QSM benchmarks¶
Destructive harvest¶
Single-tree QSMs are scored against felled-tree volumes: 72 trees from Momo Takoudjou et al. (2018), Gonzalez de Tanago et al. (2018) and Burt et al. (2021), with wood density assigned per tree and agreement statistics computed the same way for every method; rayextract is raycloudtools' tree reconstruction (Devereux et al. 2026):
| method | volume bias | rRMSE | CCC | DBH bias / rRMSE |
|---|---|---|---|---|
| rayextract | −11.8 % | 48.6 % | 0.920 | −23 % / 29 % |
| sylva | −3.8 % | 19.9 % | 0.989 | +2 % / 13 % |
Study biases are Momo −5 %, GdtM −3 %, Burt +1 %. The remaining scatter is a handful of co-dominant Peruvian trees whose second limb is barely sampled: the pipe model has to guess it. Never cap or anchor stem radii on a breast-height slice: at 1.3 m the big tropical trees are 3 m-wide buttress stars where a circle explains 15 % of the points.
Fitting on other wood inputs (graph-based labels, and all points weighted by a wood confidence) was scored the same way on 73 of these trees; none beat the passage filter (22.0 % rRMSE on that set, against 22.6-26.6 %); the table is in the QSM guide.
CHERLET plots against raycloudtools¶
On the CHERLET plots, per matched tree, sylva / raycloudtools volume medians
are Wytham 0.78 (0.86 for trees over 20 cm DBH; plot totals 126 vs 126 m³),
Ofental 0.69 (totals 74 vs 60 m³), Litchfield 0.38 (0.85 over 20 cm; totals
17 vs 22 m³) and Robson 0.19. Against the earlier chain-prior model the
plots gained ten times as many cylinders per tree and branch orders 4–5
instead of 1–3. QSM DBH agrees with the detected DBH to within ~15 % on
the temperate plots; on the rainforest saplings the wood filter's medium
step lowers it (Litchfield 0.72, Robson 0.49 against 0.83 / 0.62 with
medium_threshold=1.0). The big Robson trees that sylva models at a
tenth of rayextract's volume are detection failures (a 0.08 m stem
candidate on a 0.8 m tree) and truncated segments, not QSM ones. Segmentation
leaves points below 0.5 m unassigned unless they sit within 1 m (1.5 DBH)
of their tree's base (segment_trees(low_height=, low_radius=)): the
graph otherwise reaches ground remnants, litter and understorey along the
surface and the QSM models them as fat horizontal cylinders (a Wytham oak
lost 0.4 of 2.7 m³ that way); the rule does not change the detection
scores. The Robson gap is the single-scan rainforest crowns and trunks:
where a trunk is sparsely sampled the anisotropy wood filter keeps few of
its points, the stem goes unmeasured and its radius comes from the prior.