Different KNN parameters on road surface extraction from UAV-based point cloud under several machine learning algorithms

Abstract

Accurate road surface from a three-dimensional (3D) point cloud depends on various parameters. One crucial parameter is the set of point features. Point features enable classification by capturing characteristics of the surface on which the points are located. These features are calculated based on the closest points surrounding each point. In this study, the K-nearest neighbors algorithm (KNN) was applied to identify these closest points. The KNN algorithm requires only one input, the number of closest points (k). Eight different point features were developed using different k values, and their impact on road surface classification from the 3D point cloud was investigated. It was observed that there is no significant improvement in classification accuracy until a certain k value. However, better classification accuracy was achieved after a certain k value. The effect of different k values was also investigated under different training sample structures and machine learning (ML) algorithms. When training samples were selected from a single location as a large group, similar classification accuracy was obtained across different k values. Conversely, when training samples were chosen from various regions in smaller groups rather than a single large group, improved classification was observed as the k value increased. Additionally, it was noted that five different ML algorithms-random forest, support vector machine, generalized linear model, linear discriminant analysis, and robust linear discriminant analysis-have almost similar performance under different k values. Finally, using the optimum k value, improvements of up to 4.543% and 6.601% in accuracy and quality measures, respectively, were found.