Charles Explorer logo
🇬🇧

Semantic classification of sandstone landscape point cloud based on neighbourhood features

Publication at Faculty of Science |
2020

Abstract

The technology of airborne laser scanning enables fast and accurate gathering spatial data containing also echoes from the terrain below the vegetation canopy that is beneficial for topographic mapping of wooded sandstone landscapes in Czechia, Poland, and Germany. The challengeable task is to determine the ground points in the point cloud because commonly used filtration methods do not successfully distinguish between vegetation and rock pillars and faces.

In this paper, we replace filtration with classification approach using the features derived from characteristics of points within a neighbourhood of optimized sizes, such as eigenvalue-based features and echo ratio. Random Forest classifier is trained and tested on the manually labelled dataset with a density of almost 650 points/m2 from the Adršpach-Teplice Rocks.

The overall accuracy reaches 87% but recall and precision of non-ground points are unsatisfactory. Misclassified non-ground points are located also within trees, thus we do not consider the result as suitable for DTM processing without further processing.