Deep learning–based downscaling of ERA5-Land temperature to 250 m resolution over the Trentino–South Tyrol Alpine region

High-resolution near-surface temperature data are essential in mountainous regions, where complex topography induces strong spatial and temporal variability. However, coarse-resolution reanalysis products such as ERA5-Land (9 km) fail to represent fine-scale thermal patterns. This study presents a UNet based deep learning framework to downscale ERA5-Land 2-m air temperature to 250 m resolution at a 6-hourly temporal frequency over the Trentino–South Tyrol Alpine region. The model is trained and validated using the observation-based ALPINE-TST-250 gridded dataset, derived from more than 300 weather stations, and integrates high-resolution elevation data along with physically meaningful auxiliary predictors. Including 2-m dew point temperature significantly improves performance, reducing the RMSE from 2.32 ° C to 2.05 ° C and the MAE from 1.76 ° C to 1.54 ° C . The resulting 2011–2021 downscaled dataset successfully reproduces sub-kilometer spatial gradients and temporal variability absent in the original ERA5-Land fields. These results highlight the potential of deep learning approaches to enhance temperature representation in complex alpine terrain for climate and environmental applications.