Building segmentation is an important step in urban planning and development. In this work, we propose a new deep learning model, namely Multidimension Attention U-Net (MDAU-Net), to accurately segment building pixels and nonbuilding pixels in remote sensing images. Furthermore, we introduce a novel Multidimension Modified Efficient Channel Attention (MD-MECA) model to enhance the network discriminative ability through considering the interdependence between feature maps. Through deepening the U-Net model to a seven-story structure, the ability to identify the building is enhanced. We apply MD-MECA to the "skip connections"in traditional U-Net, instead of simply copying the feature mapping of the contraction path to the matching extension path, to optimize the feature transfer more efficiently. The obtained results show that our proposed MDAU-Net framework achieves the most advanced performance on publicly available building data sets (i.e. the precision over the Massachusetts buildings data set and WHU data set are 97.04% and 95.68%, respectively). Furthermore, we observed that the proposed framework outperforms several state-of-the-art approaches.
ASJC Scopus subject areas
- Computer Science Applications