Abstract
This paper presents a hierarchical recovery method for generating Digital Terrain Models (DTMs) from airborne LiDAR point clouds using a novel layering approach. The method registers and layers last-return points by height, creating a reference system for identifying terrain. A series of rasterized pyramid levels are generated to iteratively identify terrain points from coarse to fine scales. Cells identified as off-terrain are interpolated, and the results are further refined using first-return feedback and Inverse Distance Weighted (IDW) interpolation. This data-driven approach requires no prior knowledge of scene complexity and demonstrates strong performance in flat or gently sloping terrains.
Introduction: The Filtering Dilemma
Generating an accurate Digital Terrain Model (DTM)—a representation of the bare earth—from LiDAR data remains a significant challenge in remote sensing. While laser pulses can penetrate canopy gaps, the resulting point cloud is often a complex mixture of ground, vegetation, buildings, and noise.
Existing automated filters often face a "filtering dilemma": small search windows may mistake large flat roofs for ground, while large windows may smooth out critical small-scale features like ridges or ditches. Additionally, "low outliers" (erroneous points below the ground) and complex man-made structures like bridge ramps frequently cause classification errors. This research proposes a robust, multi-scale method to address these complexities across various landscapes.
Methodology
The proposed Multi-scale Terrain Filtering (MTF) method is structured into three primary stages designed to "peel back" non-ground features.
1. Preprocessing and Layering
The algorithm focuses on "last-return" points, which are most likely to represent the ground. It generates a height histogram of the dataset and divides points into vertical layers. This global analysis allows the system to identify and remove low-elevation outliers that typically plague local filtering algorithms.
2. Iterative Multi-scale Filtering
The system constructs a pyramid of rasterized levels. Starting at the coarsest (top) level, the lowest points in large grid cells are assumed to be terrain. As the algorithm moves down to finer scales, it compares a point's layer number and local slope against the previous level's results. Points flagged as "off-terrain" (such as buildings or trees) have their heights replaced by values interpolated from neighboring terrain points.
3. DTM Refinement
In the final stage, the method incorporates "first-return" data to calculate a normalized Digital Surface Model (nDSM). This feedback loop fine-tunes the classification before a final Inverse Distance Weighted (IDW) interpolation produces the high-resolution DTM.
Experimental Results
The MTF method was evaluated using the ISPRS WG III/3 LiDAR dataset, covering 15 diverse sample sites in Vaihingen and Stuttgart, Germany. These sites included challenging features such as steep hillsides, dense vegetation, large irregular buildings, and bridges.
Key Findings
- Accuracy: The method achieved an average accuracy rate of 85%, with peak performance reaching 96% in favorable conditions.
- Statistical Reliability: The average Kappa coefficient was 61.2%, indicating substantial agreement beyond chance.
- Comparative Performance: In a head-to-head comparison with eight established algorithms, the MTF method ranked 6th. While it outperformed several well-known methods, it lagged behind top-tier adaptive TIN models.
- Terrain Sensitivity: The algorithm proved exceptionally reliable in flat urban areas and gentle slopes. However, it struggled in areas with very steep slopes and discontinuous terrain, such as riverbanks.
Conclusion and Future Work
The iterative layering approach provides a powerful, data-driven framework for DTM generation that is resilient to low-elevation noise and large man-made structures. Its primary limitation remains the "Steep Slope Problem," where rapid elevation changes can trigger slope-based thresholds incorrectly, leading to higher Type II errors (misclassifying objects as ground).
Future research will focus on refining interpolation logic for steep terrain and reducing the manual parameter tuning required to adapt the algorithm to different environmental complexities.
Summary updated