ICANEWS

Engineering & Technology

Multi-Scale Interaction Network for Building Change Detection in Post-Earthquake Environments

This research addresses the application of change detection (CD) to post-earthquake damage assessment, focusing on inferring destroyed regions from multi-tempor

Overview

This research addresses the application of change detection (CD) to post-earthquake damage assessment, focusing on inferring destroyed regions from multi-temporal remote sensing images. The study introduces a CD dataset, TUE-CD (Turkey earthquake CD dataset), specifically designed for evaluating building damage shortly after an earthquake, and proposes a multi-scale feature interaction network (MSI-Net) to handle challenges associated with short imaging intervals and differing angles.

Research Context

Earthquakes are described as destructive natural disasters causing economic losses globally. CD methods are identified as applicable for post-earthquake damage assessment. The efficacy of CD for emergency relief post-earthquake is noted to benefit from short imaging intervals between multi-temporal remote sensing images. However, a limitation for current deep neural network-based CD methods is the absence of datasets with such short imaging intervals. A specific challenge arising from short acquisition intervals of post-event images is the varying imaging angle between different temporal images, leading to side-looking problems.

Approach

To address the need for evaluating building damage in the short term after an earthquake, the researchers created the Turkey earthquake CD dataset (TUE-CD). This dataset was developed to support immediate relief efforts and to confront issues arising from short acquisition intervals, such as different imaging angles and side-looking problems.

Simultaneously, the research presents a multi-scale feature interaction network (MSI-Net). This network is designed for efficient interaction between bi-temporal features and to mitigate the effects of side-looking problems. The MSI-Net architecture comprises three primary modules:

  • Joint Cross-Attention (JCA) modules: These modules are designed to unify channel cross-attention and spatial joint attention, aiming for sufficient feature interaction.
  • Multi-scale Offset Calibration (MOC) modules: The function of these modules is to estimate offsets for aligning bi-temporal images, utilizing multi-scale features.
  • Feature Integration (FeI) modules: These modules fuse calibrated features and multi-scale features for predicting changed areas.

Findings

Experiments were conducted on three datasets: WHU-CD, CLCD, and the newly constructed TUE-CD dataset. The results indicate that the proposed MSI-Net provides better outcomes when compared to other state-of-the-art CD methods considered in the study.

Why This Matters

The application of change detection in post-earthquake damage assessment can contribute to inferring destroyed regions from remote sensing images. The development of a CD dataset with short imaging intervals, such as TUE-CD, and methods like MSI-Net that address challenges like side-looking problems, can potentially satisfy the needs of emergency rescues after earthquakes and contribute to post-disaster immediate relief.

Key Findings