Ground truth: benchmarking LiDAR classification algorithms at scale

Every digital terrain model starts with a decision most people never see: which LiDAR returns are ground and which are not. Get that wrong and every downstream product — contours, slope, viewsheds, flood models — inherits the error. So before committing a 500 km² survey to a single algorithm, we benchmarked the three most common ground filters against a large, hand-validated reference.

The reference dataset

We assembled 1.2 billion returns spanning three deliberately awkward terrain types: dense urban with flat rooftops that masquerade as ground, closed-canopy forest where few pulses reach the floor, and coastal dunes where the ground itself undulates sharply. A team manually classified representative tiles from each so we had a trustworthy answer to compare against.

The point of choosing hard terrain is that easy terrain hides differences. On a flat ploughed field every algorithm scores well; the separation only appears where the assumptions each method makes start to break.

The three contenders

• Progressive Morphological Filter (PMF): fast, but its window-size assumptions struggle with abrupt breaks like dune crests and building edges.
• Cloth Simulation Filter (CSF): intuitive and robust on moderate terrain, sensitive to its rigidness and cloth-resolution parameters.
• Adaptive TIN densification: accurate on complex ground, the most expensive of the three per point.

What the numbers said

No single algorithm won everywhere, which was the whole point. PMF was fastest and perfectly adequate on open, gently varying ground, but its recall collapsed on dune crests where it clipped real ground as non-ground. CSF was the best all-rounder and the most forgiving on canopy, provided its cloth resolution was tuned per terrain class. TIN densification produced the cleanest urban result, correctly rejecting rooftops, but at a runtime that made it unaffordable across the full extent.

The hybrid we shipped

Production runs a terrain-aware pipeline rather than one global filter. A coarse land-cover pass routes each tile to the algorithm that wins on that terrain: PMF for open rural where speed dominates, CSF as the tuned default, and TIN densification reserved for dense urban tiles where its accuracy justifies the cost. GPU-backed workers make the expensive path viable where it is actually needed.

The lesson generalises beyond LiDAR: at national scale the right answer is rarely a single algorithm applied uniformly. It is a router that spends compute where the terrain earns it.