GA-SmaAt-GNet: Generative Adversarial Small Attention GNet for Extreme Precipitation Nowcasting

TL;DR

This paper tackles extreme precipitation nowcasting by introducing GA-SmaAt-GNet, a conditional GAN that fuses a two-stream generator (precipitation maps and binary precipitation masks) with an attention-augmented discriminator. The approach extends SmaAt-UNet (via a second encoder and CBAMs) and demonstrates superior performance for heavy rainfall on a 25-year KNMI Netherlands radar dataset, especially for 10–20 mm/h events and during summer. The authors perform comprehensive uncertainty analyses (epistemic via test-time dropout and aleatoric via learned log-variance) and provide Grad-CAM based visual explanations to illuminate which input regions drive predictions. The work highlights the value of incorporating auxiliary binary masks and attention mechanisms in meteorological nowcasting, with practical implications for warnings and risk management, while noting areas for future improvements such as alternative loss functions and additional data sources.

Abstract

In recent years, data-driven modeling approaches have gained significant attention across various meteorological applications, particularly in weather forecasting. However, these methods often face challenges in handling extreme weather conditions. In response, we present the GA-SmaAt-GNet model, a novel generative adversarial framework for extreme precipitation nowcasting. This model features a unique SmaAt-GNet generator, an extension of the successful SmaAt-UNet architecture, capable of integrating precipitation masks (binarized precipitation maps) to enhance predictive accuracy. Additionally, GA-SmaAt-GNet incorporates an attention-augmented discriminator inspired by the Pix2Pix architecture. This innovative framework paves the way for generative precipitation nowcasting using multiple data sources. We evaluate the performance of SmaAt-GNet and GA-SmaAt-GNet using real-life precipitation data from the Netherlands, revealing notable improvements in overall performance and for extreme precipitation events compared to other models. Specifically, our proposed architecture demonstrates its main performance gain in summer and autumn, when precipitation intensity is typically at its peak. Furthermore, we conduct uncertainty analysis on the GA-SmaAt-GNet model and the precipitation dataset, providing insights into its predictive capabilities. Finally, we employ Grad-CAM to offer visual explanations of our model's predictions, generating activation heatmaps that highlight areas of input activation throughout the network.

Figures (11)

• Figure 1: General overview of proposed (a) GA-SmaAt-GNet architecture with mask generation, (b) SmaAt-GNet used as generator, and (c) the Attention-Augmented Discriminator architecture.
• Figure 2: Preparation of the raw precipitation data. All pixels where there is no data available are set to zero and a center 64 by 64 cutout is taken from the center of the land mass of the Netherlands.
• Figure 3: An example of the precipitation mask generation process. For mask generation 12 consecutive radar frames are summed up to obtain the accumulated precipitation of one hour. 25 masks are generated from this accumulated precipitation, one for each integer threshold between 1 and 25 mm/h.
• Figure 4: MSE per time step calculated on the precipitation test set. The persistence model is not shown to improve visibility.
• Figure 5: Binary metrics per season for a threshold of 10 mm/h calculated on the test set. Given a specific metric, a $\uparrow$ indicates that higher values for that metric are better.