Can generative AI models downscale very rare precipitation events? An illustration of the 2020 south of France flash flood.

1 Introduction and objectives

Downscaling of climate model outputs is essential to correctly represent the distribution of precipitation, including extreme events which are often missed by coarse-resolution models. Deep learning approaches have shown some skills to downscale precipitation at a much lower cost than physically-based models. However, it is unclear how well they capture extreme events with decadal return periods, which only occur a few times in the training dataset. We propose to study how a diffusion-based model downscales a particularly intense precipitation episode which occurred in the South-East of France in October 2020, an event whose return period is estimated to be 100 years by Météo France [1].

2 Methods and data

Our diffusion model is based on the EDM framework from Karras et al [2]. It was trained on a domain of size 350 x 350 km² centered around Nice, France and uses precipitation and multiple covariates from the ERA5 dataset as predictors and precipitation from the regional CERRA reanalysis as predictand. An evaluation of the downscaling method using conventional ML metrics shows that our model is capable of capturing typical precipitation. To further evaluate the ability of the model to downscale very rare events, we downscaled the situation between October 2nd 06:00 UTC to October 3rd 00:00 UTC ten times using the generative capability of our diffusion-based model. The maximum six-hour accumulated precipitation registered in the CERRA dataset during this episode is above 120 mm, much more than the 99.99th percentile of the training period, which is 36 mm. We then evaluated the downscaled precipitation using CRPS, Fraction Skill Score, RAPSD, and PIT histograms [3,4], and checked the distribution of downscaled precipitation on a pixel basis as well as aggregated on hydrological basins. The diffusion model’s performance is compared to two deterministic baselines: a UNet model and a classical statistical downscaling method (bias correction spatial disaggregation) [5].

3 Results

The diffusion model is able to generate visually realistic samples for moderate precipitation events that better capture the spatial structure and distribution of precipitation than the considered baselines. When downscaling the considered extreme event, the diffusion model reproduces the position of intense precipitation, but underestimates or overestimates the intensity depending on samples (see figure). Because the event considered is exceptional in the studied dataset, it is difficult to know if this variability should be considered as an error of the downscaling model or if it is inherent to the distribution of fine scale precipitation conditioned on coarse-scale atmospheric covariates. Because samples are independent of their temporal predecessors, the 12 hour- and 24 hour-accumulated precipitation fields generated lack fine-scale details. We will investigate further this event and test ways to improve its simulation.

Figure 1. Example of four samples of downscaled six-hour accumulated precipitation fields (diffusion 0 to 5), compared to the corresponding ERA5 (low resolution) (top left) and CERRA (high resolution) (bottom left) precipitation for October 2nd, 12:00 to 18:00 UTC.