Downscaling Precipitation Extremes Using Physics-coupled Dynamic Data Driven Adversarial Learning

The intensity and frequency of extreme rainfall events are likely to increase under projected climate change scenarios. Given the adverse socio-economic impacts of these extreme events, we need to model their risk to develop effective policies for adaptation and mitigation. Simulating local hydrometeorological processes at the resolutions essential for assessing impacts and planning is computationally expensive using global climate models. Thus, there is a demand for efficacious downscaling from the coarse-resolution climate model outputs to the finer local scales of interest. Here, we develop a dynamic data-driven model coupled with physics, to downscale coarse-resolution climate model outputs (0.25° × 0.25°) to high-resolution (0.01° × 0.01°) rainfall. The downscaled rainfall is initially estimated by actively searching data on a manifold to learn the downscaling function incrementally using an iterative Gaussian process (GP). Upon convergence, the “first-guess” downscaled rainfall field, along with a physics-based estimation of orographic rainfall are processed by an adversarial learning framework (GAN) to refine finer-scale details. A stochastic sampling model and optimal estimation are used to correct the biases and obtain the final rainfall super-resolution fields. We assess the skill of the proposed model, using ERA5 reanalysis data and Daymet observation data at different terrain conditions (plain and hilly), and show that the downscaled rainfall closely matches the ground truth spatial patterns and extreme rainfall risk. By comparing the performance of individual components of our model (GAN, GP, and Physics) we find that the combined model outperform the individual components, and the GAN accounts for the maximum performance gain of the downscaling model.