Automatic Segmentation of Individual Grains From a Terrestrial Laser Scanning Point Cloud of a Mountain River Bed

In this article, we propose a method for instance segmentation of individual grains from a terrestrial laser scanning point cloud representing a mountain river bed. The method was designed as a classification followed by a segmentation approach. The binary classification into either points representing river bed or grains is performed using the random forest algorithm. The point cloud is classified based only on geometrical features calculated for a local, spherical neighborhood. A multisize neighborhood approach was used together with the feature selection method that is based on correlation analysis. The final classification was performed using a set of features calculated for the neighborhood size of 5, 15, and 20 cm. The achieved classification results have the overall accuracy of 85-95%, depending on the test site. The segmentation is performed using the density-based spatial clustering of applications with noise algorithm in order to cluster the point cloud based on Euclidean distances between points. The performed experiments showed that the proposed method enables us to correctly delineate 67-88% of grains, depending on the test site. However, the resulting point cloud based completeness expressed as Jaccard index is similar for each of the test sites and is approximately 88%. Moreover, the proposed method proved that it is robust to the shadowing effect.

Automated summary

This article proposes a method for instance segmentation of individual grains from a terrestrial laser scanning point cloud. The method includes a classification step using the random forest algorithm and a segmentation step using density and Euclidean distance. The experiments showed that the method accurately identifies grains and is robust to the shadowing effect.