Enhanced Building Change Detection in Aerial Imagery Using Siamese U-Net Architectures

Monitoring urban development and infrastructure changes is essential for sustainable city planning and disaster management. This paper investigates building change detection using two enhanced Siamese U-Net architectures: one utilizing feature concatenation and the other leveraging element-wise difference between bi-temporal aerial images. The LEVIR-CD dataset, comprising high-resolution paired images of urban landscapes, is employed to train and evaluate the models. Preprocessing includes slicing large images into manageable patches, applying augmentation techniques, and normalizing inputs. Experimental results demonstrate that the Siamese U-Net with difference outperforms its concatenation counterpart, achieving superior accuracy, F1-score, and intersection-over-union metrics. The findings highlight the effectiveness of dual-branch networks, particularly difference-based approaches, in capturing subtle structural transformations in complex urban environments. These results support the adoption of advanced Siamese architectures as robust tools for automated building change detection in remote sensing applications.