RELATION NETWORK FOR FULL-WAVEFORMS LIDAR CLASSIFICATION

Abstract

LiDAR data are widely used in various domains related to geosciences (flow, erosion, rock deformations, etc.), computer graphics (3D reconstruction) or earth observation (detection of trees, roads, buildings, etc.). Because of the unstructured nature of remaining 3D points and because of the cost of acquisition, the LiDAR data processing is still challenging (few learning data, difficult spatial neighboring relationships, etc.). In practice, one can directly analyze the 3D points using feature extraction and then classify the points via machine learning techniques (Brodu, Lague, 2012, Niemeyer et al., 2014, Mallet et al., 2011). In addition, recent neural network developments have allowed precise point cloud segmentation, especially using the seminal pointnet network and its extensions (Qi et al., 2017a, Riegler et al., 2017). Other authors rather prefer to rasterize / voxelize the point cloud and use more conventional computers vision strategies to analyze structures (Lodha et al., 2006). In a recent work, we demonstrated that Digital Elevation Models (DEM) is reductive of the vertical component complexity describing objects in urban environments (Guiotte et al., 2020). These results highlighted the necessity to preserve the 3D structure of the point cloud as long as possible in the processing. In this paper, we therefore rely on ortho-waveforms to compute a land cover map. Ortho-waveforms are directly computed from the waveforms in a regular 3D grid. This method provides volumes somehow “similar” to hyperspectral data where each pixel is here associated with one ortho-waveform. Then, we exploit efficient neural networks adapted to the classification of hyperspectral data when few samples are available. Our results, obtained on the 2018 Data Fusion Contest dataset (DFC), demonstrate the efficiency of the approach.