Coral Cover: BVI Necker 2026 (snorkels)

Shaded value = in the CV mean (shade deepens with cover) Grey italic = shown for comparison but left out of the CV mean: the gold standard and manual CoralNet annotations (human reference values, not models), Coralscapes (not trusted), and any value that reuses the reference's own data n = frames behind the value — = not run or not applicable (see each region's notes) Site names and method names link out (new tab)

Coral Cover: BVI Necker 2026 (snorkels)

5 sites · 324 photos

Gold standard = 35 manually scored (via CoralNet) photos from Red Dock.

Site	Date	Claude	Gold standard	DINOv3	CoralSCOP	Mean ± SD of CV	Range
Red Dock	2026-07-01	8.1n=87	6.3n=35	20.2n=93	20.8n=93	16.3 ± 7.2	8.0–20.8
Turtle Beach	2026-07-01	3.3n=67	—	15.8n=70	17.7n=70	12.3 ± 7.8	3.3–17.7
Bali Hai to Turtle Beach	2026-07-02	2.5n=89	—	13.8n=89	12.1n=89	9.5 ± 6.1	2.5–13.8
Turtle Beach round 3	2026-07-02	1.9n=16	—	19.6n=21	14.4n=21	12.0 ± 9.1	1.9–19.6
Watersport Beach to Bali Hai	2026-07-02	2.4n=51	—	11.5n=51	12.3n=51	8.7 ± 5.5	2.4–12.3
All sites (unweighted mean)		3.6	—	16.2	15.4	11.8 ± 7.0	3.6–16.2

Notes for BVI Necker 2026 (snorkels) (12)

[a] No hand-labelled frame at this area.
[b] CoralNet robot labels exist on the other 275 frames, but the source's label set has NO scleractinian class (0 of 8,250 robot points are hard coral; prep_necker.py), so the robot output is structurally unusable - blank, not 0.
[c] DINOv3 (in-house, frozen Alor/Indonesia model) = mean hard-coral probability over 20 synthesized points per photo, NO water/fish class, so not a benthic denominator (from the map build's hard_coral_cover_by_site.csv).
[d] GENUINELY NOT RUN: no Coralscapes v1/v2 output for Necker anywhere on disk (backlog 09-25; re-checked maps/bvi_necker_map_build/, model_comparison_audits/, _from_external_20260925/).
[e] CoralSCOP = per-frame sum of colony-mask area fractions (zero-filled; overlapping masks double-count), averaged over frames with an ok row. Area index, not a point count: denominator differs from any point-count reference.
[f] Not applicable: Necker's only '-edited' folder holds 2 frames (EDITED_UNEDITED_RULES.md).
[g] Claude = blind HC/NOT/UNK calls on the unedited frames, 30 CoralNet points per frame, cover = HC/30 (all-30: WATER/FISH are unknown outside the reference frames), mean of per-frame cover - the arm's own convention (RESULTS_20260919.md). 310/310 frames.
[h] Gold = Anderson's own CoralNet labels on the 35 frames whose 30 points he labelled himself (all at Red Dock), all-30 denominator, HC set as in necker_claude_arm/score_necker_claude_20260919.py.
[i] THIN SAMPLE (under 30 frames at this site): indicative only.
Mean/Range: Mean ± SD and Range are over the CV cells that enter the mean (CoralNet where independent, DINOv3, CoralSCOP unedited/edited, Claude). Gold standard, frames manually annotated in CoralNet, Coralscapes and same-measurement cells are shown but never averaged (*).
Methods: Claude (Anthropic); DINOv3 (Meta; paper; used here as the frozen backbone of the in-house classifier); Coralscapes v1 and v2 (EPFL-ECEO; dataset paper, code); CoralSCOP (Zheng et al., CVPR 2024); CoralNet.
Sites: names drop the region prefix (hover for the full name) and link to the site's CRD page, or to the regional summary page where the site has none. Date = survey date from the map pins (first–last visit); sites are listed earliest first, undated sites last.

All regions