RainScaleGAN: a Conditional Generative Adversarial Network for Rainfall Downscaling

TL;DR

RainScaleGAN introduces a conditional Wasserstein GAN with gradient penalty to downscale precipitation from coarse $2.0^{\circ}\times2.0^{\circ}$ grids to fine $0.25^{\circ}\times0.25^{\circ}$ grids in a perfect-model ERA5 framework, achieving improved spatial detail and statistical realism over RainFARM. The method uses a noise-augmented generator conditioned on low-resolution fields and a Lipschitz-discriminating critic, optimized without a content loss, and evaluated with RMSE, LSD, climatology, percentile statistics, and spectral diagnostics. Results show closer alignment with ground-truth distributions, better reproduction of extreme values, and plausible ensemble variability, highlighting RainScaleGAN’s potential for efficient probabilistic downscaling and its applicability to other atmospheric variables. This work points to a scalable, region-agnostic alternative to dynamical downscaling, enabling rapid generation of high-resolution rainfall fields for climate projections and extreme-event studies, while prompting further bias-correction, regional validation, and extension to additional variables.

Abstract

To this day, accurately simulating local-scale precipitation and reliably reproducing its distribution remains a challenging task. The limited horizontal resolution of Global Climate Models is among the primary factors undermining their skill in this context. The physical mechanisms driving the onset and development of precipitation, especially in extreme events, operate at spatio-temporal scales smaller than those numerically resolved, thus struggling to be captured accurately. In order to circumvent this limitation, several downscaling approaches have been developed over the last decades to address the discrepancy between the spatial resolution of models output and the resolution required by local-scale applications. In this paper, we introduce RainScaleGAN, a conditional deep convolutional Generative Adversarial Network (GAN) for precipitation downscaling. GANs have been effectively used in image super-resolution, an approach highly relevant for downscaling tasks. RainScaleGAN's capabilities are tested in a perfect-model setup, where the spatial resolution of a precipitation dataset is artificially degraded from 0.25$^{\circ}\times$0.25$^{\circ}$ to 2$^{\circ}\times$2$^\circ$, and RainScaleGAN is used to restore it. The developed model outperforms one of the leading precipitation downscaling method found in the literature. RainScaleGAN not only generates a synthetic dataset featuring plausible high-resolution spatial patterns and intensities, but also produces a precipitation distribution with statistics closely mirroring those of the ground-truth dataset. Given that RainScaleGAN's approach is agnostic with respect to the underlying physics, the method has the potential to be applied to other physical variables such as surface winds or temperature.

Figures (16)

• Figure 1: ERA5 daily total accumulated precipitation (left) and corresponding coarsened version (right) for two sample days (04 November 1966 and 07 October 1970). The original ERA5 examples have a spatial resolution of 0.25°$\times$0.25° and consist of 64x64 grid points. The coarsened versions, obtained with an upscaling factor of 8, have a spatial resolution of 2.0°$\times$2.0° and consist of 8x8 grid points.
• Figure 2: Information flow during the model training.
• Figure 3: The architecture of the networks composing RainScaleGAN. The input layers of the two networks (coarse image + noise source for the generator, generated/ground-truth image + corresponding coarse image for the discriminator) are not shown. (Top) Generator architecture. The upsampling layers have upsampling factors of (2,2), thereby doubling the number of rows and columns of their input. The intermediate convolutional layers have a number of kernels equal to 256, 128, 64, and 32, respectively. The final convolutional layer has a single kernel and is activated with the hyperbolic tangent function. (Bottom) Discriminator architecture. The convolutional layers, except the last one, have a number of kernels equal to 64, 128, 256, 512, respectively, and strides (2,2). Each of them halves the height and width of the field it receives as input.
• Figure 4: Evolution of image metrics throughout the GAN training process. (Top) Log-spectral distance between the spatial radial spectra of the generated precipitation and that of the corresponding true dataset. (Bottom) Root-mean-squared errors between the generated images and their corresponding true counterparts. The solid lines refer to the training dataset (1940-1998), while the dashed lines correspond to the validation dataset (1999-2010).
• Figure 5: Evolution of statistical metrics over the GAN training: root-mean-squared errors between (top) the climatology and standard deviation, and (bottom) 95th and 99th percentiles of the generated dataset with respect to the corresponding true dataset. The solid lines refer to the training dataset (1940-1998), while the dashed lines correspond to the validation dataset (1999-2010).