Aligning Discriminative and Representative Features: An Unsupervised Domain Adaptation Method for Building Damage...

Area of Science:

• Remote Sensing
• Machine Learning
• Disaster Management

Background:

• Building damage assessment is critical post-hurricane, but labeled data for machine learning classifiers is scarce and costly.
• Existing machine learning methods struggle with diverse remote sensing imageries from different disaster events due to low inter-category discrimination and high intra-category diversity.

Purpose of the Study:

• To propose an unsupervised domain adaptation method (ADRF) for building damage classification using labeled data from relevant disaster scenes.
• To address the challenge of domain shift in remote sensing imageries from different hurricane disasters.

Main Methods:

• Developed a framework utilizing aligned discriminative and representative features (ADRF).
• Employed three pipelines: a classifier for source domain data and autoencoders for both source and target domains.
• Aligned marginal distributions of source and target domains using Maximum Mean Discrepancy (MMD) to achieve domain invariance.

Main Results:

• Tested ADRF on challenging transfer tasks using hurricane Sandy, Maria, and Irma datasets.
• Achieved 71.6% accuracy in transferring from Sandy to Maria.
• Achieved 84.1% accuracy in transferring from Sandy to Irma.

Conclusions:

• The proposed ADRF method effectively improves feature discriminative and representative capabilities for cross-disaster building damage classification.
• Unsupervised domain adaptation is a viable approach to overcome data limitations in post-disaster remote sensing analysis.