Growth form and smoothed crown outlines

What changed

• 3,352,099 crowns on 3,101 survey tiles were scored for growth form. 2,771,551 (82.7%) show one; 580,548 (17.3%) read growth form not yet determined.¹
• A blind audit of the scores on 12 October 2026 checked the calls before publication against a machine-learning image-interpretation model’s labels (one pass; growth form only; no field check): 0.899 of decided audit crowns agreed with the reference labels under the audited rules (95% interval 0.828–0.956). Two kinds of call that fell short are withheld and one is reworded; with those changes the figure is 0.921 (0.848–0.980), computed after the rules were set.
• Every crown is drawn with a smoothed outline of its release 5 cells: 3,352,099 release 5 crowns and 534 crowns drawn from a record in release 5.2. Crown area, height and every estimate keep the cell area.
• Benefit estimates follow the growth form where the release had no recorded species. Estimated annual benefits rise by NZ$3,186,004 a year; stored carbon falls by 114,533 t CO₂e and avoided runoff by 483,669 m³ a year.
• 354 reviewer corrections from the ledger are accounted for, one of them applied to the map: a notable tree now stands on the crown its plaque marks.

Growth form

A classifier reads each crown’s structure in the 2024 LiDAR returns, its surroundings and image features from the 2024 aerial photograph at the crown top. It was trained on about 22,000 crowns: 19,329 Council tree register records (97% on the central isthmus), plus 2,659 crowns outside the isthmus labelled from the 2024 aerial imagery and LiDAR profiles by a machine-learning image-interpretation model (one pass; no independent second labeller yet). It was checked on 600 fresh crowns, labelled the same way, that it never saw. A separate palm detector confirms palms it can see from above. Image features are DINOv2 embeddings (Meta; Apache License 2.0); the classifier is LightGBM (Microsoft; MIT licence); the palm detector is a RetinaNet built on DeepForest (Weecology; MIT licence) and torchvision (BSD 3-Clause). Palm detection was checked against iNaturalist research-grade observations obtained through GBIF.org (doi:10.15468/ab3s5x, accessed 2 Oct 2026), each under its observer’s licence; Sources and credits gives the citations. Crowns flown on 30 April, 4 July and 21 October, when deciduous trees are in leaf, are scored by a variant trained for leaves on the trees (417,335 crowns).²

A growth form is shown only where calls of its kind agreed with the reference labels often enough when tested on labelled crowns: at least 8 in 10 for a firm call and at least 6 in 10 for a “possibly” call. Otherwise the crown reads “growth form not yet determined”. After the audit no “possibly” call is shown in this release. On the leaf-on flights only evergreen broadleaf calls and palms seen by the palm detector are shown. The map and the tree page mark every growth form as an estimate, and a species recorded by the Council or another source stays the species shown, with the growth form beside it.

Crowns not yet determined: 304,461 where the model abstained, 111,339 middle-band conifer calls outside the isthmus, 108,528 palm-type middle-band calls the palm detector did not confirm, 45,693 other calls on the leaf-on flights, where only evergreen broadleaf calls and palms seen by the palm detector are shown, and 10,527 “possibly conifer” calls on the isthmus withheld after the audit.

Outside the isthmus every deciduous call reads “Broadleaf”: there the model does not reliably tell deciduous trees from kānuka-type scrub, whose porous crowns look bare in the LiDAR returns. A palm-detector call on a crown wider than 10 m is not used on its own, because such crowns are often several trees; 1,028 crowns fall back to the model’s own call.

The 12 October audit

Before publication, about 200 crowns drawn from what the release would show, stratified by growth form, area and flight season, were labelled the same way, by the machine-learning image-interpretation model, together with abstentions and decoys, blind to the classifier’s calls, from the 2024 aerial imagery (7.5 cm) and LiDAR profiles: one pass, no independent second labeller yet. The labels were hashed before the key was opened, and the pass rules were written before any label existed. Crowns used to train, tune or test the model were left out.³

Agreement with a machine-learning image-interpretation model’s labels (one pass): 0.899 (0.828–0.956) of decided audit crowns under the audited rules; 0.78 counting undecidable crowns; growth form only; no field check. The figure is weighted to the metro (186 of 214 decided crowns) and cleared the bar of 0.85. Two kinds of call failed or nearly failed their own bar and are withheld; one wording was changed:

Crowns: the crowns each change applies to in this release. Middle-band palm-type calls are withheld on the isthmus too (13,670 crowns), as before the audit.

Smoothed outlines

Release 5 traced crowns along the 0.5 m cells of the LiDAR survey, so they looked square. Release 6 draws each crown with a smooth outline around the same cells. The crown’s area, height and every estimate keep the cell area; the outline’s own area is stored beside it as the crown outline area. Outlines keep 0.995 of the cell area at the median (1st–99th percentile 0.939–1.019; 0.997 in total). Crowns drawn from a record in release 5.2 are smoothed the same way.

The property pages count canopy from the redrawn outlines, so the map and a property agree: canopy over parcels is 185.65 km², 0.34% less than from the cells (186.28 km²). Smoothing lets neighbouring outlines touch: overlapping neighbours share 10,832 m² in all, 0.005% of the counted canopy, which the property totals can count twice.

Benefit estimates

Where a crown has no recorded species, its benefit estimates now follow the growth form. For a shown call, leaf area and the curve from height to trunk diameter are averaged over the four growth-form classes, weighted by the model’s probabilities; every other class parameter (rain interception, biomass, root-to-shoot ratio, carbon fraction) follows the shown call, as release 5 let its class drive every parameter. A crown whose growth form is not yet determined takes the evergreen broadleaf default, as release 5 did; an older image-model class, uncalibrated, is replaced by that default. Recorded species keep their estimates.⁴

Release totals, sensitivity scenario: about NZ$292 million → about NZ$295 million a year; avoided runoff about 21.0 million → about 20.5 million m³ a year; stored carbon about 3.49 million → about 3.38 million t CO₂e. The sums by cause match the change in the totals to within NZ$1, 3 m³ and 24 t CO₂e. Outlines do not enter the estimates.

Most of the rise in value comes from crowns the older image model had called conifers: 498,012 of them are now evergreen broadleaf. Most of the fall in stored carbon comes from crowns now called conifers (−301,504 t CO₂e), 111,116 of which release 5.2 valued as evergreen broadleaf trees: for the same height the conifer curve gives a thinner trunk and less biomass.

Other readings of the rule were tested: keeping the release 5.2 class for the parameters the rule does not name would add NZ$4,979,744 a year, averaging every estimate over the classes NZ$2,578,258, and treating every palm-detector crown as a palm NZ$3,134,091. The estimates remain sensitivity scenarios.

Trunk diameter

Trunk diameter (DBH) is the diameter at 1.3 m above the uphill ground. No trunk in this release was measured. Each crown holds two model estimates: one from crown size and height, unchanged from release 5, and one from the benefit equations, which from this release follows the growth form where no species is recorded (its height-to-diameter curve is weighted as described under Benefit estimates). The tree panel and tree page show their median as the central estimate, marked Estimate, with the two models as its range. A trunk measured with a tape at 1.3 m at a reviewed KYTE ground visit replaces both. Tree ferns take the palm curve until a tree-fern value is chosen.

The root protection radius on the tree page is 12 times the central trunk estimate, its area capped at 707 m² (a 15 m radius): a planning scenario, not a measured root extent.

Corrections ledger

Reviewer decisions about trees, crowns and records are kept in an append-only ledger. After the build, each decision is applied to the release or recorded with the trees it names, and the release publishes a receipt of every outcome. Each tree page lists the decisions that name the tree, its records or its crown.⁵

The applied correction moves a notable tree: Council notable tree schedule 1186 at Devonport is the plaque tree, a 191 m² crown, not the tree by the basketball court where the Council point sits. Its record now represents the plaque tree’s crown, so the map shows one marker fewer (3,574,824) and estimated benefits fall by NZ$12 a year. Moving field evidence between trees waits for a full build and is pending.

What stays the same

Every release 5.2 record, link, marker, source position, survey history, review decision and accepted KYTE visit is kept, with the release 5 growth-form fields and service rows, in the release database. Crown boundaries and sizes are those of release 5 and 5.2. The map shows 3,570 protected trees and counts 3,523,024 crowns, as in release 5.2. Low and sub-canopy vegetation and the Council recorded positions are unchanged.

Limits

• A growth form is a coarse class, not a species. Shown calls agreed with the reference labels for about 94% of the 600-crown blind test outside the isthmus (design-weighted) and 0.899 of decided crowns on the audit; the conifer and palm-type figures rest on small samples.
• The labels made from imagery, used in training, in the blind tests and in the audit, come from a machine-learning image-interpretation model in one pass, from aerial imagery and LiDAR profiles; no independent second labeller has checked them, so agreement between labellers is not measured.
• “Likely conifer” outside the isthmus is mostly tall forest crowns, 21–38 m, where pine, eucalypt and podocarp cannot be told apart from above. The audit could neither confirm nor reject its 0.80 bar.
• Tree ferns use the palm leaf area and trunk curve until a tree-fern value is chosen.
• Crowns drawn from a record in release 5.2, crowns outside the release 5 survey footprint and records without a crown have no growth form.
• Outlines smooth the release 5 cells; crowns over 60 m² remain about 20% smaller than the reference outlines they were tested against. Low and sub-canopy polygons are not smoothed.
• Benefit and trunk estimates are not independently validated; they remain sensitivity scenarios.