|
Building footprint information is crucial for urban-related applications. Due to the rapid growth of the number and types of community-level buildings, the completeness and coverage of building footprint data at the community level are lagging. Efficiently and accurately generating community-level building footprint maps for the update of community electronic maps, post-disaster assessment, navigation services, etc. is an urgent and challenging task. To fill the data gaps of community-level building footprint with high accuracy, we design a building map translator named BuildingMap- GAN that is configured with a novel generator: MACU-Net. It has multi-scale connections with channel attention and asymmetric convolution blocks (ACB), which are used for fully capturing the spatial structure information of buildings and roads. An experiment is conducted on real-world community dataset in Wuhan, China, in which the street network data of 700 communities were used for training and the street network data of 42 communities were used for verification. Compared with the two baseline models, F1 scores are improved by 6.7% and 4.5%, respectively, and IoU values increase by 5.0% and 3.2%, respectively. Our model can generate refined building footprint data with higher precision and is closer to the real building distribution.
|
