Ground classification and DTM extraction from airborne LiDAR

In the ASPRS scheme, ground is class 2 — the bare earth, stripped of everything that sits on top of it. On screen it is the quietest layer in a classified block: no rooftops, no canopy, no wires, just the shape of the land. Yet almost every terrain product a production team delivers is built on top of it. The digital terrain model, the contours drawn from it, the slope rasters, the inputs to hydro-enforcement — all inherit whatever the ground class got right, and every error it carries.

This guide is written for teams that deliver terrain products. It covers why ground is the foundation class, how the terrain model relates to the surface and canopy models built from the same scan, the main families of ground-filtering methods and where each one tends to struggle, the terrain that reliably defeats a filter, and the review a surface goes through before it ships.

Ground is the class every terrain product inherits

A digital terrain model is nothing more than an interpolation of the ground class. Feed it a clean, complete set of ground points and the surface follows the real land; feed it ground that still holds a few shrubs, a low building corner, or a scatter of noise, and every one of those intrusions becomes a bump, a ridge, or a spike in the model. Because the DTM is the parent of so many other products, the error does not stay put. A residual bump in the ground surface pushes a contour line off its true position, tilts a slope value, and can send a modelled drainage path around an obstacle that is not really there.

The reverse failure is just as costly. Filtering too aggressively shaves real terrain away — the toe of a steep bank, the edge of a road cut — and the model sags where the land does not. Ground classification balances removing everything that is not terrain against keeping everything that is, and the whole downstream chain is decided at that balance point.

DTM, DSM, and CHM: three surfaces from one scan

The same LiDAR block produces several distinct surfaces, and mixing them up is a common source of confusion on delivery. Three come up constantly, and the difference between them is entirely about which returns they are built from.

The DSM shows the world as the sensor first saw it, treetops and rooftops included. The DTM shows the land as if everything on it were lifted away. The CHM — the canopy height model — is the difference between the two: a normalized height that measures how far each feature stands above the ground, which is why it depends directly on a correct terrain model underneath. A weak ground class does not only spoil the DTM; it quietly corrupts the canopy heights computed from it.

How ground filtering actually works

There is no single algorithm behind ground classification. Several method families have grown up over the years, each built on a different assumption about what separates terrain from everything else. Most production tools implement one or more of them, and knowing how each behaves helps a reviewer read the artifacts it leaves behind.

• Progressive TIN densification starts from a sparse set of low seed points, builds a triangulated surface, and adds a point to the ground whenever it falls within set angle and distance limits of the growing triangulation. It adapts well to rolling and varied terrain, but its thresholds are sensitive: too tight and it cuts into slopes, too loose and it climbs onto large flat roofs.
• Morphological filters treat the point cloud like an image and apply an opening operation with a moving window, removing anything that rises too sharply above its surroundings. They are simple and fast, but a single window size never fits both a small shrub and a large building, so most implementations step the window through several sizes — and steep terrain still tends to be shaved or stepped.
• Cloth simulation (CSF) inverts the point cloud and drapes a simulated fabric down onto it from above; points that end up close to the settled cloth are called ground. It is intuitive and needs few parameters, which makes it popular for gentle terrain, but cliffs, sharp ridges, and steep banks are exactly where a draped cloth cannot follow the real surface.
• Learned approaches train a model on already-classified data and let it predict ground directly from the point geometry. They can pick up patterns the rule-based methods miss, but they only generalize as far as their training data reaches: terrain unlike anything they were trained on is where they slip, and their output still has to be reviewed like any other.

No method is right in every setting. A filter is chosen for the terrain and land cover of the job, tuned on representative ground, and judged by what it produces rather than by its name.

The terrain that breaks a ground filter

Most of a block usually classifies cleanly. The review time goes to a short list of situations where the assumptions behind every filter start to fray, and where a reviewer expects to find work to do.

• Steep slopes and abrupt breaks. Cliffs, road cuts, and quarry walls climb faster than a slope-based filter expects, so terrain gets shaved off as if it were an object, or the filter loosens and lets objects through. Steep ground is where over- and under-filtering meet.
• Dense canopy with few ground returns. Under closed forest, only a small fraction of pulses reach the ground and return. The DTM has to be interpolated across wide gaps, and the temptation to pull low vegetation into ground to fill them is strong — which raises the surface instead of describing it.
• Buildings on slopes. A large flat roof on sloping land can look, to a filter, exactly like a terrace of real ground. Big industrial rooftops are a classic source of ground that is really a building.
• Low vegetation hugging the terrain. Tall grass, shrubs, and crops sit only centimetres above the soil, too close for a height threshold to separate cleanly, and they lift the surface into a soft, spongy blanket if they are not caught.
• Bridges and causeways. A bridge deck is not ground — it spans over the terrain — so it is held as class 17 and kept out of the surface, or the DTM bulges up to the deck and blocks the channel beneath. A causeway built as a solid earth embankment, by contrast, is genuine terrain and stays in ground. Telling the two apart is a judgement the filter cannot make on its own.

How reviewers check a ground surface

A ground class is not accepted because a filter finished running. It is accepted because someone looked at the surface it produced and could not find fault with it. Reviewers lean on a small set of views that make filtering errors visible, because the errors are almost impossible to see in a raw point cloud.

• Hillshade the DTM. A shaded-relief render of the terrain surface exposes what a point view hides: pits where a spurious low point punched a hole, bumps where vegetation or a building corner survived, and terracing or stepping left by window-based filters. Most first-pass problems show up here.
• Cut cross-sections through trouble spots. A thin profile through a bank, a bridge, or a patch of forest shows in one glance whether the ground points follow the true surface or drift up into canopy and structures. Sections are the fastest way to confirm a suspected error.
• Compare flightline overlaps. Where two strips cover the same ground, they should agree. A surface that shifts between overlapping lines points to a classification or adjustment problem that a single-strip view would never reveal.
• Spot-check against known surfaces. Survey checkpoints, flat road surfaces, and other places whose elevation is independently known give a direct read on whether the ground sits where it should.
• Look at the tile edges. Ground defined slightly differently on either side of a tile boundary leaves a seam in the merged surface. Consistent edges are what let neighbouring tiles form one continuous model.

Fitting ground extraction into a production workflow

At project scale, ground is extracted block by block. The area is split into tiles, each tile is processed with an overlap buffer so the filter has context beyond its own edges, and the tiles are merged back afterward. The buffer is what keeps the surface continuous across the seams the reviewer will later inspect.

It helps to keep two activities separate. Tuning parameters — choosing a filter and adjusting its thresholds — is done once on representative ground and then applied across the block. Reviewing outputs is a different job: reading the surface the settings produced and deciding whether it is good enough to ship. Collapsing the two, tweaking parameters tile by tile while reviewing, is how a project loses consistency, because the definition of ground quietly drifts from one tile to the next.

Whatever the tuning, the review step stays. No filter, rule-based or learned, produces a surface that can be shipped unseen, and the point of a good workflow is to make that review fast rather than to skip it. This is the shape of the ground work in Vecten Desktop, whose VGround module handles ground and terrain extraction and presents its results as reviewable outputs — hillshaded surfaces, sections, and editable classes — so the person signing off on the DTM is looking at the terrain, not just trusting that a process ran. The filter does the bulk of the work; the review is where the surface earns its place in a deliverable.