Wasserstein GAN-based precipitation downscaling with optimal transport for enhancing perceptual realism

Abstract High-resolution (HR) precipitation prediction is essential for reducing damage from heavy rainfall; however, HR precipitation forecasts using process-driven numerical weather prediction models remain challenging. For precipitation downscaling, this study proposes using conditional Wasserstein Generative Adversarial Network (WGAN) which matches the distribution of generated images to real ones, enabling it to capture higher-order statistics and sharper structures. In this study, we conduct experiments under a perfect approach in which models are trained using sets of HR precipitation and corresponding low-resolution precipitation obtained by downsampling the HR data. In contrast to a conventional neural network trained with mean squared error, the WGAN generated visually realistic precipitation fields with fine-scale structures even though the WGAN exhibited slightly lower performance on conventional evaluation metrics. The learned critic of WGAN correlated well with human perceptual realism. Case-based analysis revealed that large discrepancies in critic scores can help identify both unrealistic WGAN outputs and potential artifacts in the reference data. These findings suggest that the WGAN framework not only improves perceptual realism in precipitation downscaling but also offers a new perspective for evaluating and quality-controlling precipitation datasets.