Suggesting a Deep Learning-based Point Cloud Segmentation Framework to Classify Ground and Above-Ground Targets in a 3D Point Cloud acquired by a Full-waveform Terrestrial Laser Scanner over a Coastal wetland

Abstract:

Point cloud segmentation is the process of classifying points derived from the terrestrial or airborne laser scanners or other scanning devices. In addition to three dimensional coordinates, some terrestrial laser scanners provide additional attributes related to each individual target located in the path of the transmitted laser pulse. These attributes, such as echo index, amplitude of the echo, deviation factor of the echo and the point intensity can help to enhance the performance of underlying point cloud segmentation algorithms. However, due to the complex nonlinear correlation between those attributes, complexity of the scene, and high inter-class similarity and intra-class variability in back-scattered cross-section for multiple target categories, point cloud segmentation algorithms may not achieve the desired performance. In this study, a deep learning-based framework for point cloud segmentation is proposed. The framework combines raw waveform information provided by a full-waveform terrestrial laser scanner (TLS), for each target, with all other attributes related to the target to explore the best discriminative features in a high-dimensional feature space provided by a deep convolutional neural network. The study shows that the proposed framework can significantly increase the overall accuracy of point cloud segmentation in a coastal wetland to discriminate points from the bare ground from above ground targets such as points related to vegetated areas.