Feature Selection for Airbone LiDAR Point Cloud Classification

Abstract

The classification of airborne LiDAR data is a prerequisite for many spatial data elaborations and analysis. In the domain of power supply networks, it is of utmost importance to be able to discern at least five classes for further processing—ground, buildings, vegetation, poles, and catenaries. This process is mainly performed manually by domain experts with the use of advanced point cloud manipulation software. The goal of this paper is to find a set of features which would divide space well enough to achieve accurate automatic classification on all relevant classes within the domain, thus reducing manual labor. To tackle this problem, we propose a single multi-class approach to classify all four basic classes (excluding ground) in a power supply domain with single pass-through, using one network. The proposed solution implements random forests and gradient boosting to create a feature-based per-point classifier which achieved an accuracy and F1 score of over 99% on all tested cases, with the maximum of 99.7% for accuracy and 99.5% for F1 score. Moreover, we achieved a maximum of 81.7% F1 score for the most sparse class. The results show that the proposed set of features for the LiDAR data cloud is effective in power supply line classification.