Tree detection and segmentation

The detect → segment → prune pipeline is scored against manually segmented reference plots (per-point tree labels) as instance segmentation: a reference tree counts as found when its best-overlapping sylva tree has IoU ≥ 0.5; predicted trees outside the reference coverage are ignored. raycloudtools' rayextract trees (Devereux et al. 2026) output is scored the same way.

Site (points)	sylva F1	raycloudtools F1	sylva TP / ref	sylva mean IoU
LITCH_CHERLET (8 M)	0.96	0.97	124 / 127	0.98
ROBSON_CHERLET (9 M)	0.57	0.55	91 / 149	0.73
WYTHAM_CHERLET (14 M)	0.77	0.73	125 / 180	0.85
OFENTAL_CHERLET (15 M)	0.57 (0.63 with min_quality_short=0.15)	0.73	56 / 89	0.69

(Those figures are the per-site runs recorded earlier.)

Scored the way the benchmark scores

Cherlet et al. (2026) evaluate only the trees that lie at least 90 % inside the test sub-plot (128, 181, 89 and 150 of them, which is what the in_plot_th0.90 folders hold), and a prediction whose best overlap is with an edge tree is neglected rather than counted against precision. Matching is Hungarian on the IoU matrix, true positive at IoU ≥ 0.5. Scoring against every labelled instance instead, and calling every edge-tree hit a false positive, makes the numbers look far worse than they are.

Scored with a port of the benchmark's own evaluation code (qforestlab/TreeInstSegEval: IoU on point sets rounded to 1 cm, Hungarian matching on the evaluated trees, a true positive at IoU > 0.5), on the test splits, with min_height set to the 5 m the references themselves use (6 m at Robson Creek; prune_trees keeps 3 m by default). The figures an earlier version of this page gave (0.976, 0.732, 0.667, 0.528) could not be reproduced with that code and are superseded. With the current defaults (power=6):

test split	eval trees	recall	precision	F1	Rayextract F1	best F1 of the other four methods
Litchfield (savanna)	128	0.945	0.984	0.964	0.919	0.930 (TreeLearn)
Wytham (temperate broadleaf)	181	0.702	0.751	0.726	0.570	0.540 (TreeLearn)
Ofental (conifer)	89	0.753	0.827	0.788	0.744	0.497 (TreeLearn)
Robson Creek (rainforest)	150	0.513	0.475	0.494	0.480	0.247 (ForAINet)

The other methods' values are those of Cherlet et al. (2026), Table 2 (Rayextract, Treeiso, SSSC, and TreeLearn and ForAINet fine-tuned on all four plots). Sylva's are computed here with the same protocol, so the comparison is indicative rather than independent.

power (the exponent on graph edge length) was raised from 4 to 6 after a sweep on the benchmark's validation areas, never on the test splits: F1 there went 0.514 → 0.563 at Robson Creek, 0.636 → 0.660 at Ofental and 0.581 → 0.588 at Wytham, and Litchfield did not change. On the test splits, scored once with the setting fixed, it gave 0.479 → 0.494, 0.682 → 0.788, 0.707 → 0.726 and 0.964 → 0.964. Settings that did not help on the validation areas: gravity, wood_costs (better at Ofental only), a weaker or no height_prior, and two forms of a crown-size prior from stem allometry (crown radius proportional to DBH, applied to every path or only to points two trees compete for), which truncated real crowns or fragmented trees.

What remains. Stem detection is nearly complete (one of the 150 Robson Creek trees is never detected); the losses are in dividing crowns between neighbours. Most missed trees are shorter than their neighbours and are absorbed by them: 49 of 76 at Robson Creek, 44 of 57 at Wytham and 21 of 31 at Ofental before the change of power.

The reporting threshold matters more than anything else we tuned, because the references only label trees: at Litchfield, moving it from 3 m to 5 m took F1 from 0.810 to 0.976 under the earlier scoring, without losing a single tree, since everything it drops is shrub. Keep the threshold you segment with separate from the one you report, and set the second to match whatever your inventory calls a tree.

Seeds used to be too wide. A tree starts from the points within seed_radius of its stem at seed_height; at 0.5 m those discs overlap wherever stems are close, and the stronger seed takes the other's crown. Of the 102 Wytham trees missed before, 79 had been swallowed by a neighbour whose stem stood a median of 0.3 m away: coppice stools and low forks, not distant trees. Halving seed_radius to 0.25 m and merge_radius to 0.2 m finds 22 more trees (130 → 152), lifts mean IoU 0.83 → 0.85 and precision 0.88 → 0.90, and costs Litchfield nothing (135 → 136).

A sparser graph and a steeper cost help on top of that: k 10 → 6 and power 3 → 4 (the exponent on edge length, so a path pays more for a long hop between two crowns; since raised to 6, above). Found trees: Litchfield 136, Wytham 152 → 154, Robson 105 → 110, Ofental 63 → 68, with fewer spurious trees everywhere.

A finer graph pays for itself where the stand is dense: the cloud is thinned to voxel_size before the graph is built, and 0.05 → 0.03 m gives Wytham 154 → 156, Robson 110 → 117 and Ofental 68 → 72, each with fewer false trees and a better IoU, for about 1.7× the time. Litchfield does not move at any resolution. 0.02 m gains two more at Wytham for 2.5× the time, which did not seem worth making the default.

Other things tried, none of which helped: a denser graph (k=20, Wytham 136), gravity 0.3 (140), wood costs (142), dropping the height prior (139 at Wytham, 61 at Ofental), a wider height-prior radius (62 at Ofental), no angle penalty (48 at Ofental, much worse), max_edge 0.5 (no change), and no pruning at all (142), so the limit is not the candidate list.

Where the remaining loss is. Robson and Ofental are still the hard ones, and for different reasons. At Ofental 41 of the 54 missed trees are absorbed by one neighbour, and the missed trees are short (median 10.1 m) where the matched ones are tall (16.4 m): suppressed conifers under dominants, whose stems are detected (81 % have a candidate within 0.5 m, median 0.05 m) but whose crowns are taken by the tree above. Robson keeps 222 reference trees against 655 candidates, so it over-detects in the understorey as well as merging. Both point at the cost model rather than at detection or pruning.

Of Robson's 212 spurious trees, 123 sit almost entirely (> 80 % of their points) on what the reference calls ground (rainforest understorey, median 6.2 m tall), and only 40 are fragments of a labelled tree. The same tension as the savanna shrubs: real vegetation the benchmark does not count as a tree. prune_trees(min_quality_short=0.15) trims about a quarter of them for two found trees.

Two ideas from other systems were implemented and measured, and both made things worse on this data:

• A lateral-offset prior. raycloudtools multiplies every edge by 1 + g·(horizontal offset from the seed)², which is how it stops a dominant reaching sideways into a neighbour. Sylva has it as gravity. Ofental barely moves (F1 0.667 → 0.674 at g = 1, with mean IoU falling 0.718 → 0.694) and Robson falls apart: 0.521 → 0.487 → 0.450 → 0.437 for g = 0.1, 0.3, 1. In raycloudtools the term multiplies a cost already normalised by canopy height; here it stacks on length⁴, so a 10 m-wide rainforest crown pays about ×31 on every edge.
• Edge weight as the gap between clusters rather than the distance travelled. This is TLS2trees' (Wilkes et al. 2023) documented fix for a path that prefers a suppressed tree's base. Grouping points into 0.25 m cubes and pricing each edge by the closest approach between two groups gives Ofental F1 0.667 → 0.158 and Robson 0.521 → 0.150. The reason is instructive: with gap weights, groups that touch cost nothing, so in a closed canopy the whole plot becomes one cheap component and a single seed takes it. TLS2trees avoids this by running the gap graph on wood points only, with clusters from 0.2 m slices, and attaching foliage afterwards from the branch tips of the finished skeletons. Done that way it might work; done naively it does not, and the code was removed rather than shipped as a knob.