A Deep Learning–Based Multi-Variable Spatiotemporal Downscaling Approach for High-Resolution Tropical Cyclone Detection

Low spatiotemporal resolution of global climate models (GCMs) outputs such as  CMIP6 models limits accurate detection of tropical cyclone (TC). Traditional statistical downscaling has difficulties in resolving non-linear relationships among different variables, while dynamical downscaling with regional high-resolution models is computational expensive and often distorts the results due to different dynamics framework with the GCMs. Here, we proposed a deep-learning based Multi-Variable Spatiotemporal Downscaling Generative Adversarial Network (MV-STD-GAN). It simultaneously spatiotemporally downscales five essential variables (sea level pressure, 300hPa/500hPa geopotential height, 10m zonal/meridional wind) closely related to TC detection. Trained on high- and low-resolution ERA datasets, it substantially improves the detection of observed TC and significantly outperforms traditional and other deep learning baselines, when subject to the same detection algorithm. It can also be successfully applied to low-resolution CMIP6 models, detecting TC activities very similar to the corresponding high-resolution models.