A Context-Structural Feature Decoupling Change Detection Network for Detecting Earthquake-Triggered Damage

Identifying damaged areas after earthquakes is essential for emergency rescue and rebuilding efforts. Remote sensing image change detection can facilitate these tasks by analyzing features differences between bitemporal/multi-temporal optical images. However, the existing methods still face challenges due to the unpredictable timing, locations, and magnitudes of earthquakes. Firstly, the spectral features of optical images vary over time, making it difficult to differentiate actual damage from areas appearing damaged due to spectral noise. Secondly, variations in earthquake characteristics can cause different types and scales of damage, from individual buildings in urban areas to extensive landslides in mountain areas. Overcoming these variations to establish capabilities for detecting multiple types of damage is also critical. To tackle these challenges, this study introduces a change detection network based on the decoupling of context-structural features (referred to as CSDNet). CSDNet uses a CNN-transformer architecture and decouples image features into structural and context branches through a feature difference extractor. Subsequently, the structural branch extracts reliable features through multi-scale consistency feature fusion, ensuring robust damage detection across multiple scales. Context branch facilitates long-range supervision and reduces the impacts of pseudo-changes through partial feature exchange between bitemporal images. Furthermore, to enhance CSDNet’s adaptability across different scenarios and explore CSDNet’s potential for solving multiple tasks within a unified framework, an adaptive module has been designed to dynamically fuse structural and context features. Experiments were conducted on two representative tasks: individual building damage with partial rebuilding and large-scale landslides, to investigate CSDNet’s capabilities. As a result, CSDNet achieved F1-Scores of 92.2% and 90.2% on the two tasks respectively, indicating the effectiveness of CSDNet.