Super-Resolution of Climate Projections using NASA Satellite Images at 1-km Resolution

Proceedings of The 3rd World Conference on Climate Change and Global Warming

Year: 2023



Super-Resolution of Climate Projections using NASA Satellite Images at 1-km Resolution

Grzegorz Chlodzinski, Alka Dagar, Shruti Iyyer, Yuri Katz, Saurabh Paul, Gaurav Singh




Climate projections coming from General Circulation Models (GCMs) outline the impact of climate change decades into the future, However, GCM projections are typically at a scale that is too coarse for local risk assessment. To address this limitation, NASA Earth Exchange (NEX) provides climate projections at 25km resolution using statistical downscaling (SD) techniques. We propose deep-learning based image super-resolution (SR) techniques that further enhance the resolution of the NEX product to a scale of 1km. Computer-vision SR techniques, by design, are better suited to generalize spatially and are not limited to availability of ground-truth observational data at high spatial-resolution, a key requirement for SD. We focus on surface air temperature and demonstrate how SR models trained on NASA satellite data, available globally at 1km resolution, outperforms interpolation-based techniques while capturing local variability of temperature. Additional data channels, such as elevation, are seamlessly incorporated into the deep learning architecture, improving fidelity of the output. The trained model is then applied to enhance the resolution of local temperature profiles projected up to 2099. The proposed approach is generic and can be applied to other hazards, such as precipitation, thus opening the possibility of local risk assessment using high-resolution fundamental climatological variables.

keywords: climate change, climate hazards, deep learning, general circulation models, image super-resolution