DeepLab-Bam: An Accurate Building Extraction Method from High-Resolution Remote Sensing Images

Buildings, as the primary spatial hubs for human activities, are intrinsically connected to various aspects of human society. Accurately extracting their spatial location and geometric structure from high-resolution remote sensing images is crucial for various applications. However, the performance of existing building extraction methods is limited due to ignorance of significant intra-class variability and small inter-class differences. To achieve accurate building extraction from high-resolution remote sensing images, we propose a novel building extraction network model, named DeepLab-Bam. Based on the DeepLabV3+ backbone, we first design a novel CSC module by combining channel attention, spatial attention, and context-aware attention through multiplication, effectively enhancing the model’s ability to capture contextual relationships. Then, an improved ASPP module is proposed with strip pooling, to enhance the model’s ability to acquire receptive field and semantic information at different scales. Finally, an attention mechanism based upsampling module is proposed to improve the retention and prominence of important features. Experiments on the WHU-Building dataset demonstrate that DeepLab-Bam outperforms DeepLabV3+ with improvements of 0.74%, 1.79%, and 3.07% in terms of intersection ratio, accuracy, and precision, respectively. It also shows significant improvements over Unet, PsPnet, and Segnet. The results confirm that DeepLab-Bam delivers superior extraction accuracy and robustness in building extraction from high-resolution remote sensing images.