Detecting Spatiotemporal Dynamics of Western European Heatwaves Using Deep Learning

Heatwaves over western Europe are increasing faster than elsewhere, which recent studies have attributed at least partly to changes in atmospheric dynamics. To understand the dynamical drivers of heatwaves, we develop a heatwave classification based on their spatiotemporal evolution over western Europe. We use the KNMI large-ensemble time scale (KNMI-LENTIS) large ensemble dataset of 1600 years of present-day climate and find approximately 14 000 heatwaves. To detect heatwaves, we use the Generalized Density-based Spatial Clustering of Applications with Noise (GDBSCAN) algorithm to cluster together grid points based on their surface temperature extremes. For each heatwave, we extract a 5-day window of daily sea level pressure and streamfunction at 250 hPa, capturing their near-surface and upper-level atmospheric circulation. We use a 3D variational autoencoder (VAE) to extract compact representations from this high-dimensional circulation dataset. We reduce the dimensionality to a 128-dimensional latent space, which we analyze with several algorithms (t-distributed stochastic neighbor embedding, k-means, and the Gaussian mixture model). This yields physically distinct types of heatwaves that are interpretable with known circulation patterns, i.e., U.K. high, Scandinavian high, Atlantic high, and Atlantic low. Our results indicate that the VAE can extract meaningful features from high-dimensional climate data. We find that the heatwave phase space, as found with opt-SNE, is continuous with the four circulation regimes at its corners but with soft boundaries between these circulation regimes, indicating that heatwaves are best categorized in a probabilistic way.