Scholarlink

<p>The rapid and accurate three-dimensional (3D) data acquisition of urban fabric, including building information, can be achieved using airborne laser scanning (ALS), a cutting-edge 3D mapping technique. This paper presents a Deep Residual U-Net framework (DRU-Net3D) for urban building identification using 3D ALS point cloud data with three well-designed steps, namely point-based feature descriptor generation, point-based feature descriptor to image conversion, and DRU-Net3D method training and testing considering different levels of scene complexity and heterogeneity in the ALS datasets. The proposed DRU-Net3D model is straightforward to implement, where accurate building segmentation is effectively achieved in several complex cases, such as connected and overlapped objects with one of the objects as building; partially acquired building geometries and data gaps; different shapes, sizes and orientation of the buildings present in the ALS datasets. The proposed framework achieved an average Intersection-over-Union (IoU) of 81.56%, an overall accuracy of 94.27%, and an average F1-score of 82.16% across the evaluated airborne LiDAR datasets, demonstrating consistent segmentation performance under varying urban scene complexities. The proposed method was compared with various state-of-the-art techniques, namely U<b>-</b>Net, SegNet and DeepLabv3+, and demonstrated improved performance on the evaluated datasets. </p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DRU-Net3D: Deep Residual U-Net framework for urban building identification using three-dimensional airborne laser scanning point cloud data

  • 期刊论文

  • 发表日期：

    2026年2月21日

查看全文
  • Dheerendra Pratap Singh,
  • Manohar Yadav

摘要

The rapid and accurate three-dimensional (3D) data acquisition of urban fabric, including building information, can be achieved using airborne laser scanning (ALS), a cutting-edge 3D mapping technique. This paper presents a Deep Residual U-Net framework (DRU-Net3D) for urban building identification using 3D ALS point cloud data with three well-designed steps, namely point-based feature descriptor generation, point-based feature descriptor to image conversion, and DRU-Net3D method training and testing considering different levels of scene complexity and heterogeneity in the ALS datasets. The proposed DRU-Net3D model is straightforward to implement, where accurate building segmentation is effectively achieved in several complex cases, such as connected and overlapped objects with one of the objects as building; partially acquired building geometries and data gaps; different shapes, sizes and orientation of the buildings present in the ALS datasets. The proposed framework achieved an average Intersection-over-Union (IoU) of 81.56%, an overall accuracy of 94.27%, and an average F1-score of 82.16% across the evaluated airborne LiDAR datasets, demonstrating consistent segmentation performance under varying urban scene complexities. The proposed method was compared with various state-of-the-art techniques, namely U-Net, SegNet and DeepLabv3+, and demonstrated improved performance on the evaluated datasets.