DDNet: disaster damage detection for buildings based on dual-temporal joint attention network

Area of Science:

• Remote Sensing
• Computer Vision
• Disaster Management

Background:

• Accurate building damage assessment post-disaster is vital for emergency response.
• Satellite imagery analysis, involving building localization and damage classification, is a common approach.
• Existing methods face challenges with class imbalance, boundary deviations, and feature extraction from temporal images.

Purpose of the Study:

• To propose a novel two-stage network, DDNet, for enhanced building damage detection.
• To address limitations in building localization and damage classification accuracy.
• To improve the overall effectiveness of post-disaster building damage assessment.

Main Methods:

• A two-stage network (DDNet) was developed.
• The first stage uses differential upsampling and unified focal loss for improved building localization.
• The second stage incorporates a joint attention module for effective feature extraction from multi-temporal satellite images.

Main Results:

• DDNet achieved a total F1 score of 79.56% on the xBD dataset.
• Localization F1 score reached 86.38%, and damage classification F1 score was 76.64%.
• The network effectively mitigated class imbalance and improved feature mining.

Conclusions:

• The proposed DDNet framework significantly enhances building damage assessment accuracy.
• The two-stage approach effectively addresses challenges in localization and classification.
• DDNet shows strong performance for post-disaster building damage detection using satellite imagery.