AI-Based Super-Resolution Approaches to Reconstruct Temperature Fields of Climatic Data Using Deep Neural Networks

High-fidelity climate analysis is heavily constrained both by the computational costs of high-resolution dynamical simulations and by the general scarcity of fine-scale observational data for model evaluation. To address these limitations, in this work we propose a deep learning framework for spatial climate downscaling using an AI-based super-resolution (SR) approach. We develop Deep Neural Network architectures designed to upscale 2 m surface temperature fields, bridging the gap between gridded global reanalysis data at lower scale (1.0°) and higher scale (0.25°). We developed three DNN models, specifically ResNet, RNN, and ConvRNN, trained and validated on ECMWF ERA5 data from the early years of the current millennium (i.e., 01/01,01/04,01/07,01/10 of 2001,2003,2005 at 00:00) and evaluated over various periods, including 1992, a historical period influenced by the 1991 Mount Pinatubo eruption (a large volcanic eruption which is known to have had impacts on the climate due to the massive stratospheric injections of sulfur dioxide). The framework is tested also on the Berkeley Earth global observational dataset and on outputs from a MPI-ESM1-2-LR historical global climate simulation and compared with traditional interpolation methods, to verify the models generalization capability. The models reached high levels of accuracy, with errors lower than 0.5 K and a similarity index higher than 0.97, successfully capturing the spatio-temporal dynamics of temperature fields. Overall, the novelty of this work is in three main aspects: the implementation of DNN capable to capture the spatio-temporal dynamics and not only the static frame-by-frame processes, the combination of conventional pixel-based error metrics with morphological validation frameworks based on isoline pattern and gradient orientation comparison, and the introduction of optimized training strategies (i.e., early stopping), to improve learning mechanisms and balance computational costs.