Domain knowledge integration into deep learning for typhoon intensity classification.

Area of Science:

• Meteorology
• Computer Science
• Artificial Intelligence

Background:

• Accurate typhoon intensity estimation is crucial for disaster preparedness.
• Traditional methods often rely on manual analysis or less sophisticated algorithms.
• Deep learning offers potential for automated and improved intensity classification.

Purpose of the Study:

• To develop a high-accuracy deep learning technique for estimating typhoon intensity from satellite imagery.
• To incorporate meteorological domain knowledge into the deep learning model.
• To validate the model's performance and interpretability.

Main Methods:

• Utilized the Visual Geometric Group's (VGG-16) deep learning model.
• Applied fisheye distortion preprocessing to satellite images to enhance key typhoon features.
• Employed t-distributed stochastic neighbor embedding (t-SNE) for feature visualization and gradient-weighted class activation mapping (Grad-CAM) for interpretability.

Main Results:

• Achieved significantly higher classification accuracy compared to previous studies, even with sequential-split validation.
• Verified that fisheye preprocessing aids feature extraction and cluster formation in model feature maps.
• Grad-CAM highlighted relevant regions (typhoon eye, surrounding clouds), indicating expert-like analysis.

Conclusions:

• The proposed deep learning technique, enhanced with meteorological domain knowledge and fisheye preprocessing, significantly improves typhoon intensity classification accuracy.
• The integration of domain knowledge overcomes limitations of purely data-driven deep learning approaches.
• The model demonstrates interpretable results, aligning with expert meteorological insights.