BEARNet: A Novel Buildings Edge-Aware Refined Network for Building Extraction From High-Resolution Remote Sensing Images

Abstract

Accurately extracting buildings from high-resolution remote sensing images is important to obtain urban information, and promote the development of smart cities. At present, the knowledge-driven building extraction methods mostly rely on building prior knowledge of manual design, resulting in low automation and poor versatility. The data-driven methods usually rely too much on training samples, resulting in insufficient pertinence for the features of building extraction and low generalization ability of the model. Therefore, this study proposes a novel buildings edge-aware refined deep learning network (BEARNet) for building extraction from high-resolution remote sensing images. The network takes the building edge as a priori knowledge, learns the building edge features by decoupling the building body and edge, and further optimizes the network by combining the multi-objective loss function to strengthen the pertinence of building edge features extraction. Experimental results show that on the WHU building dataset, which is less difficult to extract buildings, compared with other methods, BEARNet has the highest Precision, F1, IoU and OA values of 97.70%, 97.42%, 95.3% and 98.67%. On the Massachusetts building dataset, where building extraction is difficult, BEARNet has the highest Precision, Recall, F1, IoU and OA compared to other methods, with values of 84.92%, 85.27%, 85.09%, 75.82% and 93.99%, respectively. Our proposed method is more accurate in extracting complex shapes and dense small-scale buildings, and the building edges are more refined and complete.