Deep-SiamChange: A Multi-Scale Attention-Based Siamese Network for Robust Structural Change Detection in Urban Environments

Abstract: The automatic identification of structural changes in urban environments through bitemporal satellite imagery presents substantial challenges stemming from environmental noise, illumination variations, and the inherent complexity of distinguishing genuine construction alterations from transient phenomena. Traditional change detection methodologies frequently succumb to the "noise challenge," wherein variable sun angles, atmospheric interference, and seasonal vegetation fluctuations generate false positives that obscure authentic building modifications. This investigation introduces Deep-SiamChange, a novel architecture that integrates a Siamese encoder with multi-scale attention mechanisms and convolutional block attention modules to achieve time-invariant and noise-robust feature extraction. The proposed framework processes bitemporal imagery through twin neural pathways with shared weights, ensuring consistent feature extraction logic across temporal intervals. A Feature Pyramid Network captures structural details across multiple scales, enabling the detection of both minor residential extensions and substantial industrial developments. The integration of channel and spatial attention mechanisms filters environmental noise by emphasizing geometric structural patterns while suppressing illumination-related artifacts. Experimental evaluation on the LEVIR-CD benchmark dataset, comprising 637 high-resolution bitemporal image pairs, demonstrates that Deep-SiamChange achieves an F1-score improvement from 83.9% to 87.3% compared to baseline implementations. The architecture exhibits particular effectiveness in mitigating misregistration errors and maintaining detection accuracy under varying illumination conditions. These findings establish Deep-SiamChange as a practical solution for urban governance applications, including automated illegal construction monitoring, property tax assessment, and post-disaster structural assessment.

Keywords: Change Detection, Remote Sensing, Siamese Networks, Attention Mechanisms, Multi-Scale Feature Fusion, Urban Analytics