GeoAI Dataset for Building Change Detection from CAS500-1

Building change detection (BCD) is essential for urban planning, disaster response, and environmental monitoring. Existing datasets, such as WHU-CD, LEVIR-CD, and SYSUCD, as well as overseas satellite imagery datasets, have limitations in reflecting South Korea’s distinctive high-density urban morphology. Domestic aerial datasets, including AI-based national park change monitoring and urban BCD datasets, also suffer from varying sensor resolutions and limited applicability to satellite-based monitoring. To overcome these challenges, we introduce a high-resolution BCD dataset based on Compact Advanced Satellite 500-1 (CAS500-1), launched in 2021. This dataset utilized 0.5 m/pixel satellite imagery collected from 2021 to 2024 across diverse regions, including Seoul, Busan, and Daejeon. The dataset construction involved precise image alignment, RGB band synthesis, histogram stretching for contrast enhancement, and polygon-based labeling to identify urban changes. Additionally, format conversions were applied to ensure deep learning compatibility, generating 3,803 paired image patches (256×256 pixels). To validate the dataset, we employed ChangeMamba, a model integrating a visual state space model for spatiotemporal change detection. The model trained on our dataset achieved an F1 score of 0.79 and an intersection over union (IoU) of 0.66, outperforming models trained on different BCD datasets in cross-dataset tests, where the latter scored significantly lower (e.g., F1, 0.18 and IoU, 0.10 for the model trained on LEVIR-CD). Moreover, our dataset demonstrated improved consistency over domestic aerial datasets, which often suffer from heterogeneous resolutions and limited scalability for long-term monitoring. By leveraging high-resolution satellite imagery, our dataset provides a robust alternative to both overseas and domestic aerial datasets, enabling precise urban monitoring with reduced reliance on aerial surveys. It serves as a critical resource for advancing BCD research in South Korea. Future research will focus on automating the labeling process, expanding temporal coverage, and improving model performance through reinforcement learning on pre-trained models using this dataset.