3D Edge Convolution in Deep Neural Network Implementation for Land Cover Semantic Segmentation of Airborne LiDAR Data

Abstract

3-dimensional data contains more informative visualization than a 2-dimensional one. LiDAR sensor produces 3D data or point cloud data. There have been many implementations of LiDAR data such as for building detection, urban area modeling, and land cover analysis. This study will analyze land cover because of its substantial benefits. The purpose of this study is to produce semantic segmentation of land cover from LiDAR data by implementing EdgeConv Algorithm from Dynamic Graph Convolutional Neural Network (DGCNN). The dataset in this study is LiDAR data of Kupang, one of the areas in Indonesia. This work achieves the average accuracy of 67.76% for DGCNN better than the state-of-the-art method PointNet (previous method) with 64.97% by implementing the point cloud dataset from LiDAR data of Kupang. This is because the edge convolution could recognize the global shape structure and local neighborhood information so that it could improve the segmentation performance result.