Fine classification of near-ground point cloud based on terrestrial laser scanning and detection of forest fallen wood

Abstract

Terrestrial Laser Scanning (TLS) can effectively describe complex forest scenes. This study aimed to classify ground point cloud within the height of 1.3 m into ground, vegetation, fallen wood, and standing trunk based on the TLS obtained from fallen wood plots in Daxing'anling. Fallen wood cloud point was segmented and merged. The optimal 3D neighborhood of each individual point was calculated through the Shannon entropy constructed by linearity, planarity, and scattering to avoid the difference in cloud density and the morphological characteristics introduced by occlusion. Shannon entropy could be maximized across the increasing kNN with an interval of 5 points. The optimal neighborhood size was used to compute the covariance eigenvalues for constructing 3D and 2D features. Key features were selected following the recursive feature elimination criteria, and a random forest classification algorithm was used to classify the points. Noise removal approach was applied to the fallen wood points classified by self-adjusting kNN features, and random sample consensus (RANSAC) segmentation was implemented to segment cylinders. Fallen wood cylinders were selected and merged depending on the axis direction less than 12° and the distance less than 0.1 m between each other. The overall classification accuracies of self-adjusting kNN method in plots A, B, and C were 93.17%, 94.52%, and 95.16%, respectively, and corresponding Kappa coefficients were 0.8771, 0.9145, and 0.9242, respectively. The overall accuracies of non-self-adjusting kNN were 92.65%, 89.09%, and 92.99%, and the Kappa coefficients were 0.8684, 0.8909, and 0.9299. Point cloud of plots B and C was classified using the model we trained using plot A. The classification accuracies of plots B and C were 62.38% and 59.80%, and the user precisions of fallen wood point cloud were 79.31% and 48.06%. All fallen woods had the same number as the ground measurement, and the parameters of fallen wood could be estimated roughly. Compared with the non-self-adjusting kNN method, the near-ground point cloud classification accuracy was improved by the self-adjusting kNN point cloud feature. Classification of plots B and C using the training result of plot A suggested that the selected key features in the complex forest could explain the dependent variable well. RANSAC could effectively segment the cylinder and estimate the parameters of the fallen wood. This research is significant for extracting parameters of the existing work. Further ecological research will be considered accordingly.