A matrix-variate mixture model for clustering weather data with seasonal trends

Michael Fop, University College Dublin

Co-authors: Ganesh Babu, University College Dublin; Claire Gormley, University College Dublin; Volodymyr Melnykov, University of Alabama

Abstract: Long-term weather changes, driven by human activities such as fossil fuel consumption, deforestation, and industrialization, require comprehensive analysis to understand climate shifts. Traditional classification methods, such as the Köppen-Geiger classification (KGC), group local weather into climate zones relying on threshold-based categorization of weather variables like temperature and precipitation, but fail to capture temporal variations. To address this limitation, a mixture of matrix-variate Gaussian models with sine seasonality (MMGS) is introduced for clustering locations based on weather data. This approach considers weather variables recorded over time, representing each location as a matrix where the dimensions correspond to different variables and to sequential time points. The proposed MMGS models component means using a parsimonious sine function, effectively capturing seasonality and long-term trends, enabling a clear analysis of temporal weather changes within each cluster. Applied to U.S. weather station data, MMGS identifies interpretable regions characterized by different weather trends over time.