LiDAR Ground Classification Simulator: Terrain Filtering Algorithms

simulator intermediate ~10 min
Loading simulation...
Ground: 42% — moderate canopy coverage area

With slope threshold 15°, height difference 1 m, and cell size 10 m on moderate terrain, approximately 42% of points are classified as ground. Commission error is 3% and omission error is 5%, suitable for DEM generation at 1 m resolution.

Formula

slope = arctan(Δz / Δxy) (point-to-point slope)
Δh_threshold = s_max × d + h_offset (adaptive height filter)
DEM_resolution = √(area / n_ground) (minimum DEM cell size)

Separating Ground from Above

A raw LiDAR point cloud contains returns from every surface the laser pulse encounters: treetops, building roofs, power lines, vehicles, and the ground itself. Separating ground from non-ground is the critical first step in producing bare-earth Digital Elevation Models used for flood mapping, infrastructure planning, and archaeological prospection. The challenge is that no single rule perfectly distinguishes a steep hillside from a building wall, or a boulder from a car — making this one of the most studied problems in geospatial science.

Progressive Morphological Filtering

The most widely used approach builds the ground surface iteratively from coarse to fine. Starting with large grid cells, the algorithm identifies the lowest point in each cell as a ground seed. It then progressively decreases cell size, adding points that satisfy slope and height-difference criteria relative to the existing ground surface. Points too far above the interpolated surface are classified as non-ground. This simulation visualizes the classification process and shows how parameter choices affect the resulting ground surface.

Error Trade-offs

Every classification algorithm faces a fundamental trade-off between commission errors (falsely including non-ground objects) and omission errors (falsely rejecting ground points). Relaxing thresholds captures more legitimate ground but also admits low vegetation and terrain-following structures. Tightening thresholds cleans the ground surface but erodes edges, clips hilltops, and removes rough terrain features. The simulation lets you explore this trade-off interactively by adjusting slope, height, and cell-size parameters.

Adaptive Classification

Modern production workflows use spatially varying parameters: tight thresholds in flat urban areas (where buildings must be excluded) and relaxed thresholds in mountainous terrain (where steep slopes are legitimate ground). Machine learning approaches trained on manually classified reference datasets now augment or replace rule-based filtering in complex environments. The goal remains the same: extract an accurate bare-earth surface that faithfully represents the terrain beneath all surface cover.

FAQ

What is ground filtering in LiDAR?

Ground filtering (or ground classification) is the process of separating LiDAR point cloud returns into ground and non-ground categories. Ground points are used to generate bare-earth Digital Elevation Models (DEMs), while non-ground points represent vegetation, buildings, and other above-ground objects. Accurate filtering is essential for flood modeling, cut/fill calculations, and archaeological detection.

How does progressive morphological filtering work?

Progressive morphological filtering starts with a coarse grid (large cell size) and identifies the lowest points as ground seeds. It then progressively increases resolution, adding points that meet slope and height criteria relative to already-classified ground points. Each iteration refines the ground surface at finer scales, balancing ground detection with noise rejection.

What causes classification errors?

Commission errors (false ground) arise from low vegetation, terrain-following features like walls, and steep slopes below the threshold. Omission errors (missed ground) occur when legitimate ground in rough terrain exceeds the height threshold, or when dense canopy prevents last returns from reaching the surface. Parameter tuning must balance these competing errors.

What terrain types are hardest to classify?

Steep mountainous terrain, dense tropical forest with minimal ground returns, urban areas with complex building footprints, and coastal cliffs with overhanging features are the most challenging. Each requires different parameter settings, and many operational workflows use terrain-adaptive parameter sets that vary spatially across the survey area.

Sources

Embed

<iframe src="https://homo-deus.com/lab/lidar-remote-sensing/terrain-classification/embed" width="100%" height="400" frameborder="0"></iframe>
View source on GitHub