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

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

Problem Statement
Earthquakes cause massive building collapses and human casualties, necessitating rapid and reliable damage assessment for emergency rescue operations. Change detection (CD) using multi-temporal remote sensing (RS) images is a critical tool for this purpose. While deep neural network (DNN)-based methods have achieved impressive performance in general CD tasks, they face significant limitations when applied to post-earthquake scenarios with short imaging intervals. Specifically, when post-event images are acquired shortly after a disaster (e.g., within days), the varying imaging angles between pre- and post-event images lead to "side-looking" problems. These geometric mismatches, particularly in areas with high-rise buildings, cause spatial misalignments that degrade the accuracy of existing CD methods. Furthermore, a lack of specialized datasets with short acquisition intervals hinders the development and evaluation of methods tailored for immediate post-disaster relief.

Methodology: MSI-Net
To address these challenges, the authors propose a Multi-Scale Feature Interaction Network with Offset Calibration (MSI-Net). The architecture is built upon a weight-sharing Siamese encoder (based on ResNet-18) to extract multi-scale features from bi-temporal RS images. The network integrates three novel modules:

1. Joint Cross-Attention (JCA) Module: Designed to facilitate sufficient interaction between bi-temporal features across different scales. It comprises two sub-blocks:

   • Channel Cross-Attention (CCA): Enhances information exchange between temporal images by calculating cross-temporal attention maps, making feature distributions more similar.
   • Spatial Joint Attention (SJA): Utilizes both cross-attention and self-attention mechanisms to explore spatial detail distributions and capture long-range dependencies. It concatenates queries from both temporal images to simultaneously achieve self-attention and cross-attention, effectively modeling global spatial relationships.
   • Note: Shallow features, which contain less semantic information, bypass the JCA module and are fed directly into the calibration module.

2. Multi-Scale Offset Calibration (MOC) Module: This module specifically targets the side-looking mismatch problem. It estimates an offset field between the bi-temporal features by convolving and concatenating them. Using this offset field, the module performs adaptive alignment via bilinear interpolation to correct spatial misalignments caused by different imaging angles. This calibration mitigates the negative impact of side-looking issues on change detection accuracy.

3. Feature Integration (FeI) Module: This module fuses the calibrated features from the MOC module with the multi-scale features from the JCA module. It employs a projection mechanism to reduce mismatches before concatenating the features for final change map prediction.

The network is optimized using a weighted cross-entropy loss function to address the class imbalance between changed and unchanged pixels.

Key Contributions
The paper presents three primary contributions:

1. TUE-CD Dataset: The authors constructed a new Change Detection dataset, the Turkey Earthquake CD dataset (TUE-CD), specifically for assessing building damage in the short term after an earthquake. It contains 1,656 pairs of 256×256 bi-temporal RS images captured by the WorldView-2 satellite within 5 days of the February 6, 2023, magnitude 7.8 earthquake in Turkey. The dataset explicitly includes samples with side-looking mismatches, providing a benchmark for short-interval disaster assessment.
2. MSI-Net Architecture: The proposal of a network specifically designed to suppress the impact of side-looking issues through the integration of JCA, MOC, and FeI modules, enabling better alignment of features from different temporal RS images.
3. Novel Module Design: The design of the JCA module (combining spatial joint and channel cross-attention) for sufficient feature interaction and the MOC module for accurate spatial calibration of bi-temporal features to mitigate spatial mismatches.

Experimental Results
The authors conducted extensive experiments on three datasets: WHU-CD (building changes), CLCD (cropland changes), and the newly constructed TUE-CD. MSI-Net was compared against nine state-of-the-art DNN-based CD methods (including FC-EF, BIT, GAS-Net, and HANet).

• WHU-CD: MSI-Net achieved the best performance in Precision (96.66%), mF1 (95.58%), mIoU (91.81%), and Overall Accuracy (99.27%).
• CLCD: MSI-Net attained the highest scores in Recall, mF1, mIoU, and OA, outperforming the second-best method (HANet) by 0.6% in mF1.
• TUE-CD: In the challenging earthquake scenario with severe side-looking issues, MSI-Net achieved the best mF1 (78.02%), mIoU (68.48%), and OA (96.16%). It demonstrated a 1.13% and 1.24% improvement in mF1 and mIoU over the second-best method (AMTNet), respectively.
• Ablation Studies: Experiments confirmed that removing any of the JCA, MOC, or FeI modules led to performance degradation. Specifically, removing the MOC module resulted in blurred edges and incomplete change regions due to uncorrected side-looking mismatches, while removing JCA led to incorrect estimation of irrelevant areas.

Significance and Claims
The paper claims that MSI-Net provides a more effective solution for post-earthquake building damage assessment compared to existing state-of-the-art methods, particularly in scenarios involving short imaging intervals where side-looking mismatches are prevalent. By introducing the TUE-CD dataset, the authors aim to fill the gap in available data for short-term disaster response. The proposed network's ability to efficiently interact with bi-temporal features and calibrate spatial offsets allows it to generate more complete edge information and reduce false positives in complex disaster scenarios. The results suggest that MSI-Net is a robust tool for immediate relief efforts following earthquakes.