Published: 1 Feb 2026
Extreme rainfall events can lead to severe flooding, infrastructure damage, and major economic losses. Predicting these events even minutes to hours in advance is crucial for early warning systems and disaster prevention.
In this work, we introduce GA-SmaAt-GNet, a deep learning framework designed to improve extreme precipitation nowcasting using radar data and generative adversarial networks.
Extreme precipitation events have become more frequent in many regions, including the Netherlands. Such events can overwhelm drainage systems and lead to urban flooding.
In operational meteorology, nowcasting focuses on predicting weather conditions in the short term, typically from a few minutes up to two hours ahead.
Traditional forecasting approaches include:
Numerical Weather Prediction (NWP)
• physics-based atmospheric models
• computationally expensive
• limited ability to capture small-scale rainfall events
Machine learning models, especially deep learning, provide an alternative by learning patterns directly from historical radar observations.
To improve predictions for extreme rainfall events, the proposed framework combines two ideas:
1️⃣ SmaAt-GNet generator architecture
2️⃣ Generative adversarial learning
This combination allows the system to produce more realistic precipitation predictions, especially during high-intensity rainfall events.
The generator network builds on the SmaAt-UNet architecture, which is an efficient UNet-based model designed for precipitation nowcasting.
In this work, the architecture is extended to include precipitation masks as an additional input.
Precipitation masks represent binarized rainfall regions, helping the model better identify areas with active precipitation.
By incorporating this additional information, the network can better learn the spatial structure of extreme rainfall events.
The framework uses a Generative Adversarial Network (GAN) setup.
This consists of two networks:
Produces predicted precipitation maps.
Attempts to distinguish between:
real radar observations
generated predictions.
The discriminator includes an attention mechanism that focuses on important spatial regions of the radar images.
During training, the generator learns to produce increasingly realistic rainfall predictions in order to “fool” the discriminator.
The model was evaluated using a real precipitation dataset from the Netherlands, covering more than 25 years of radar observations.
Extreme precipitation events are defined as rainfall intensities exceeding 20 mm per hour, which is significant because many drainage systems cannot handle larger rainfall rates.
The experiments compare GA-SmaAt-GNet with several established deep learning architectures, including:
RainNet
SmaAt-UNet
EarthFormer.
The experiments show that the proposed model improves precipitation nowcasting performance.
Major results include:
• improved prediction accuracy for extreme rainfall events
• better spatial structure in generated precipitation maps
• enhanced detection of high-intensity precipitation regions
The GAN-based training also helps produce more realistic rainfall patterns, especially during severe weather events.
To interpret the model’s predictions, the study uses Grad-CAM visualization techniques.
Grad-CAM generates activation heatmaps showing which regions of the input radar images influence the model’s prediction.
These visualizations reveal that the model focuses on areas with strong precipitation activity, confirming that it learns meaningful meteorological patterns.
Extreme precipitation events remain one of the most challenging problems in short-term weather prediction.
GA-SmaAt-GNet demonstrates that combining:
efficient UNet-based architectures
precipitation masks
generative adversarial training
can significantly improve the prediction of high-intensity rainfall events.
Such models represent an important step toward AI-driven early warning systems for severe weather.
This blog post is based on the research paper,
Eloy Reulen, Jie Shi, Siamak Mehrkanoon, GA-SmaAt-GNet: Generative adversarial small attention GNet for extreme precipitation nowcasting, Knowledge-Based Systems, 2024.