3D-SegNet: A deep learning framework for three-dimensional airborne laser scanning point cloud segmentation for building identification

Abstract

Abstract The airborne laser scanning (ALS), a state-of-the-art 3D mapping technique is used for the fast and comprehensive three-dimensional (3D) data acquisition of urban environment. In this paper, a 3D-SegNet method is presented for identification of buildings using 3D ALS point cloud data. This method is mainly divided into two main steps: data preprocessing, and SegNet convolutional neural network: Urban building segmentation. In data preprocessing, the various LiDAR and geometric features are generated using point-wise 3D analysis in local spherical neighborhood. These features are processed and rasterized into feature images. Feature images along with buildings masks are used for the proposed 3D-SegNet model training and testing. The proposed 3D-SegNet model is straightforward to implement, where accurate segmentation of buildings are effectively dealt in several complex cases, such as buildings with varying dimensions, incomplete building geometry and data gaps; overlapped and connected objects with one of the objects as building, etc. The 3D-SegNet method performance for buildings segmentation was reported as average IOU, accuracy and F1-score of 76.19%, 91.19% and 77.45%, respectively employing the method on two datasets having different scene complexity. The proposed method is straightforward to implement and can be used as standard tool in urban planning strategies formation.