Super-resolution for downscaling climate data

Abstract

A common task in Earth Sciences is to infer climate information at local and regional scales from global climate models. An alternative to running expensive numerical models at high resolution is to use statistical downscaling techniques. Statistical downscaling aims at learning empirical links be- tween the large-scale and local-scale climate, i.e., a mapping from a low-resolution gridded variable to a higher-resolution grid that incorporates observational data. Seasonal climate predictions can forecast the climate vari- ability up to several months ahead and support a wide range of societal activities. The coarse spatial resolution of seasonal forecasts needs to be downscaled or refined to the local scale for specific applications. In this study, we present super-resolution (SR) techniques for the task of downscaling climate variables with a focus on temperature over Catalonia. Our models are trained using high and medium resolution ( ~ 5 and ~ 25 km) gridded climate datasets with the ultimate goal of increasing the resolution of coarse resolution ( ~100 km) seasonal forecasting systems. Taking the gridded data from ~100 to ~5 km implies a 20x upscaling factor. It is worth pointing out that handling such large upsampling factor is not typical in computer vision, where most applications factors while 16x is considered as extreme SR.