Downscaling Methods Combining Super-Resolution and Multi-task Learning

In recent years, deep learning techniques have been introduced into the meteorological field due to their efficient prediction capabilities, especially in downscaling techniques, which accelerate the modelling process of meteorological data by establishing a link between low-resolution meteorological data and high-resolution local climate features. While downscaling methods for single meteorological variables have limitations, this study proposes MtISD, a downscaling framework that combines enhanced deep super-resolution network (EDSR) and multi-task learning (MTL). This framework reveals the intrinsic connection between multiple meteorological variables by training them simultaneously, thus improving the accuracy and generalisability of downscaling results.accuracy and generalisability of the downscaling results. Compared with traditional univariate downscaling methods, the MtlSD framework is better able to handle multiple meteorological variables, such as temperature, precipitation, wind speed, etc., and extends to more complex variables, such as humidity, pressure, and radiation, while maintaining its efficiency. This approach provides an innovative and practical solution to overcome the limitations of existing downscaling algorithms.