Tree segmentation methods

Segment Trees offers two ways to split a cloud into individual trees. They answer the same question from opposite ends of the tree, and which one works depends on which end the scan saw.

TreeIso — from the stems up

TreeIso (Xi & Hopkinson, 2022) runs in three stages:

1. 3D over-segmentation. The cloud is voxel-decimated and cut into many small clusters by cut-pursuit, a graph method that groups points that are close together.
2. 2D grouping. A second cut-pursuit, in plan view, assembles those clusters into candidate trees. The 2D reg. strength (λ₂) controls how large these groups grow.
3. Merging by similarity. Groups that look like fragments (no stem-like base, a small height span) are merged into the neighbor they most resemble.

It relies on seeing trunks and branches, which a terrestrial scan records well. An airborne scan of a closed canopy records almost none of either. Stage 2 is then left grouping crown fragments by how close they are in plan view, and touching crowns are merged. On a 40 × 40 m plantation plot with trees about 3.5 m apart, stage 2 reduced 4,706 well-formed stage-1 clusters to 16 groups for roughly 90 trees. Of seven settings tried, the best kept 60 of 91 trees separate, and only by breaking many others into pieces.

Canopy height — from the crowns down

The canopy height (CHM) method finds each tree where an airborne scan sees it best, at the top of its crown:

1. Height above ground. Every point's height above the terrain. The terrain comes from the cloud's labeled ground points when it has them. Otherwise it comes from its lowest returns on a 5 m grid, taking the lowest return within one cell in each direction, so a cell under an unbroken crown borrows a neighbor's ground return instead of reading the crown's underside as ground.
2. Canopy height model (CHM). A grid holding the highest point in each cell. The cell is one twelfth of Crown scale, fine enough to resolve the dip between two neighboring crowns. It follows the crown size, not the point density: a coarser cell for sparser data would merge exactly the trees a sparse scan most needs kept apart. Cells no return landed in are filled from their nearest neighbor when the hole is at most 1 m wide; wider holes are real gaps in the canopy. Small pits, where a pulse slipped through the canopy, are filled, and the grid is lightly smoothed (one cell).
3. Treetops. Every cell that is the highest in a small window around itself, and taller than Min tree height, is a treetop. The window is 0.55 × Crown scale across. Two neighboring tops are kept apart by the smoothing and the valley between them as much as by the window, so the window is smaller than a crown. A flat top counts as one.
4. Crowns. A watershed grows each crown outward and downward from its treetop until it meets a neighboring crown in the valley between them. A small distance penalty keeps a tall crown from spilling across a shallow valley into a shorter neighbor's top.
5. Points. Every point takes the ID of the crown above it, so trunks and understory under a crown join that tree. Points under no crown (lower than Min tree height, or in a gap) stay unassigned (0).

It needs no trunks and works on the whole cloud in one pass. The 8.4-million-point, 150 × 170 m plantation above takes about 7 seconds.

How the settings were calibrated

Everything scale-dependent follows Crown scale (S): treetop window 0.55 × S, cell S / 12, smoothing one cell. A tree counts as found when its predicted crown overlaps the true one by at least half (intersection over union, IoU ≥ 0.5). The rule and the 2.5 m default were checked against ground truth on one real stand and four synthetic ones, so that the default is not tuned to one site:

• A poplar plantation, airborne (8.4 M points, trunks 2.4–2.7 m apart), scored against six hand-labeled trees by per-tree intersection-over-union (IoU). S = 2.0–2.5 m kept all six separate at IoU 0.80–0.86, the ceiling set by one-cell crown boundaries. S = 2.75 m merged one pair and 3.0 m merged three. TreeIso on the same scan scored IoU 0.04: each labeled tree was a small part of a much larger instance.
• A synthetic plantation of touching crowns at 3.5 × 4 m, at 8–60 returns/m²: S = 3.0–3.5 m recovered all 20 trees; 2.5 m found 20–25 (a little over-split); 2.0 m about 44.
• A synthetic even-height stand, trees about 3.1 m apart, 8–12 m tall: S = 2.0–3.5 m found 99–100% of trees.
• A synthetic open natural stand, irregular positions about 4.4 m apart, 10–25 m tall: S = 2.5 m was best, 91% of trees. Values near the trunk spacing (4–5 m) dropped to 68–76%. This is why the setting is a crown scale and not a spacing.
• A synthetic dense, mixed-height natural stand, trees about 2.5 m apart, 6–20 m tall: the best value (2.0 m) found only 64%, and 2.5 m found 56%, even though 93% of the trees are visible from above. See the limit below.

A window that grows with local tree height, the usual airborne approach (Popescu & Wynne, 2004), was also tried. It gained at most 5 points on any stand, so it was not adopted.

The errors are lopsided. A too-small S splits crowns, which is easy to see. A too-large S merges trees, and it starts early: on the poplars, 2.75 m already merged a pair. Hence the advice to err low.

What it assumes

• One top per tree. A crown with several separate leaders (a multi-stemmed tree, a wide flat-topped crown) can be split into several trees. Raise Crown scale, or seed it once.
• Dense stands of mixed heights are its weak spot. Where short and tall trees are packed together, no single crown scale suits both: a small value splits the big crowns, and a large one swallows the small trees beside them. Expect roughly half to two thirds of trees right there; seed the ones that matter.
• Trees are not stacked. A tree growing under another's crown is part of the taller tree. It is invisible from above, and the CHM is built from above.
• Crown boundaries are one cell wide. Points are assigned by the grid cell they fall in, so a strip about one CHM cell wide along each boundary can land in the neighboring tree.

Seeds

A seed replaces the automatic treetops within half the Crown scale of it, and all other automatic treetops are kept. So you only need to seed the trees that came out wrong: two seeds on a merged pair split it, and one seed on a split crown joins it. A seed on a trunk rather than the exact top still works, as long as it is inside the crown. Seeded trees are numbered 1…n in the order the seeds were placed, and the automatic ones follow.

TreeIso seeds differ: each segment is given to its nearest seed, so a seed can join pieces but cannot split a segment that already holds several trees.

Which one?

	TreeIso	Canopy height
Needs	Visible stems and branches	Visible crown tops
Suits	Terrestrial / mobile scans; understory trees	Airborne / drone scans; closed canopies
Fails when	Crowns touch and stems are unseen — trees merge	Trees overtop each other; multi-leader crowns split
Speed	Tens of seconds to minutes; tiles large plots	Seconds, any size
Main knob	2D reg. strength (λ₂)	Crown scale

The multi-trunk warning ("instance(s) contain more than one trunk") is a TreeIso diagnostic. It counts stems in each instance's base, which only means something when the scan records stems, so it is not run for Canopy height.