Spatio-temporal downscaling of precipitation data using a conditional generative adversarial network

Natural disasters caused by cyclones, hail, landslides or floods are directly related to precipitation. Global climate models are an important tool to adapt to these hazards in a future climate. However, they operate on spatial and temporal discretizations that limit the ability to adequately reflect these fast evolving, highly localized phenomena which has led to the development of various downscaling approaches .

Conditional generative adversarial networks (cGAN) have recently been applied as a promising downscaling technique to improve the spatial resolution of climate data. The ability of GANs to generate ensembles of solutions from random perturbations can be used to account for the stochasticity of climate data and quantify uncertainties. 

We present a cGAN for not only downscaling the spatial, but simultaneously also the temporal dimension of precipitation data as a so-called video super resolution approach. 3D convolutional layers are exploited for extracting and generating temporally consistent  rain events with realistic fine-scale structure. We downscale coarsened gauge adjusted and climatology corrected precipitation data from Germany from a spatial resolution of 32 km to 2 km and a temporal resolution of 1 hr to 10 min, by applying a novel training routine using partly normalized and logarithmized data, allowing for improved extreme value statistics of the generated fields.

Exploiting the fully convolutional nature of our model we can generate downscaled maps for the whole of Germany in a single downscaling step at low latency. The evaluation of these maps using a spatial and temporal power spectrum analysis shows that the generated temporal and spatial structures are in high agreement with the reference. Visually, the generated temporally evolving and advecting rain events are hardly classifiable as artificial generated. The model also shows high skill regarding pixel-wise error and localization of high precipitation intensities, considering the FSS, CRPS, KS and RMSE. Due to the underdetermined downscaling problem a probabilistic cGAN approach yields additional information to deterministic models which we use for comparison. The method is also capable of preserving the climatology, e.g., expressed as the annual precipitation sum. Investigations of temporal aggregations of the downscaled fields revealed an interesting effect. We observe that structures generated in networks with convolutional layers are not placed completely at random, but can generate recurrent structures, which can also be discovered within other prominent DL downscaling models. Although they can be mitigated by adequate model selection, their occurrence remains an open research question.

We conclude that our proposed approach can extend the application of cGANs for downscaling to the time dimension and therefore is a promising candidate to supplement conventional downscaling methods due to the high performance and computational efficiency.