Time-Aware UNet and Super-Resolution Deep Residual Networks for Spatial Downscaling

Abstract

Satellite observations of atmospheric pollutants are often available only at coarse spatial resolution, which limits their use in local-scale environmental analysis. Spatial downscaling methods aim to transform such data into high-resolution fields. In this work, two widely used deep learning architectures—the super-resolution deep residual network (SRDRN) and the encoder–decoder-based UNet—for spatial downscaling, are extended with a lightweight temporal module that encodes observation time using either sinusoidal or radial basis function representations and integrates temporal features with spatial information. The proposed time-aware extensions are evaluated in a case study on ozone downscaling over Italy. Results show that, with only a minor increase in computational cost, incorporating temporal information significantly improves downscaling accuracy and accelerates model convergence.