Source-linked AI summary
Growing a Stand, Not a Tree: Joint Canopy Generation Reproduces Crown Shyness
Guang Yang, Fengchen Liu
TL;DR
Crown shyness poses a generative-modeling question because its defining gaps exist between neighboring trees rather than within one tree. The paper uses flow matching with cross-tree attention on simulation stands certified to contain irreducible coupling, and compares it with matched independent generation. The joint model halves clearance error on unseen densities, retains its advantage outside training, and agrees with selected field measurements after one-scalar calibration.
Problem
Crown shyness is an inter-tree pattern, creating a test of whether learned set generation can model structure that no individual tree owns.
Method
The paper formulates canopy generation as joint set generation with flow matching and attention between trees, using stands whose coupling is certified irreducible to per-tree geometry.
Results
0.415 m versus 0.894 m clearance distribution error gives joint generation a 53.6% improvement over identical-capacity independent generation on unseen spacings, with the advantage persisting outside training.
Takeaways & Limitations
Crown shyness is reproduced as a stand-level property, while field agreement is obtained for gap magnitude and crown asymmetry after one scalar calibration.
Takeaways & Limitations
The model is trained on jittered-lattice stems and fails to generalize to hard-core Poisson stems; directional claims are also bounded by a single 23-tree field reference.
Abstract
from arXiv · showhide
In closed forests, neighboring tree crowns often stop short of touching, leaving a network of narrow gaps known as crown shyness. The pattern belongs to the stand rather than to any single tree, which makes it a natural probe of a question in generative modeling: can a learned model produce a set of objects whose defining structure exists only between them? We formulate stand-level canopy generation as set generation with a flow-matching model, in which attention between trees is the only channel through which coupling can arise. Trained on stands grown by a resource-competition simulation that is provably not reducible to per-tree geometry, the joint model halves the clearance distribution error of an identical-capacity model that generates each tree alone, and the advantage persists at stem densities outside the training range. Against field measurements of a tropical oak forest, a single calibrated scalar yields held-out agreement in gap magnitude and crown asymmetry. The directional statistics of the gaps are controlled by stem placement rather than by the growth rule, and match the field once stem jitter is calibrated. Crown shyness, in both the simulation and the learned model, is a property of the stand and not of the tree.
1. Introduction
The paper treats crown shyness as an inter-tree pattern and asks whether a learned generator can produce it jointly rather than impose it after independent tree generation. Matched experiments and field comparison show that coupling through attention improves unseen-density generation and supports held-out realism.
- Motivation: Crown shyness consists of narrow gaps shaped jointly by neighboring crowns, so no single tree carries the pattern.The gaps disappear when either tree is removed from the scene.
- Research gap: Independent object generation typically adds interactions afterward, whereas this paper tests whether inter-object structure emerges inside joint generation.Procedural overlap-resolution can produce plausible images but does not make the result depend on jointly generated objects.
- Research question: The paper asks whether a learned model can generate whole stands so crown shyness emerges between trees.The question requires ruling out failures in the data, comparison, and realism tests.
- Data validation: 74 to 100% of stems respond to perturbing one tree’s height across 5 spacings, certifying genuine coupling in the training distribution.The response survives 8-fold growth-step refinement, while the best independent-crown truncation still misses by 6.6 growth steps.
- Joint generation: 0.415 m versus 0.894 m clearance distribution error gives the joint model a 53.6% improvement over identical-capacity independent generation on unseen spacings.The comparison uses 12 seeds, p = 0.00049, and the advantage persists at both denser and sparser stands outside training.
- Field validation: One calibrated scalar yields held-out agreement with tropical oak field measurements in gap spread and crown asymmetry, while gap direction depends on stem placement.Calibrated stem jitter brings the directional statistic into the field band without changing gap magnitude or crown shape.
2. Related Work
Prior work generates individual trees or jointly arranges rigid scene objects, while procedural ecosystems resolve interactions with explicit rules. This paper targets learned stand generation in which neighboring crowns deform one another through emergent coupling.
- Single-tree generative models: Learned tree-generation work covers developmental rules, shape spaces, point clouds, voxels, reconstruction, and inventory-conditioned flow matching.A survey found twenty-three learned single-tree generators and no learned model generating a stand jointly.
- Single-tree generative models: TreeFlow conditions on species, platform, and height but has no mechanism for one tree to influence another.The paper therefore does not treat it as a baseline for joint stand generation.
- Procedural ecosystems: Procedural ecosystem systems produce convincing forests by growing or placing plants independently and resolving interaction with explicit rules.Their relevance here is defining a boundary: rule-produced gaps encode the hand-written procedure rather than a learned relationship.
- Set generation: Existing set-generation and scene-synthesis methods provide permutation-invariant attention and continuous-generation machinery, but their rigid assets do not respond in shape to neighbors.This paper uses standard components to isolate whether attention buys emergent inter-object structure at matched capacity.
3. A Teacher That Cannot Be Faked
The teacher simulation couples crowns through occupancy and abrasion, producing gaps as an emergent equilibrium rather than prescribing gap widths. Strict ablations and irreducibility tests show that its crown shapes cannot be reduced to independent tree geometry or stem positions alone.
- Coupled growth with contact abrasion: Occupancy freezes growth at contact, while abrasion retracts contacting sectors, jointly creating gaps without allowing crown overlap.Gap width emerges from the balance between growth pressure and contact loss; the model records 0.00 m overlap in all used runs.
- Visual comparison: Figure 2 places real, teacher-generated, and learned canopies side by side; at unseen spacing, the generated model reproduces the teacher’s gap-channel structure.The comparison is viewed from above and colored by height.
- Both mechanisms are necessary: The two mechanisms are nonredundant: occupancy sets where crowns stop, abrasion determines how far they retreat, and both are required for positive clearance.With coupling off, clearances are negative; with abrasion off, crowns meet at +0.028 m; with both on, the first three clearances are +0.364 m, +0.390 m, and +0.421 m.
- Both mechanisms are necessary: A strict gate requires positive clearance toward the three nearest neighbors while preserving canopy cover, and 14 of 16 swept configurations pass.A single die-back fraction covers every tested spacing, indicating a regime rather than a narrow tuning.
- Irreducibility: The best-fitting Voronoi-style truncation still misses the teacher by 0.55 m per sector, or 6.6 growth steps, showing stem positions alone cannot reproduce its crowns.The discrepancy follows from contested-space resolution depending on both trees’ heights through the vigour factor.
4. Method: Joint Generation of a Stand
The paper generates crown shapes jointly for local stands, using attention to model interactions between neighboring trees. It evaluates the resulting inter-crown clearances with a shared canonical estimator and controlled comparisons.
- Stand representation: A stand is represented as an unordered set of 24 nearby trees, each carrying stem position and a 24-direction crown-radius vector.Positions are normalized by spacing so the generator receives local neighborhood geometry.
- Flow-matching model: The flow-matching target is each crown’s directional deviation from its mean radius, so the model generates crown shape conditional on crown size.The deviation profile is generated while the mean radius is supplied as an input feature.
- Flow-matching model: Transformer blocks use multi-head attention across the stand, allowing trees to exchange information during joint generation.Sampling integrates the learned velocity field from noise in 100 Euler steps.
- Controlled comparison: Ablation figures compare joint generation and growth mechanisms using identical stems and crown draws, showing how attention and abrasion affect canopy gaps.The figures report clearance and crown-intersection comparisons for the joint, ablated, and mechanism-removed conditions.
- Controlled comparison: Disabling attention creates an identical-capacity independent model in which each tree is generated alone, isolating inter-tree coupling as the experimental variable.Both arms use the same data, seeds, schedule, and 300-epoch training procedure.
- Evaluation: The shared evaluation instrument measures nearest-neighbor crown-edge clearance, with positive values denoting gaps and negative values denoting interpenetration.Evaluation uses the Wasserstein-1 distance between generated and teacher clearance distributions, alongside noise-floor and direction-blind reference levels.
5. Experiments
Experiments compare matched-capacity joint and independent generation, validate the teacher against field measurements, and separate growth-rule effects from stem-placement effects. Joint generation improves clearance modeling beyond training spacings, while field agreement depends on one calibration and directional realism depends on stem placement.
- Experimental setup: Training uses jittered-lattice stands at 3.5–6.5 m spacings, with unseen evaluation at 4.0, 5.0, and 6.0 m and extrapolation at 3.0 and 7.5 m.Learning experiments fix the die-back fraction at 0.06; field comparisons use a separately calibrated value of 0.16.
- Joint versus independent generation: 0.415 m clearance distribution error for joint generation versus 0.894 m for independent generation on unseen spacings, a 53.6% improvement over 12 seeds.The matched-capacity models use identical data, parameters, and schedule; shuffling stem positions degrades the joint model by 2.50×.
- Extrapolation: 62.5% improvement at 3.0 m and 41.8% at 7.5 m show that the joint advantage persists at both ends outside the training range.The pooled Wilcoxon p-value is 0.00049, and the joint model remains below the direction-blind baseline at both spacings.
- Stand-level structure: Joint and uncoupled stands built from identical stems and crown draws show the gap network closing when coupling is removed.Figure 7 presents the stand across 5, 4, and 3 m mean stem spacings, with 85%, 92%, and 96% of crowns contacting a neighbour.
- Field validation: A single calibration to median clearance yields held-out clearance spread at 1.57× the field spread and crown elongation of 1.450, both inside observed bands.Crown elongation was not calibrated; the uncoupled twin is rejected at 31.3× the field spread.
- Directional gap statistics: Directional anisotropy is too low on the 4% jitter lattice but enters the field band when stem jitter rises from 4% to 8%, while gap magnitude and crown shape remain in-band.The attribution experiment holds the teacher, die-back, estimator, and seed discipline fixed, assigning the anisotropy shortfall to stem placement rather than the growth rule.
6. Conclusion
The paper shows that crown shyness can emerge from jointly generated trees when the training data contains irreducible inter-tree coupling. The result is supported by matched-model comparisons, field validation, and explicit scope limits.
- Attention between trees halves clearance error on unseen stem densities relative to an identical-capacity independent model.The joint model retains its advantage outside the training range.
- A perturbation test, discretization control, and best-case reduction establish that the training simulation cannot be rewritten as per-tree geometry.The best-case reduction still misses by 6.6 growth steps.
- One calibrated scalar yields held-out agreement with field measurements in gap magnitude and crown shape.The field comparison uses measurements rather than visual plausibility.
- Directional realism is attributed to stem placement rather than the growth rule.The attribution experiment separates directional effects of stem arrangement from those of growth.
- The study is limited to two-and-a-half-dimensional crowns, jittered-lattice stems, one 23-tree field reference, and calibrated jitter amplitude.Generalization across stem processes and extension to full 3-D crowns remain open.
- Measured statistics are generated from frozen result files by reproducible scripts, with design constants and evaluation artifacts traceable to code or sources.The simulation, training code, estimator, result files, and audit tooling are slated for release.