Characterizing Temperature Extremes and their relationship with Land-Atmosphere Interactions in the Mediterranean Using Deep Learning: Insights from Past and Future Climate Data

Temperature extremes, among others climate extremes, are becoming a growing concern for both researchers and policymakers due to their significant and devastating impact on socio-economic sectors (such as agriculture), infrastructure, ecosystems and human health. Recent trends indicate an increase in the frequency and intensity of extreme temperature events, particularly in the Mediterranean region, which is highly vulnerable to climate change [1]. According to the Intergovernmental Panel on Climate Change (IPCC) Sixth Assessment Report [2], the Northern Hemisphere is experiencing more frequent and intense weather extremes, leading to greater economic and human losses. Understanding and quantifying temperature extremes, both historically and under future climate scenarios, is crucial for developing strategies to mitigate these impacts and strengthen resilience.

To gain deeper insights into potential changes in the frequency, duration, and intensity of heatwaves in the Mediterranean region, we apply a novel approach developed within the InterTwin project. Rather than relying on traditional statistical methods, this study employs a generic machine learning (ML) framework based on Convolutional Variational Auto Encoders (CVAE), enabling a more efficient and scalable analysis of extreme temperature events. The unsupervised anomaly detection approach identifies anomalies by measuring reconstruction errors (the difference between the original and reconstructed images). To assess the robustness of this method, we compare the results against traditional statistical techniques using extreme temperature indices defined by Expert Team on Sector-Specific Climate Indices (https://climpact-sci.org/indices/). Heatwaves are analyzed based on summer daily maximum temperature from the Regional Climate Model EBU-POM, with a horizontal resolution of 0.5°. EBU-POM, developed at the Faculty of Physics, University of Belgrade, is a part of the Med-CORDEX framework within the CORDEX international initiative. Additionally, heatwaves are examined for both past and future climate conditions up to 2100, following the RCP8.5 climate scenario. Given the strong influence of land-atmosphere interactions on climate extremes in the Mediterranean region, where soil moisture affects air temperature [3], we also assess the correlation between land-atmosphere coupling metrics and heatwave characteristics in past and future climates.

This approach highlights the potential of machine learning in climate research, offering more efficient and scalable methods for analyzing climate extremes in the Mediterranean region, particularly in handling large time series from climate ensembles.

[1] P. Lionello and L. Scarascia, The relation between climate change in the Mediterranean region and global warming, Reg. Environ. Change 18, 1481 (2018).

[2] IPCC, Climate change 2021: The Physical Science Basis. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change, 2(1), p.2391 (2021).

[3] I. Lazić, et al., Impact of soil texture in coupled regional climate model on land-atmosphere interactions, Theor. Appl. Climatol. 156, 165 (2025).

👉Acknowledgment:
This research was supported by the Science Fund of the Republic of Serbia, No. 7389, Project “Extreme weather events in Serbia - analysis, modelling and impacts” - EXTREMES

👉 Poster link