Urban scene classification using features extracted from photogrammetric point clouds acquired by uav

Abstract

Images and LiDAR point clouds are the two major data sources used by the photogrammetry and remote sensing community. Although different, the synergy between these two data sources has motivated exploration of the potential for combining data in various applications, especially for classification and extraction of information in urban environments. Despite the efforts of the scientific community, integrating LiDAR data and images remains a challenging task. For this reason, the development of Unmanned Aerial Vehicles (UAVs) along with the integration and synchronization of positioning receivers, inertial systems and off-the-shelf imaging sensors has enabled the exploitation of the high-density photogrammetric point cloud (PPC) as an alternative, obviating the need to integrate LiDAR and optical images. This study therefore aims to compare the results of PPC classification in urban scenes considering radiometric-only, geometric-only and combined radiometric and geometric data applied to the Random Forest algorithm. For this study the following classes were considered: buildings, asphalt, trees, grass, bare soil, sidewalks and power lines, which encompass the most common objects in urban scenes. The classification procedure was performed considering radiometric features (Green band, Red band, NIR band, NDVI and Saturation) and geometric features (Height-nDSM, Linearity, Planarity, Scatter, Anisotropy, Omnivariance and Eigenentropy). The quantitative analyses were performed by means of the classification error matrix using the following metrics: overall accuracy, recall and precision. The quantitative analyses present overall accuracy of 0.80, 0.74 and 0.98 for classification considering radiometric, geometric and both data combined, respectively.