Deep learning for filtering the ground from ALS point clouds: A dataset, evaluations and issues

Abstract

The capability of partially penetrating vegetation canopy and efficiently collecting high-precision point clouds over large areas makes airborne laser scanning (ALS) a valuable tool for various geospatial applications. However, automated ground filtering (GF), one fundamental and challenging step for most ALS applications, has remained a widely researched yet unsolved problem for decades. The recent breakthroughs in supervised deep learning (DL) techniques, which rely on sufficient and high-quality labeled datasets, provide a new solution to better solve this problem. Unfortunately, public 3D geospatial datasets are scarce, especially for those tailored for the landform-scale GF task. Moreover, whether advanced deep neural networks (DNNs) can be well-scaled to the problem of GF remains an open question. To comprehensively advance the development of effective DL-based GF pipelines, we first publish an ultra-large-scale GF dataset built upon open-access ALS point clouds of four different countries worldwide, which covers over 47 km2 and nine different terrain scenes. Then, multiple attractive advantages of DL techniques in GF are evaluated through extensive experimental comparisons with traditional GF methods on the presented dataset. Furthermore, we reveal several issues faced by generalizing existing advanced 3D DNNs into GF tasks with a series of in-depth experimental analyses. Finally, some promising directions for future research are suggested in response to the identified challenges. Our dataset, named OpenGF, is available at https://github.com/Nathan-UW/OpenGF.