Building Change Detection in Earthquake: A Multi-Scale Interaction Network and A Change Detection Dataset

Researchers have developed MSI-Net, a deep learning model for detecting building damage in post-earthquake satellite imagery, and introduced the TUE-CD dataset based on the Turkey earthquake. The solution addresses the challenge of analyzing remote sensing images with short time intervals and varying imaging angles to support emergency response operations.

This research addresses a critical gap in disaster response technology by combining machine learning innovation with practical humanitarian application. Earthquakes cause massive economic damage and loss of life, yet emergency responders often lack timely damage assessments needed to allocate rescue resources effectively. The new Turkey Earthquake Change Detection (TUE-CD) dataset fills an important void by providing satellite imagery with short acquisition intervals—the realistic conditions faced during post-disaster assessments rather than the longer intervals in existing datasets.

The technical challenge here is significant. When satellites capture images in rapid succession following a disaster, they photograph from different angles due to orbital mechanics, creating geometric distortions that complicate analysis. The MSI-Net architecture tackles this through three integrated components: joint cross-attention modules that simultaneously process channel and spatial information, multi-scale offset calibration that geometrically aligns images despite viewing angle variations, and feature integration modules that fuse information across scales. This multi-layered approach represents meaningful progress in handling the complex real-world conditions of post-disaster imaging.

For the broader AI and remote sensing industry, this work demonstrates how specialized datasets paired with purpose-built architectures can solve domain-specific problems. The model's superior performance on multiple benchmarks suggests it could enhance disaster response capabilities worldwide. Organizations managing emergency relief operations, satellite imagery providers, and government agencies overseeing disaster management would benefit from integrating such technology. The research validates that AI systems designed around authentic operational constraints outperform general-purpose models, establishing a template for similar specialized applications in climate monitoring, infrastructure assessment, and environmental management.

• →MSI-Net achieves superior change detection performance by addressing geometric distortions in short-interval post-earthquake satellite imagery
• →The new TUE-CD dataset provides realistic training data with rapid acquisition intervals matching actual emergency response timelines
• →Joint cross-attention and multi-scale offset calibration modules effectively handle the technical challenges of analyzing images from different orbital angles
• →The model shows transferable capabilities, outperforming baselines on three separate datasets including existing benchmarks
• →AI-driven damage assessment from satellite data could significantly improve emergency response coordination and resource allocation