Towards a more realistic and detailed deep-learning-based radar echo extrapolation method

Abstract

Deep-learning-based radar echo extrapolation methods have achieved remarkable progress in the precipitation nowcasting field. However, they suffer from a common notorious problem—they tend to produce blurry predictions. Although some efforts have been made in recent years, the blurring problem is still under-addressed. In this work, we propose three effective strategies to assist deep-learning-based radar echo extrapolation methods to achieve more realistic and detailed prediction. Specifically, we propose a spatial generative adversarial network (GAN) and a spectrum GAN to improve image fidelity. The spatial and spectrum GANs aim at penalizing the distribution discrepancy between generated and real images from the spatial domain and spectral domain, respectively. In addition, a masked style loss is devised to further enhance the details by transferring the detailed texture of ground truth radar sequences to extrapolated ones. We apply a foreground mask to prevent the background noise from transferring to the outputs. Moreover, we also design a new metric termed the power spectral density score (PSDS) to quantify the perceptual quality from a frequency perspective. The PSDS metric can be applied as a complement to other visual evaluation metrics (e.g., LPIPS) to achieve a comprehensive measurement of image sharpness. We test our approaches with both ConvLSTM baseline and U-Net baseline, and comprehensive ablation experiments on the SEVIR dataset show that the proposed approaches are able to produce much more realistic radar images than baselines. Most notably, our methods can be readily applied to any deep-learning-based spatiotemporal forecasting models to acquire more detailed results.