Post-disaster building damage assessment based on improved U-Net.

Area of Science:

• Remote Sensing
• Computer Vision
• Disaster Management

Background:

• Accurate post-disaster building damage assessment is crucial for effective disaster response.
• High-resolution remote sensing technology offers a foundation for damage information extraction.
• Existing methods struggle with building positioning inaccuracies and classification challenges.

Purpose of the Study:

• To develop an improved deep learning network for accurate building damage assessment.
• To address limitations in building segmentation and classification in previous methods.
• To enhance the extraction of post-disaster building damage information.

Main Methods:

• Designed a two-stage network based on U-Net: a U-Net for building segmentation and a Siamese U-Net for damage classification.
• Incorporated Extra Skip Connection and Asymmetric Convolution Block for multi-scale building segmentation.
• Utilized Shuffle Attention to correlate pre- and post-disaster building information.
• Trained and tested the network on the xBD dataset.

Main Results:

• Achieved a balanced F1-score of 0.8741 for building localization.
• Achieved a balanced F1-score of 0.7536 for building damage classification.
• Demonstrated superior overall performance compared to existing building damage assessment methods.

Conclusions:

• The proposed two-stage network significantly improves building damage assessment accuracy.
• The network shows good generalization capabilities across multiple disaster scenarios.
• This approach provides a more reliable method for rapid post-disaster building damage information extraction.