Learning Digital Terrain Models From Point Clouds: ALS2DTM Dataset and Rasterization-Based GAN

Abstract

Despite the popularity of deep neural networks in various domains, the extraction of digital terrain models (DTMs) from airborne laser scanning (ALS) point clouds is still challenging. This might be due to the lack of dedicated large-scale annotated dataset and the data-structure discrepancy between point clouds and DTMs. To promote data-driven DTM extraction, this paper collects from open sources a large-scale dataset of ALS point clouds and corresponding DTMs with various urban, forested, and mountainous scenes. A baseline method is proposed as the first attempt to train a Deep neural network to extract digital Terrain models directly from ALS point clouds via Rasterization techniques, coined DeepTerRa. Extensive studies with well-established methods are performed to benchmark the dataset and analyze the challenges in learning to extract DTM from point clouds. The experimental results show the interest of the agnostic data-driven approach, with sub-metric error level compared to methods designed for DTM extraction. The data and source code is provided at https://lhoangan.github.io/deepterra/ for reproducibility and further similar research.

Generating DTM from DSM Using a Conditional GAN in Coastal Urban Areas of Japan

“…The present RMSE is smaller than that reported in similar previous studies, although simple comparisons cannot be made because the error is expected to strongly depend on the target area characteristics (e.g., topography and coverage of buildings and trees) [9]. In the previous study that used cGAN to translate DSM into DTM [12], RMSE was 1.6 m for the area with an elevation histogram similar to that for Hyogo in the present study. The spatial resolution of their DSM was also similar to ours.…”

“…Although RGB images, DSMs, and classification results for ground and nonground objects were used as input, the RMSE was 1.5-2 m. They concluded that the model would be trained to output the average elevation in the dataset. In recent years, an algorithm combining U-Net structure networks and dedicated pre-and post-data processing [9], as well as cGAN [12], have offered good results even with DSM-only input. The present study is consistent with those studies.…”

“…Yashima et al [10] used the pix2pix model [11], which performs image conversion using conditional adversarial generation networks (cGANs); however, the data used and the investigation of results were very limited. Lê et al [12] also developed a method based on cGAN. Although errors in their results were not small, they showed that the GAN feature (i.e., networks of generators and discriminators are trained competitively) contributed to improving accuracy.…”

A Discordance Analysis in Manual Labelling of Urban Mobile Laser Scanning Data Used for Deep Learning Based Semantic Segmentation

“…Labelled point clouds are widely applied to train and test DL methods, both in indoor and outdoor environments. Some researchers process point clouds through a rasterization (Guiotte et al, 2020;Lê et al, 2022;Paz Mouriño et al, 2021) and voxelization (Tchapmi et al, 2017;Xu et al, 2021). In this way, advantages from well-studied Convolutional Neural Networks (CNNs) can be applied to point cloud structures.…”

The assessment of elevation data consistency. A case study using the ALS and georeference database in the City of Kraków

“…Their findings shed light on the performance of these algorithms and their varying accuracy levels, influenced by factors, such as model resolution, ground slope, and point cloud density. In the domain of deep learning, Lê et al (2022) addressed the extraction of DTMs from ALS point clouds using a deep neural network approach named DeepTerRa. They collected a large-scale dataset of ALS point clouds and corresponding DTMs, training a deep neural network for direct DTM extraction through rasterisation techniques.…”