The lack of pixel-level labeling limits the practicality of deep learning-based building semantic segmentation. Weakly supervised semantic segmentation based on image-level labeling results in incomplete object regions and missing boundary information. This paper proposes a weakly supervised semantic segmentation method for building detection. The proposed method takes the image-level label as supervision information in a classification network that combines superpixel pooling and multi-scale feature fusion structures. The main advantage of the proposed strategy is its ability to improve the intactness and boundary accuracy of a detected building. Our method achieves impressive results on two 2D semantic labeling datasets, which outperform some competing weakly supervised methods and are close to the result of the fully supervised method.

Semantic segmentation is one of the hot topics in the field of remote sensing image intelligent analysis. Deep convolutional neural network (DCNN) has become a mainstream technology in semantic segmentation due to its powerful semantic feature representation. The emergence of high-resolution remote sensing imagery has provided massive detail information, but difficulties and challenges remain in the “feature representation of fine geo objects” and “feature distinction of easily confusing geo objects.” To this end, this article focuses on the distinguishing features of geo-object details and proposes a novel DCNN-based semantic segmentation. First, the cascaded relation attention module is adopted to determine the relationship among different channels or positions. Then, information connection and error correction are used to capture and fuse the features of geo-object details. The output feature representations are provided by the multiscale feature module. Besides, the proposed model uses the boundary affinity loss to gain accurate and clear geo-object boundary. The experimental results on the Potsdam and Vaihingen datasets demonstrate that the proposed model can achieve excellent segmentation performance on overall accuracy and mean intersection over union. Furthermore, the results of ablation and visualization analyses also verify the feasibility and effectiveness of the proposed method.