AN EFFICIENT DEEP LEARNING APPROACH FOR GROUND POINT FILTERING IN AERIAL LASER SCANNING POINT CLOUDS

Abdul Nurunnabi, Felix Norman Teferle, Jonathan Li, Roderik Lindenbergh, Addisu Hunegnaw2021 · DOI

Abstract

Ground surface extraction is one of the classic tasks in airborne laser scanning (ALS) point cloud processing that is used for three-dimensional (3D) city modelling, infrastructure health monitoring, and disaster management. Many methods have been developed over the last three decades. Recently, Deep Learning (DL) has become the most dominant technique for 3D point cloud classification. DL methods used for classification can be categorized into end-to-end and non end-to-end approaches. One of the main challenges of using supervised DL approaches is getting a sufficient amount of training data. The main advantage of using a supervised non end-to-end approach is that it requires less training data. This paper introduces a novel local feature-based non end-to-end DL algorithm that generates a binary classifier for ground point filtering. It studies feature relevance, and investigates three models that are different combinations of features. This method is free from the limitations of point clouds’ irregular data structure and varying data density, which is the biggest challenge for using the elegant convolutional neural network. The new algorithm does not require transforming data into regular 3D voxel grids or any rasterization. The performance of the new method has been demonstrated through two ALS datasets covering urban environments. The method successfully labels ground and non-ground points in the presence of steep slopes and height discontinuity in the terrain. Experiments in this paper show that the algorithm achieves around 97% in both F1-score and model accuracy for ground point labelling.

Introduction

Generating accurate Digital Terrain Models (DTMs) requires "ground point filtering"—the process of separating points that hit the actual earth from those that hit "off-ground" objects like trees, buildings, and cars. While Deep Learning (DL) has become the gold standard for this task, it faces major hurdles in the context of point clouds. Point clouds are inherently unstructured and vary in density, making them difficult to process with standard Convolutional Neural Networks (CNNs) designed for regular grids. Furthermore, popular "end-to-end" models require massive amounts of labeled training data to learn spatial patterns from scratch. This research proposes a feature-based approach that works directly on raw points without needing to convert them into 3D voxels or 2D rasters, significantly reducing data requirements.

Methodology

The researchers proposed a three-step workflow centered on a Fully Connected Neural Network (FCNN).

Feature Engineering

Instead of feeding raw $(x, y, z)$ coordinates into the network, 28 local features were calculated for every point. These features are categorized into Covariance Features (3D), Z-Features (2D), and LiDAR Features (1D).

Table of local geometric features used for point cloud classification.

Model Selection and Architecture

Three different combinations of these features were tested:

  • Model 1: All 28 features.
  • Model 2: A subset used in previous multi-scale studies.
  • Model 3: A refined set of 19 features emphasizing shape characteristics (like planarity and eigentropy) over raw eigenvalues.

The architecture consists of a deep network with 5 hidden layers (50 neurons each) using ReLU activation, Batch Normalization, and a Sigmoid output layer for binary classification (Ground vs. Non-Ground).

Experimental Setup

The model was evaluated on two distinct urban ALS datasets:

  • AHN Data (Netherlands): An open-source dataset (~20 pts/m²) featuring complex urban objects like residential-commercial buildings and large churches.
  • ACT Data (Luxembourg): A 2019 LiDAR survey (~15 pts/m²) used to test robustness against height discontinuities and steep terrain.

Results and Discussion

The proposed method demonstrated high precision and reliability across both test sites, achieving approximately 97.3% accuracy on the AHN dataset and 98.6% on the ACT dataset. F1-scores consistently reached around 97% for both ground and non-ground labeling.

Handling Complex Terrain

A key strength of the algorithm is its ability to label points in areas that traditionally challenge filters, such as steep slopes and sudden height discontinuities.

Demonstration of ground point filtering in the presence of a steep slope.

Efficiency via Feature Selection

Interestingly, Model 3 (19 features) outperformed Model 1 (28 features), suggesting that redundant features can introduce noise. By pre-calculating geometric features, the neural network does not have to "learn" geometry from scratch, making the training process faster and more effective with smaller training sets compared to end-to-end architectures.

Conclusion and Future Work

While the binary classification is highly successful, the approach requires domain knowledge to manually select and calculate features. Future research will focus on multi-class classification to identify specific object types (vegetation, buildings) and automating the selection of the neighborhood radius for feature calculation.

Summary updated

Abstract

Ground surface extraction is one of the classic tasks in airborne laser scanning (ALS) point cloud processing that is used for three-dimensional (3D) city modelling, infrastructure health monitoring, and disaster management. Many methods have been developed over the last three decades. Recently, Deep Learning (DL) has become the most dominant technique for 3D point cloud classification. DL methods used for classification can be categorized into end-to-end and non end-to-end approaches. One of the main challenges of using supervised DL approaches is getting a sufficient amount of training data. The main advantage of using a supervised non end-to-end approach is that it requires less training data. This paper introduces a novel local feature-based non end-to-end DL algorithm that generates a binary classifier for ground point filtering. It studies feature relevance, and investigates three models that are different combinations of features. This method is free from the limitations of point clouds’ irregular data structure and varying data density, which is the biggest challenge for using the elegant convolutional neural network. The new algorithm does not require transforming data into regular 3D voxel grids or any rasterization. The performance of the new method has been demonstrated through two ALS datasets covering urban environments. The method successfully labels ground and non-ground points in the presence of steep slopes and height discontinuity in the terrain. Experiments in this paper show that the algorithm achieves around 97% in both F1-score and model accuracy for ground point labelling.
BibTeX
@inproceedings{nurunnabi2021efficient,
  author = {Abdul Nurunnabi and Felix Norman Teferle and Jonathan Li and Roderik Lindenbergh and Addisu Hunegnaw},
  title = {{AN EFFICIENT DEEP LEARNING APPROACH FOR GROUND POINT FILTERING IN AERIAL LASER SCANNING POINT CLOUDS}},
  booktitle = {˜The},
  year = {2021},
  doi = {10.5194/isprs-archives-xliii-b1-2021-31-2021},
}
Figure