Dataset Construction and Semantic Segmentation of Buildings and Roads from CAS500-1 Images

The current status of buildings and roads is an essential element of land monitoring and management, and is used in various fields such as urban planning, disaster management, and change detection. With the recent advancement of deep learning technology, high-performance semantic segmentation techniques have become commonplace, and research on applying them to building and road detection has begun to be actively conducted. However, most of these studies are conducted overseas, so there are still limitations in applying them to Korean field. Therefore, in this study, we constructed a building and road detection dataset optimised for Korea using high-resolution Compact Advanced Satellite (CAS500) images and digital maps, and developed a detection model using the latest semantic segmentation models, the unified perceptual parsing network (UPerNet) based on the sifted window (Swin) transformer and the masked-attention mask transformer (Mask2Former), and compared and evaluated their performance. In particular, a systematic refinement process was used to minimise the spatiotemporal discrepancies between the satellite imagery and the digital maps, and a temporary dataset consisting of only pure digital map labels was created separately to evaluate the effectiveness of the dataset built in this way. The results of the building and road detection experiments showed that the Mask2Former model using the Swin-L backbone performed the best with a building intersection over union (IoU) of 77.93 and a road IoU of 74.85. In addition, the model trained with the optimised dataset in this study performed qualitatively and quantitatively better than the model trained with the pure digital map dataset, proving its effectiveness. The methodology and results of this study are expected to contribute to the efficiency of land status information construction and monitoring in Korea, and furthermore, it is expected that it can serve as a practical basis for the advancement of land spatial information services and the establishment of related policies.