Building Change Detection in Aerial Imagery Using End-to-End Deep Learning Semantic Segmentation Techniques

Automatic building change detection is essential for updating geospatial data, urban planning, and land use management. The objective of this study is to propose a transformer-based UNet-like framework for end-to-end building change detection, integrating multi-temporal and multi-source data to improve efficiency and accuracy. Unlike conventional methods that focus on either spectral imagery or digital surface models (DSMs), the proposed method combines RGB color imagery, DSMs, and building vector maps in a three-branch Siamese architecture to enhance spatial, spectral, and elevation-based feature extraction. We chose Hsinchu, Taiwan as the experimental site and used 1:1000 digital topographic maps and airborne imagery from 2017, 2020, and 2023. The experimental results demonstrated that the data fusion model significantly outperforms other data combinations, achieving higher accuracy and robustness in detecting building changes. The RGB images provide spectral and texture details, DSMs offer structural and elevation context, and the building vector map enhances semantic consistency. This research advances building change detection by introducing a fully transformer-based model for end-to-end change detection, incorporating diverse geospatial data sources, and improving accuracy over traditional CNN-based methods. The proposed framework offers a scalable and automated solution for modern mapping workflows, contributing to more efficient geospatial data updating and urban monitoring.