Deep Learning for Downscaling Tropical Cyclone Rainfall to Hazard-Relevant Spatial Scales

Flooding, driven in part by intense rainfall, is the leading cause of mortality and damages from the most intense tropical cyclones (TCs). With rainfall from TCs set to increase under anthropogenic climate change, it is critical to accurately estimate extreme rainfall to better support short-term and long-term resilience efforts. While high-resolution climate models capture TC statistics better than low-resolution models, they are computationally expensive. This leads to a trade-off between capturing TC features accurately, and generating large enough simulation data sets to sufficiently sample high-impact, low-probability events. Downscaling can assist by predicting high-resolution features from relatively cheap, low-resolution models. Here, we develop and evaluate a set of three deep learning models for downscaling TC rainfall to hazard-relevant spatial scales. We use rainfall from the Multi-Source Weighted-Ensemble Precipitation observational product at a coarsened resolution of similar to 100 km, and apply our downscaling model to reproduce the original resolution of similar to 10 km. We find that the Wasserstein Generative Adversarial Network is able to capture realistic spatial structures and power spectra and performs the best overall, with mean biases within 5% of observations. We also show that the model can perform well at extrapolating to the most extreme storms, which were not used in training. Plain Language Summary Tropical cyclones (TCs) are often associated with intense winds, but it is actually their associated rainfall and flooding that cause the majority of mortality and damages. A warmer atmosphere is able to hold more water vapor and therefore we expect to see increases in rainfall from TCs with global warming. To better support resilience efforts, it is critical to model current and future TC rainfall, but climate models at standard resolution struggle to do this accurately. Running climate models at very high resolution produces better results, though this requires significant computational resources meaning that fewer high-impact, low-probability TCs can be generated. Other methods, called downscaling models, are used to provide a computationally cheaper alternative by generating high-resolution TC-specific data rather than an entire global climate data set. In this study, we develop a set of deep learning models which can generate high-resolution rainfall data from low-resolution rainfall data. To do this, we train our models on data from observational data sets that have data for the period 1979-2020. We find that the Wasserstein Generative Adversarial Network performs the best over the metrics studied and is able to reproduce the most extreme storms that were not used in training.