Last Return

One pulse, several echoes — and the decision about which one is the ground

When a lidar pulse enters a forest canopy, it does not arrive at the ground as a single event. The beam strikes leaves, then branches, then understorey shrubs, losing energy at each encounter and generating a partial return from each surface it intersects. A pulse fired into dense woodland may yield four or five discrete echoes before — or instead of — reaching bare soil. The question the classifier must then answer is: which of these is ground?

The convention is to treat the last detectable return as the most likely candidate for the ground surface. That intuition is mostly correct. The final echo in a pulse train is, by definition, the deepest one the system recorded — and if any energy reached the soil, it is that echo. But "last" and "ground" are not synonyms. Dense leaf litter, a mossy boulder, a fallen trunk lying flat across the slope: all can intercept a pulse before it reaches mineral earth, and all generate a last return that the classifier will provisionally call ground until something contradicts it.

Classifiers approach the problem statistically rather than pulse by pulse. The most widely used algorithm — progressively densified filtering, developed around work by Axelsson in the late 1990s and still at the core of most commercial and open-source pipelines — builds a thin triangulated mesh from the lowest points in each local neighbourhood, then tests adjacent returns for inclusion by progressively relaxing the height tolerance above that seed surface. Points that fall within tolerance are reclassified as ground; points above it are held as vegetation or structure. The method is iterative: each accepted point tightens or extends the mesh, and the threshold parameters — maximum angle, maximum distance — are tuned by the operator against the terrain type.

Flat ground with open canopy barely tests the classifier. Steep ground under closed evergreen canopy is where the method earns its difficulty. On a sharp slope, a return from a low branch can sit at roughly the same absolute elevation as a ground return from a few metres uphill, and the classifier may accept it. In dense conifer stands, where pulse penetration is thin and the last returns are sparse, the seed mesh itself may be poorly constrained. Artefacts — pits, raised patches — appear in the bare-earth model at the densest sections of canopy, and they are not always obvious in the final product.

The practical implication for anyone relying on a bare-earth model under closed canopy is to treat classifier output as a probabilistic surface, not a measurement. The point cloud is the measurement; the ground classification is an inference, and the quality of that inference varies with canopy density, flight parameters, and the operator's choice of threshold.