TL;DR
This paper introduces a novel recursive filtering algorithm for terrestrial laser scanning (TLS) data, specifically designed to accurately extract ground surface information in challenging steep-slope areas. By combining a slope-based seed point extraction with an adaptive Principal Component Analysis-Triangulated Irregular Network (PCA-TIN) approach, the method significantly improves Digital Terrain Model (DTM) generation accuracy in these complex terrains, outperforming existing algorithms.
What Problem Does This Paper Address?
Landslides pose a significant threat, causing substantial human and economic losses. Accurate Digital Terrain Models (DTMs) are crucial for understanding and predicting landslide risks, especially in mountainous and urbanized areas with steep slopes. Terrestrial Laser Scanning (TLS) offers high-resolution data ideal for capturing detailed terrain features. However, existing ground filtering algorithms, largely developed for Airborne Laser Scanning (ALS), struggle to accurately extract ground points from TLS data in steep-slope environments. This is due to challenges like abrupt slope changes, varying point densities, and the presence of dense vegetation, which can lead to the misclassification of non-ground points as ground, resulting in inaccurate DTMs.
Methods
The authors propose a new recursive filtering method that addresses the limitations of existing algorithms for steep-slope areas. The core of their approach involves a two-step process:
- Slope-Based Seed Point Extraction:
- Lowest Point Extraction: The point cloud is divided into a virtual grid, and the lowest point within each grid cell is identified as a potential seed point.
- Slope Angle Calculation: An initial DTM is generated from these seed points and interpolated. Slope angles are then calculated from this interpolated DTM.
- Axis Rotation for Steep Slopes: For cells exceeding a defined slope angle threshold (60° in their experiments), the algorithm rotates the axis based on a fitted plane to better capture ground points on abrupt slope changes. This process extracts additional seed points, particularly from the sides of steep features.
- Adaptive PCA-TIN:
- Initial TIN Generation: A Triangulated Irregular Network (TIN) is constructed using the extracted seed points. To preserve terrain information on steep slopes, Principal Component Analysis (PCA) is applied to rotate the point cloud to its principal axis before TIN generation. This helps to avoid losing crucial lateral view information.
- Iterative Point Addition: The algorithm iteratively adds candidate points to the TIN based on their vertical distance to the existing TIN and whether they fall within the TIN plane. Points below a threshold are classified as ground.
- TIN Reconstruction: The TIN is regenerated with the newly added ground points, and the PCA rotation is reapplied. This process repeats until no more candidate points are classified as ground, resulting in a final DTM.
Datasets
The study utilized terrestrial laser scanning (TLS) data acquired from two distinct steep-slope study areas in South Korea:
- Site A (Gwangju): Characterized by rock and minimal vegetation on the slope itself, with vegetation on the outskirts.
- Site B (Busan): Significantly covered with dense vegetation on the slope.
The data was acquired using a Riegl VZ-2000 scanner. Portions of the full datasets were extracted for analysis, resulting in point clouds of approximately 2.5 million points for Site A (covering 5252 m²) and 4 million points for Site B (covering 1378 m²). The paper also mentions using NAVER Map street-view images for visual comparison and Google Earth images for top-view context. For accuracy assessment, reference DTMs were generated by manually removing noise from the point clouds.
Key Insights and Ideas
- Steep slopes are a unique challenge: Standard ground filtering algorithms, often designed for ALS data and flatter terrains, fail to adequately capture the complex topography of steep slopes when applied to TLS data.
- Lateral view is critical: The oblique view of TLS data provides crucial lateral information on steep slopes that is often lost in 2.5D filtering methods. PCA rotation before TIN generation is key to preserving this information.
- Adaptive approach is essential: A fixed slope threshold is insufficient. The algorithm's ability to adapt by rotating the axis for steeper sections and iteratively refining the TIN is vital for accurate ground extraction.
- Vegetation is a persistent hurdle: Even with advanced filtering, dense vegetation remains a significant challenge, leading to commission errors (non-ground points classified as ground).
Key Results
The proposed algorithm demonstrated strong performance in steep-slope areas:
- Accuracy: The total accuracy values for the generated DTMs, measured by Root Mean Squared Error (RMSE), were 1.84 cm for Site A and 2.13 cm for Site B.
- Comparison: The proposed method outperformed other tested algorithms (ATIN, WLLI, CSF) in terms of RMSE and data missing ratio for Site A. For Site B, while WLLI showed a slightly lower RMSE, the proposed method achieved a significantly lower data missing ratio.
- Visual Inspection: Qualitative results showed that the algorithm effectively extracted ground points in steep areas, even where slopes changed abruptly, and represented significant terrain features well. However, some commission errors were observed in areas with low point density or thick vegetation.
Limitations and Future Work
- Limitations:
- Commission Errors: The algorithm exhibited commission errors, misclassifying non-ground points (vegetation) as ground, particularly in edge areas where gradients changed abruptly and in areas with low point density or thick vegetation.
- Memory and Time Efficiency: The authors acknowledge that future research should aim to improve memory consumption and time efficiency, which are common challenges with large point cloud datasets.
- Future Work:
- The authors suggest that future research should focus on improving the efficiency of the algorithm and further addressing the challenges posed by dense vegetation.
Summary updated