A multiscale building detection method based on boundary preservation for remote sensing images: Taking the Yangbi M6.4 earthquake as an example

Abstract

Building detection is an important topic in remote sensing image applications. This study proposes a multiscale building detection method based on boundary preservation, to detect building roofs from a large number of high-resolution remote sensing images at a fast speed. A lightweight network extracts a feature map and a feature pyramid network (FPN) aggregates low- and high-level features. A multidimensional attention network (MDA) enhances building features and weakens the complex background information. The method uses four branches to extract buildings: classification, box, direction box, and mask. We manually labeled approximately 870,000 buildings of different types and selected about 300 1 ​km ​× ​1 ​km image plots of different ground objects without buildings, to construct positive and negative sample sets for 27 provinces in China. The accuracy and recall of test results of the proposed method are 12.4% and 3.6% higher than those of Mask R–CNN, respectively, while its accuracy and detection time of segmentation results are 6% and about 30% higher than those of Mask R–CNN, respectively. Pre-disaster buildings were extracted using the proposed method in several key provinces across the country, which were applied to quick assessment in emergency work for Yangbi earthquake. The method was used on post-disaster UAV images as well, achieving 95.61% precision and 91.36% recall of detection results. Experiments show that: the detection method and its results are beneficial to reduce manual interpretation time significantly, and detections on pre- and post-disaster images can be compared to help identification of damaged buildings.