Using High-Resolution Remote Sensing Images for Extracting Buildings Through the Jointly Improved YOLOv8 Deep Learning Network and Object-Oriented Analysis

Remote sensing images are crucial for acquiring ground information, with buildings being significant ground targets. This paper addresses the limitations of traditional methods in extracting buildings from high-resolution remote sensing images, which often fail to maintain regional integrity and produce clear edges. To tackle this issue, we propose a method that combines edge-guided multi-scale segmentation for building extraction. Firstly, we enhance the Dense Extreme Inception Network for Edge Detection (DexiNed) with an attention mechanism to detect building edges and perform post-processing on results. Next, we optimize the multi-scale segmentation results by incorporating edge feature information to obtain final segmentation results. Lastly, we improve the popular YOLOv8 network for classification. The experiments were conducted on WHU Aerial Building Dataset and Massachusetts Buildind Dataset, achieving Intersection over Union (IoU) of 91.95% and 74.91 % respectively. Compared to other methods, our approach enhances the accuracy of building extraction, produces more complete building regions with clear edges, validating the effectiveness of our proposed method.