AI-based downscaling in the Mediterranean for improved understanding of observed and projected climate change

The Mediterranean region has been identified as a hotspot for climate change, with increasing frequency and intensity of extreme events such as heatwaves, heavy precipitation, and drought episodes. In this line, regional-to-local climate information is key to better comprehend the observed and projected changes in a global warming scenario.

 

This study develops an AI-based downscaling architecture to obtain high-resolution information on essential climate variables. The model is based on an initial regression using a (deterministic) U-Net, and then we assess the potential of improving the residuals through a generative approach based on diffusion. In a first stage, we calibrate the deep-learning model using large-scale and thermodynamic predictors from the ERA5 reanalysis (1950-2024) and surface target variables (e.g., tasmax, precipitation) from different observational and reanalysis-based products, including ERA5-Land. Different evaluation metrics are quantified, including systematic biases, anomaly correlations, and extreme indices at seasonal and annual scales (e.g., TXx, Rx1day). 

 

Once the model is tested for efficiency, the potential for several applications arises, including explainability metrics to be used in climate attribution studies (e.g., saliency maps) and replication into general circulation models from the CMIP6 experiment in different scenarios. Preliminary results show the AI-model is able to represent the climatology of the target variables, highlightinghting the added value of deep-learning downscaling in capturing several aspects of the Mediterranean climate, with focus on extremes, while presenting sensitivity to observational reference and model parameters. 

 

This methodology provides tailored climate information useful when providing climate services, complementing other data sources available at the regional scale. In this way, multi-model and multi-method approaches are recommended to address the uncertainty in future projections in the Mediterranean region, especially for extreme events.