Automatic Building Extraction from Multispectral LiDAR Using Novel 3D Spatial Indices and Deep Learning Point CNN

In the dynamic landscape of urban development, the demand for precise urban extraction methods is paramount. A multispectral light detection and ranging (LiDAR) system simultaneously captures spatial geometry and multi-wavelength intensity data, allowing for three-dimensional point cloud segmentation and feature extraction, accurate urban extraction using a Multispectral LiDAR (MS-LiDAR) point cloud is vital. Feature extraction from MS-LiDAR point cloud has evolved by dint of deep learning scientific progress. Automated building extraction from MS-LiDAR point clouds without converting point cloud data to a raster is the main idea in this paper. This investigation used Teledyne Optech Titan multi-spectral LAS point clouds. A statistical outlier removal (SOR) filter removed noise. LAS Teledyne Optech Titan multispectral ALS point clouds are used in the study. A representative example of a complex urban environment from the National Center for Airborne Laser Mapping (NCALM) of a residential area near the University of Houston. A statistical outlier removal (SOR) filter removed noise. Then point clouds from 532, 1024, and 1550 nm were merged. Via the cloth simulation filter (CSF) filtering algorithm, ground and above-ground points were separated. The Point CNN algorithm was used in this study to extract buildings from multispectral LiDAR point clouds directly. This was studied in nine different scenarios. A new index-based methodology, normalized difference feature Index (NDFI), is proposed to enhance the building extraction accuracy from multispectral LiDAR point cloud by using green and mid-infrared (MIR) bands. The deep learning building extraction was applied based on the proposed new urban index and point convolutional neural network (Point-CNN) algorithm. In addition to its conceptual ease and low processing cost, the suggested index is efficient at extracting buildings. In Deep Network, introducing a new urban index NDFI yielded better accurate findings (OA of 95.27%) than NDVI’s 95.11%. Finally, this strategy was tested with real-world data, proving its viability. Overall results show that the point-CNN algorithm combined with the new urban index is both accurate and well-suited for building extraction.