On the use of convolutional neural networks for downscaling daily temperatures over southern South America in a climate change scenario

Global Climate Models (GCMs) depict a notable influence of climate change on southern South America (SSA). Future regional-to-local information for adaptation and mitigation policies can be obtained by downscaling over GCMs outputs, increasing the resolution of the climate projections. Current statistical downscaling approaches in the region [e.g., Generalised Linear Models (GLMs)] need to undergo "human-guided" feature selection, which is one of the main sources of uncertainty. Here, we explore the advantages and limitations of using Convolutional Neural Networks (CNNs) in SSA to downscale daily minimum and maximum temperatures. For this purpose, we elaborate three different experiments: a cross-validation (CV) in the present climate; downscaling the historical and RCP8.5 scenarios of the EC-Earth; a pseudo-reality experiment to measure the extrapolation skill. CV-experiment results show no remarkable differences between CNNs and GLMs, although the non-linearity of the CNNs improved the representation of the extreme aspects of temperatures. Additionally, we use eXplainable Artificial Intelligence to prove that co-linearities are better handled in CNNs. The pseudo-reality experiment shows a good extrapolation skill of CNNs, especially when the activation functions are linear. Overall, the automatic skill of CNNs to deal with co-linearities in predictor data-against conventional approaches-together with the plausible climate change projections obtained-verified with the pseudo-reality experiment-make them attractive to be used for downscaling beyond their non-linear nature. These results enforce the idea of incorporating CNNs into the battery of downscaling techniques over SSA and provide experimental guidelines with prospects to be utilised in climate change studies.