Detecting Extreme Temperature Events Using Gaussian Mixture Models

Extreme events are rare atmospheric phenomena that cause significant damage to humans and natural systems, but detecting extreme events in the future in a changing climate can be challenging. Traditionally, temperature distributions were assumed to follow a normal distribution and certain thresholds were used to define extreme events. However, the mean and the variance of temperatures are expected to change in a future climate, which might limit the application of traditional methods for detecting extreme events.

We found that daily maximum surface temperature data can be described accurately using a multimodal distribution. In this study, we therefore used a statistical method called Gaussian Mixture Models (GMM) to fit a multimodal distribution to daily near-surface maximum air temperature data from simulations participating in the Coupled Model Intercomparison Project Phase 6 (CMIP6) for 46 Intergovernmental Panel on Climate Change (IPCC) land regions. GMM allowed us to use the parameters from the Gaussian distribution fitted to the higher temperatures to define the thresholds for the return period of extreme events. We analysed the change in the return periods of extreme temperature events in study regions compared to the historical period (1980-2010) under future Global Warming Levels (GWL) of 1.5°C, 2°C, 3°C and 4°C for each Shared Socioeconomic Pathways (SSP) scenarios.