statistical modelling to the spatiotemporal analysis of multivariate heatwave dynamics using deep learning. A spatiotemporal Variational Autoencoder (VAE) was trained on year-round eleven-day heatwave samples from the ERA5 reanalysis dataset from 1941-1990 over the North Atlantic region, incorporating nine atmospheric variables including temperature, wind, humidity, cloud cover, radiation, and geopotential height. The VAE encodes each multivariate event into a compact latent space, which is then clustered using GMM to identify characteristic heatwave regimes. Then, the VAE was tested with heatwave samples from 2001-2022 to analyse atmospheric patterns before and during the Western European heatwaves. Notably, recent summer heatwaves form a distinct and previously unseen cluster, suggesting a shift in atmospheric circulation patterns in response to climate change. Composite anomaly maps reveal coherent pre-onset signatures across variables, indicating that the VAE can extract physically interpretable features from complex, high-dimensional data. Together, the two studies presented in this dissertation demonstrate the value of unsupervised learning in climate science. While GMM provide a flexible and interpretable statistical framework for quantifying changes in the frequency and distribution of extreme temperature events, the VAE offers a novel approach for understanding the underlying physical drivers of heatwaves. By applying these methods to observational reanalysis data and climate model outputs, this dissertation contributes to advancing our understanding of how extreme temperature events are evolving in a warming world. The findings not only reveal the increasing risk of high-impact extreme temperature events but also highlight the importance of combining statistical and physical perspectives to better characterize, monitor, and ultimately predict these events.