Deep Learning Super-Resolution for Temperature Data Downscaling: A Comprehensive Study Using Residual Networks

Shailesh Kumar JhaVivek Gupta^*

• Indian Institute of Technology Mandi, Mandi, India

Extreme weather events like heatwaves, cyclones, floods, wildfires and droughts are becoming more frequent because of climate change. Climate change causes shifts in biodiversity, which at regional scale impacts agriculture, forest ecosystem, and the water resources. However, to study those impacts at the regional scale, the spatial resolution provided by the general circulation models (GCMs) and reanalysis products are not adequate. This study evaluates advanced deep learning models for downscaling ECMWF Reanalysis v5 (ERA5) 2 m temperature data by a factor of 10 (i.e., ~250 km to ~25 km resolution) for the region spanning 50° to 100° E and 0° to 50° N. We concentrate on gradually improving downscaling models with help of residual networks. We compare the baseline Super-Resolution Convolutional Neural Network (SRCNN) model with two advanced models: Very Deep Super-Resolution (VDSR) and Enhanced Deep Super-Resolution (EDSR) to assess the impact of residual networks and architectural improvements. The results indicate that VDSR and EDSR significantly outperform SRCNN. Specifically, VDSR increases the Peak Signal to Noise Ratio (PSNR) by 4.27 dB and EDSR by 5.23 dB. These models also enhance the Structural Similarity Index Measure (SSIM) by 0.1263, and 0.1163, respectively, indicating better image quality. Furthermore, improvements in the 3°C error threshold are observed, with VDSR and EDSR showing increases of 2.10% and 2.16%, respectively. Explainable AI technique called saliency map analysis provided insights into model performance. Complex terrain areas, such as the Himalayas and Tibetan Plateau, benefit the most from these advancements. These findings suggest that advanced deep learning models employing residual networks, like VDSR and EDSR, significantly enhance temperature data accuracy over SRCNN. This approach holds promise for future applications in downscaling other atmospheric variables.