A Multi-Temporal Remote Sensing Image Dataset for Building Change Detection in Chinese Regions

Over the past few decades, a large collection of feature learning and segmentation methods have been developed for building change detection (BCD) in multi-temporal remote sensing images. Focused on this, various datasets were built up for analyzing changes in different countries, lacking attention to different regions in China. In this paper, a new Chinese Region BCD (CRBCD) dataset was established, which collects image pairs between 2017-2022 years in 31 cities of China from Google Earth. In specific, the dataset covers changing buildings in multiple types of regions, including urban, sub-urban, rural and scenic areas. Besides, the differences between constructed dataset and some existing typical BCD datasets are comprehensively compared and analyzed. In addition, considering the local spatial consistency distribution characters, a structure prior guided convolutional neural network (SPG-CNN) framework is introduced here. Both the multi-scale superpixel segmentation and high-frequency derived edge information are inserted to the CNN, improving the difference feature learning capability. The quantitative and visual comparison results on both the public dataset and the proposed dataset demonstrate the effectiveness of the proposed SPG-CNN. The CRBCD dataset can be found at https://github.com/cpfhnu/CRBCD.