Regularized Building Boundary Extraction From Remote Sensing Imagery Based on Augment Feature Pyramid Network and Morphological Constraint

Automatic building boundariesextraction methods are important for urban planning, change monitoring, and smart city construction. In this article, we propose a regularized building boundaries extraction from remote sensing imagery based on augment feature pyramid network (AFPN) and morphological constraint. First, we build an AFPN, which can provide more accurate and dense global features for semantic segmentation tasks to avoid the loss of feature information. Second, the building shape is manually divided into linear and curved by analyzing the morphological characteristics. The extraction results are regularized to achieve the refined expression of contour according to different types of building shapes. Finally, we conduct experiments on the benchmark dataset to test the availability of the proposed approach. The results showed that the F1-score and intersection over union (IOU) reached 93.7% and 88.8%, respectively. Besides, our proposed approach is compared with some of the excellent research in recent years of models, such as PSPNet, Unet++, RefineNet, and DeconvNet. On the benchmark dataset, the proposed method increases the IOU by 0.9–2.7% and improves the F1-score by 0.2–2.5%. In addition, the results prove that our method considering morphological constraint can achieve better visual effects.