National-scale mapping of building footprints using feature super-resolution semantic segmentation of Sentinel-2 images

Since buildings are closely related to human activities, large-scale mapping of individual buildings has become a hot research topic. High-resolution images with sub-meter or meter resolution are common choices to produce maps of building footprints. However, high-resolution images are both infrequently collected and expensive to obtain and process, making it very difficult to produce large-scale maps of individual buildings timely. This paper presents a simple but effective way to produce a national-scale map of building footprints using feature super-resolution semantic segmentation of sentinel-2 images. Specifically, we proposed a super-resolution semantic segmentation network named EDSRNASUnet, which is an end-to-end network to generate semantic maps with a spatial resolution of 2. 5 m from real remote sensing images with a spatial resolution of 10 m. Based on the dataset consisting of images from 35 cities in China, we quantitatively compared the proposed method with three methods under the same framework and qualitatively evaluated the identification results of individual buildings. In addition, we mapped building footprints within the entire China at 2. 5 m-resolution using Sentinel-2 images of 10 m resolution. The density of building footprints varies considerably across China, with a gradual increase in building footprints from west to east, i. e. from the first step of China’s terrain to the third one. We detected over 86. 3 million individual buildings with a total rooftop area of approximately 58, 719. 43 km2. The number of buildings increased from 5. 73 million in the first step of China’s terrain, through 23. 41 million in the second step of China’s terrain, to 57. 16 million in the third step of China’s terrain. The area of buildings also increased from 3318. 02 km2 through 13, 844. 29 to 41, 557. 12 km2. The Aihui-Tengchong line, a dividing line representing the regional distribution of China’s population, also divides the regional distribution of Chinese buildings. Our approach has a more open and practical application because of the medium-resolution images and platform with open access. Results are available to the community (https: //code. earthengine. google. com/? asset=users/flower/2019China).