Published: 1 Jan 2026
Short-term weather prediction, also known as weather nowcasting, aims to forecast meteorological events in the near future with high spatial resolution. These predictions are extremely important for many applications such as aviation, agriculture, transportation, and disaster management.
In this research, we introduce Broad-UNet, a deep learning architecture designed to improve weather nowcasting by capturing multi-scale spatial patterns from satellite imagery.
Weather forecasting plays a crucial role in modern society. Accurate predictions help industries such as agriculture, mining, and construction plan their operations and reduce economic losses caused by unexpected weather events.
Traditional nowcasting methods mainly rely on:
Numerical Weather Prediction (NWP)
These models simulate atmospheric physics but require large computational resources.
Optical Flow techniques
These methods estimate motion between image frames but struggle to capture the physical dynamics of weather systems.
Recent advances in data-driven deep learning models have opened new possibilities for improving short-term weather forecasting.
Broad-UNet extends the standard UNet architecture to better capture multi-scale weather patterns.
The model introduces two key innovations:
Instead of using a single convolution kernel, the model processes data through parallel convolution branches with different kernel sizes.
These branches capture patterns at different spatial scales.
The outputs are then combined to create richer feature representations.
Broad-UNet integrates an ASPP module in the bottleneck of the network.
ASPP uses multiple convolution filters with different dilation rates, allowing the network to capture information at different spatial resolutions.
This enables the model to extract:
local features
global context information
simultaneously.
Broad-UNet was evaluated on two different nowcasting tasks.
Predict rainfall maps 30 minutes ahead using radar imagery.
Predict cloud presence 15 to 90 minutes ahead using satellite imagery.
The datasets include:
radar rainfall data from the Royal Netherlands Meteorological Institute (KNMI)
satellite cloud images from EUMETSAT.
Experiments show that Broad-UNet performs better than standard UNet and several other deep learning models.
Key findings include:
improved prediction accuracy for precipitation maps
better short-term cloud cover predictions
more efficient feature extraction using multi-scale filters
For precipitation nowcasting, Broad-UNet achieved the lowest Mean Squared Error (MSE) among all compared models.
Weather phenomena occur at many spatial scales:
small cloud formations
large storm systems
regional precipitation patterns
Broad-UNet captures these patterns by combining:
parallel convolution branches
multi-scale dilation filters
hierarchical feature extraction
This allows the model to understand both local weather structures and larger atmospheric patterns.
Broad-UNet demonstrates how multi-scale feature learning can significantly improve short-term weather prediction.
By extending the classic UNet architecture with:
multi-branch convolution blocks
ASPP modules
parameter-efficient convolutions
the model achieves better performance while remaining computationally efficient.
This architecture represents an important step toward real-time AI-driven weather forecasting systems.
This blog post is based on the research paper:
Jesús García Fernández, Siamak Mehrkanoon, Broad-UNet: Multi-scale feature learning for nowcasting tasks, Neural Networks, 2021.