Deep learning-based spatial downscaling and its application for tropical cyclone detection in the western North Pacific

Abstract

Resolution of global climate models (GCMs) significantly influences their capacity to simulate extreme weather such as tropical cyclones (TCs). However, improving the GCM resolution is computationally expensive and time-consuming, making it challenging for many research organizations worldwide. Here, we develop a downscaling model, MSG-SE-GAN, based on the Generative Adversarial Networks (GAN) together with Multiscale Gradient (MSG) technique and a Squeeze-and-Excitation (SE) Net, to achieve 10-folded downscaling. GANs consist of a generator and a discriminator network that are trained adversarially, and are often used for generating new data that resembles a given dataset. MSG enables generation and discrimination of multi-scale images within a single model. Inclusion of an attention layer of SE captures better underlying spatial structure while preserving accuracy. The MSG-SE-GAN is stable and fast converging. It outperforms traditional bilinear interpolation and other deep-learning methods such as Super-Resolution Convolutional Neural Networks (SRCNN) and MSG-GAN in downscaling low-resolution meteorological data in assessment metrics and power spectral density. The MSG-SE-GAN has been used to downscale the TC-related variables in the western North Pacific in the low-resolution GCMs of HadGEM3-GC31 and EC-Earth3P, respectively. The downscaled data show highly similar TC activities to the direct outputs of the high-resolution HadGEM3-GC31 and EC-Earth3P, respectively. These results not only suggest the validity of the MSG-SE-GAN but also indicate its possible portability among low-resolution GCMs.