Evaluate the Impact of Class Granularity in Point Cloud Semantic Segmentation on DTM Accuracy

Abstract. The Digital Terrain Model (DTM) is considered an essential component in various applications, including road design, urban planning, terrain analysis, and, environmental monitoring. LiDAR data is known to have very high accuracy therefore it is considered the most reliable source for DTM generation. However, accurately filtering the LiDAR data for the ground classification remains a challenge. This study explores the impact of class granularity on semantic segmentation and its effect on the accuracy of DTM generation sourced to LiDAR data. The RandLA-Net has been used for semantic segmentation, and for the training, the DALAS dataset which comprises various terrains and structures is used. The process is comprised of training the deep learning models on datasets that are classified into two schemes. The first scheme is a course with 2-classes (ground and non-ground) and the second scheme is a finer one with 8-classes (ground, vegetation, cars, trucks, powerlines, fences, poles, and buildings). The trained models are applied to three datasets to evaluate the result of the granularity on the accuracy of the DTM. The results show that the accuracy of the DTM based on the 2-class model outperforms the accuracy of the DTM obtained via the 8-classes model, as indicated by the used statistical measures such as RMSE, mean and standard deviation (STD). The semantic segmentation of the 8-classes model shows more misclassification, especially in complex urban areas, especially in distinguishing the ground points from non-ground objects. The study emphasizes the trade-off between class granularity and DTM accuracy, which shows that simpler classification schemes will lead to a better result of the DTM generation results.