Point Cloud Classification and Accuracy Analysis Based on Feature Fusion

Abstract

A method for land-cover classification was proposed based on the fusion of features generated from waveform data and point cloud respectively. It aims to partially overcome the ineffectiveness of many traditional classifiers caused by the fact that point cloud is lacking spectral information. The whole flowchart of the method is as follows: Firstly, Gaussian decomposition was applied to fit an echo full-waveform. The parameters associated with the Gaussian function were optimized by LM (Levenberg-Marquard) algorithm. Six and thirteen features were generated to describe the waveform characteristics and the local geometry of point cloud, respectively. Secondly, a random forest was selected as the classifier to which the generated features were input. Relief-F was used to rank the weights of all the features generated. Finally, features were input to the classifier one by one according to the weights calculated from feature ranking, where classification accuracies were evaluated. The experimental results show that the effectiveness of the fusion of features generated from waveform and point cloud for LiDAR data classification, with 95.4% overall accuracy, 0.90 kappa coefficient, which outperform the results obtained by a single class of features, no matter whether they were generated from point cloud or waveform data.

Keywords

• |airborne lidar|full-waveform data|feature fusion|land-cover classification