Deep Learning for Heatwaves

Abstract

Due to anthropogenic activities, the global climate has changed significantly during the last decades, leading to a rise in both frequency and intensity of weather extremes around the world. Weather extremes, such as heatwaves and droughts, have devastating effects on human health, ecosystems, infrastructure, food security, and more. Some modes of climate variability, are known to influence multiple regions around the world simultaneously, impacting e.g. global food production through simultaneous extremes. Furthermore, due to the complexity of the climate system, some regions around the world have seen a stronger increase in extremes than elsewhere. Europe is one such hotspot region, with observed temperature trends exceeding those of state-of-the-art global climate models (CMIP6). This mismatch is problematic, because it limits the confidence of climate model projections and experiments. Changes in atmospheric circulation have been identified as the prime reason behind this discrepancy. This thesis firstly investigates possible drivers of the observed atmospheric circulation changes in key hotspots in the Northern Hemispheric midlatitudes. I show that the observed long-term circulation trends originate from two distinct regional atmospheric wave patterns over the US-Atlantic sector and over the Eurasian sector. Long-term trends of each regional wave pattern explain 15% and 26% of the observed regional surface temperature trends, in the two sectors respectively. Furthermore, the wave patterns are preceded by distinct sea surface temperature (SSTs) and outgoing longwave radiation anomalous imprints. Concerningly, CMIP6 simulations do not capture the magnitude of the observed circulation trends, nor the observed patterns and drivers. Secondly, this thesis investigates which types of atmospheric circulation drive extreme heat specifically over western Europe. The key innovation in the approach is the development of a Deep Learning framework to objectively detect spatiotemporal dynamically different heatwave types, trained on large-ensemble climate model data. Four types of dynamically distinct heatwaves are identified: the Atlantic Low, the Atlantic High, the UK High, and the Scandinavian High. I apply the Deep Learning framework to observed heatwaves in the climate reanalysis dataset ERA5, to study changes in the last decades. I use novel interpretability techniques to identify the key physical features of the heatwaves in the latent space of the deep learning model. Moreover, the changes in the latent space indicate that each dynamically distinct type of heatwave is undergoing different changes in its circulation, amplifying certain blocking features. Additionally, Atlantic Plume and Atlantic Low types of heatwaves have increased in frequency over time. Thirdly, this thesis investigates how changes in the climate system can lead to globally synchronous crop failure events. To do so, I identify high impact crop failure using Archetype Analysis; a data driven approach which captures the most extreme and contrasting states in global crop production. I show that both individual synchronous global crop failures, e.g. Soy bean and Maize, as well as multi-crop failures are linked to compounding hot-dry conditions on land and specific SST fingerprints. This fingerprint consists of La Niña like SST anomalies in the tropical Pacific and negative Pacific Decadal Oscillation conditions. This is particularly concerning as I find an increase in these SST fingerprints since 1979, while previous research indicates a mismatch in SST trends between observations and model projections under anthropogenic forcing. This highlights the deep uncertainty originating from the mismatch between climate models and observations to do with global teleconnections, with potentially severe implications for many affected regions around the world. Finally, this thesis synthesizes the results within a wider scientific context, discussing the implications of the mismatch between observations and climate model simulations, and how Artificial Intelligence methods can be used to improve our understanding of extreme events in a changing climate.