Published: 1 Jan 2026
Predicting rainfall in the next minutes or hours, also known as precipitation nowcasting, is one of the most important challenges in weather forecasting. Accurate short-term predictions help aviation, transportation, energy systems, agriculture, and even everyday planning.
In this post, I discuss our work on SmaAt-UNet, a compact deep learning architecture designed for efficient precipitation nowcasting.
Most operational weather forecasts rely on Numerical Weather Prediction (NWP) models. These systems simulate atmospheric physics using complex equations describing temperature, pressure, and wind dynamics.
Although NWP models are powerful, they have two major limitations:
They require massive computational resources
They are less suitable for very short-term forecasts (minutes to a few hours)
Nowcasting methods instead rely on recent observational data, such as weather radar images, to predict near-future conditions.
Deep learning has emerged as a powerful tool for this type of data-driven prediction.
Weather radar systems continuously measure precipitation intensity across large geographical areas. These radar scans can be represented as rainfall maps, where each pixel corresponds to rainfall intensity at a particular location.
In our approach:
The model receives 12 radar images representing the last hour of precipitation observations.
The task is to predict the rainfall map 30 minutes into the future.
This transforms precipitation forecasting into a spatiotemporal image prediction problem.
To solve this problem, we use the UNet architecture, originally designed for image segmentation.
UNet has a characteristic encoder–decoder structure:
Encoder (downsampling):
extracts hierarchical spatial features
Decoder (upsampling):
reconstructs high-resolution output
Skip connections link encoder and decoder layers so the model retains spatial information.
The architecture forms a U-shaped network, enabling the model to capture both large-scale weather patterns and fine local details.
The proposed Small Attention-UNet (SmaAt-UNet) modifies the original UNet architecture to make it more efficient while maintaining strong predictive performance.
Two main improvements are introduced.
We incorporate Convolutional Block Attention Modules (CBAM) into the encoder layers.
Attention mechanisms help the network identify the most relevant parts of the radar images by focusing on:
intense rain cells
moving storm structures
spatial regions with significant precipitation activity.
This allows the network to prioritize important features while suppressing irrelevant information.
The second improvement replaces standard convolution operations with depthwise-separable convolutions.
This technique separates spatial filtering from channel mixing, drastically reducing the number of parameters in the network.
As a result:
Original UNet: about 17 million parameters
SmaAt-UNet: about 4 million parameters
This represents a four-fold reduction in model size while maintaining comparable performance.
The experiments showed that:
Deep learning models significantly outperform simple baseline methods such as persistence forecasting.
The proposed SmaAt-UNet performs nearly as well as the original UNet.
The new architecture requires only one quarter of the parameters of the original network.
This demonstrates that efficient architectures can maintain high prediction quality while greatly reducing computational requirements.
The model was evaluated using radar precipitation data from the Royal Netherlands Meteorological Institute (KNMI). The dataset contains radar rainfall maps collected every 5 minutes over several years, resulting in hundreds of thousands of samples.
The prediction task consisted of forecasting rainfall intensity 30 minutes into the future for each pixel of the radar map.
Performance was evaluated using several metrics including:
Mean Squared Error (MSE)
Precision
Recall
F1-score
Critical Success Index (CSI)
Reducing model size opens new opportunities for practical deployment.
Smaller networks can:
run faster in real-time forecasting systems
require less training time
consume less computational energy
be deployed on edge devices or mobile platforms.
Efficient architectures are therefore an important step toward sustainable and scalable AI-based weather forecasting systems.
SmaAt-UNet shows that it is possible to build compact and efficient neural networks for precipitation nowcasting without sacrificing performance.
By combining attention mechanisms and depthwise-separable convolutions, the model achieves strong predictive accuracy with significantly fewer parameters than standard architectures.
Such efficient deep learning models could play a key role in enabling real-time, data-driven weather forecasting systems in the future.
This post is based on the research paper:
Kevin Trebing, Tomasz Staǹczyk, Siamak Mehrkanoon, SmaAt-UNet: Precipitation nowcasting using a small attention-UNet architecture, Pattern Recognition Letters, 2021.