How is ground classified in a point cloud? From noise removal to a bare-earth model
Only some of the millions of points in a point cloud actually touch the ground.
Field takeaways
- Contours, sections and volumes are made from bare earth, which is why the ground class comes first.
- Noise is removed before the ground filter; a single false low point opens a pit in the surface.
- Filter parameters are chosen for the terrain; settings that work in a city fail in the mountains.
- Bridges, slopes, dense scrub and building edges are checked by hand, on cross-sections.
- If there are no points under the trees, no filter can create ground there; that area is surveyed in the field.
Why the ground class comes first
A point cloud holds everything the sensor saw: ground, treetops, roofs, vehicles, wires. Türkiye's Large-Scale Map and Map Information Production Regulation draws the line with two terms: a digital surface model represents the surface including vegetation and structures, while a digital terrain model represents the bare earth with those removed. Contours, sections, cut-and-fill volumes and drainage are computed from bare earth. In the ASPRS LAS standard, ground points are class 2, and the terrain model is built from that class. In an unclassified cloud a treetop behaves just like a hilltop.
Remove the noise first
Faulty points are removed before the ground filter runs. High noise (birds, dust, reflections) usually stands out and does little harm. The real danger is low points that fall below the ground: multipath reflections or measurement errors can produce points metres below the surface. Because most ground filters start from the lowest points, a single false low point opens a pit in the surface and makes the filter reject the real ground points around it. LAS 1.4 has class 7 for low noise and class 18 for high noise; moving noise into those classes is more traceable than deleting it.
How automatic filters work
Common approaches fall into three families. Morphological filters compare each point with the lowest surface in a moving window and enlarge the window step by step. Progressive TIN densification is the method Axelsson described in 2000: it builds a triangle network from the lowest point in each grid cell, then adds points to the ground whose distance and angle to a triangle stay within thresholds. Cloth simulation turns the cloud upside down and drops a virtual cloth onto it; the points the cloth touches count as ground. All of them have parameters: the cell size is usually chosen from the largest building in the area, and slope and distance thresholds are set for the terrain. Settings that work in a city fail in the mountains.
Where the automatic result goes wrong
On open, flat ground the automatic result is often good enough. Errors cluster in particular places: at bridges and overpasses the deck is taken for ground or the road beneath disappears; on steep slopes and road cuts the edge of the slope is clipped and rounded off; in dense low vegetation and crops the top of the plants is taken for ground; walls, retaining structures and parked vehicles get mixed into the ground; water surfaces return very few points. The best way to find these places is to look for bumps and pits in a hillshade view and to cut cross-sections through suspicious spots. Correcting means changing the class of points by hand on the section.
LiDAR versus photogrammetry
LiDAR pulses can pass through gaps in the leaves; one pulse can return several echoes, and the last returns often reach the ground beneath the trees. A photogrammetric point cloud records only the surface visible in the photos. If there are no ground points under dense forest or scrub, no filter can create them; classification only separates the points that exist. In such areas the ground must be surveyed in the field with GNSS or a total station, and those measurements added to the terrain model.
The terrain model and checking it
The terrain model is built from the ground points as a triangle network (TIN) or a regular grid; the grid spacing should not be finer than the point density supports. Contours and volumes come from this model. Before delivery, test the model against independent measurements: survey ground points with GNSS or a total station in different cover types, such as open ground, grass and under trees, and report the height differences separately for each cover type. In Türkiye, the same regulation also checks the accuracy of contours on maps against sections measured in the field, with points along each section no more than 20 m apart.
Open the point cloud, pull out the ground.
MapLab Survey 3D views LAS and LAZ files in 3D, takes sections and filters by class, suggests a ground classification you can correct by hand, and produces a terrain surface, cut-and-fill volume and contours as DXF. Windows, on request.
Frequently asked questions
Can ground classification be fully automatic?
On open, flat ground the automatic result is often good. Bridges, slopes, dense scrub and building edges need manual checking and correction.
Why are there no ground points under the trees?
A photogrammetric cloud records only the visible surface. LiDAR often reaches the ground through returns that pass between the leaves; if neither has points there, the ground must be surveyed in the field.
Does a bridge count as ground?
In a terrain model the bridge deck is usually not ground; LAS 1.4 has a separate class (17) for bridge decks. Which rule applies should be written in the delivery specification.
What does MapLab Survey 3D do in ground classification?
It opens LAS and LAZ files in 3D, takes sections and filters by class; it suggests an automatic ground classification and lets you correct it by hand. From the ground it produces a terrain surface, cut-and-fill volume and contours as DXF. It runs on Windows and is available on request.
Technical references
The field guidance in this article is aligned with the technical and official references below.
- ASPRS LAS Specification
- Axelsson (2000), DEM generation from laser scanner data using adaptive TIN models, ISPRS Archives XXXIII-B4
- Zhang et al. (2016), An Easy-to-Use Airborne LiDAR Data Filtering Method Based on Cloth Simulation, Remote Sensing 8(6):501
- Harita Genel Müdürlüğü, Büyük Ölçekli Harita ve Harita Bilgileri Üretim Yönetmeliği
Related: LAS and LAZ files · Orthophotos and 3D models · Cut-and-fill calculation · MapLab Survey