A benchmark GaoFen-7 dataset for building extraction from satellite images

This study constructs a high-resolution (0.65 m) GaoFen-7 (GF-7) Building dataset utilizing the Chinese GF-7 imagery from Tianjin, Lanzhou, Chongqing, Ningbo, Guangzhou, and Shenzhen. The dataset comprises 5175 pairs of 512×512 image tiles (including 3106 pairs in the training set, 1034 pairs in the validation set, and 1035 pairs in the test set), totally covering an area of 573.17 km² in China. It contains a total of 170,015 buildings, out of which 145,327 buildings (84.8%) are distributed in urban areas, while the remaining 24,688 buildings (15.2%) are distributed across rural areas. These high-quality labels are mainly generated through manual digitization.