Performance of unsupervised machine learning methods using chi-squared weights for LiDAR point cloud filtering in urban areas

Abstract

ABSTRACT In this study, we compared the LiDAR filtering performances of unsupervised machine learning methods, such as linkage, K-means, and self-organizing maps, for urban areas to provide a practical guide to researchers. The input parameters (x-y-z and intensity) were normalized and weighted using a chi-squared independence test to improve the classification accuracy. The best successful results were obtained using the weighted linkage method in terms of the total error of 13.53%, 3.96%, and 1.07% for the three samples, respectively. In comparison with other approaches, methods weighted by chi-squared have significant potential for classification and filtering and outperform many popular approaches.