Deep learning for downscaling tropical cyclone rainfall

Flood hazards from Tropical cyclones (TCs) are frequently the leading cause of mortality and damages. Current research indicates that TC rainfall will increase by 7 % per degree of ocean warming with a greater proportion of them being extreme . It is vital to understand how this increase in rainfall will translate to flood risk from tropical cyclones by first accurately modelling TC rainfall under present climatic conditions.

General circulation models struggle to reproduce TC rainfall fields so downscaling models are often used to generate more realistic TC rainfall data. Increasingly, rainfall downscaling studies have adopted deep learning techniques from the Computer Vision field to achieve comparable results to traditional methods at a fraction of the computational cost. Initially, convolutional neural networks (CNNs), specifically U-NETs, showed promise in precipitation downscaling. But more recently, the use of Generative models has been explored following the success of GANs in classical image super-resolution problems compared to CNNs. Generative approaches have shown potential at reproducing the fine spatial detail and stochastic nature of precipitation.

Here, we develop upon the WGAN and Variational Autoencoder GAN (VAEGAN) from Harris et al. (2022) and apply it to rainfall data from TCs to increase the resolution of rainfall measurements from 100 km resolution to 10 km resolution.

Overall, the Wasserstein GAN, performed better than other methods, the variational autoencoder GAN, U-Net and bilinear interpolation, across all diagnostics explored. We showed that for regular TCs the WGAN had the most realistic power spectra for all wave numbers, closely followed by the VAEGAN which only deviated for scales of around 5 pixels or fewer. The U-Net and Bilinear Interpolation methods both reproduced power spectra poorly compared to observations, with significant differences present from wave numbers greater than 3. We found that the WGAN had the lowest mean bias overall with errors around the core of the TC within 5 % error, while the VAEGAN had a dry bias of over 5 % outside of the inner core region. Both models had a low negative bias in standard deviation of between 0-5 %.

When looking at the 100 most extreme samples, beyond the intensity of storms used in training the models, the WGAN is able to produce results of similar quality to those for TCs of intensities used in training, except for predictions having too low spread. This indicates that if the WGAN were trained on the full observational dataset, it could perform well for storms more intense than those previously observed, which is important for judging the model's robustness. Conversely, the power spectra of the VAEGAN became more unrealistic and predictions more artificial. There were some very large errors present in VAEGAN at the upper end of the extreme test set which demonstrates the importance of evaluating models on the most extreme, unseen, cases.  Overall, these results show that generative approaches have the potential to generate TC rainfall fields with a high degree of accuracy.