Broad-UNet: Multi-scale feature learning for nowcasting tasks.

Area of Science:

• Meteorology and Atmospheric Sciences
• Computer Science and Artificial Intelligence
• Remote Sensing

Background:

• Accurate weather nowcasting is crucial for numerous human activities.
• The problem of weather nowcasting is increasingly being addressed using advanced computational models.
• Satellite imagery provides a rich data source for high-resolution meteorological predictions.

Purpose of the Study:

• To introduce a novel deep learning architecture, Broad-UNet, for enhanced weather nowcasting.
• To adapt the image-to-image translation framework for meteorological prediction tasks.
• To improve the efficiency and accuracy of short-term, high-resolution weather forecasts.

Main Methods:

• Developed Broad-UNet, a novel architecture based on the UNet model.
• Incorporated asymmetric parallel convolutions and Atrous Spatial Pyramid Pooling (ASPP) module into Broad-UNet.
• Treated weather nowcasting as an image-to-image translation problem using satellite imagery.

Main Results:

• Broad-UNet effectively combines multi-scale features using fewer parameters than the core UNet model.
• The model demonstrated superior performance in precipitation map and cloud cover nowcasting tasks.
• Numerical results indicated that Broad-UNet achieved more accurate predictions compared to existing architectures.

Conclusions:

• Broad-UNet is an efficient and accurate deep learning model for weather nowcasting.
• The proposed architecture advances the state-of-the-art in short-term meteorological prediction using satellite data.
• The integration of multi-scale feature extraction enhances the capability of nowcasting models.