Reef cover by method

Hard-coral cover (%) at each site from every method run so far, one table per region, newest trip first (23 regions, 333 sites). Built 29 September 2026 from the method output files; the Red Sea Claude and CoralSCOP runs are still in progress.

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)

Halmahera 2026

23 sites · 3,593 photos

No gold standard.

SiteDateClaudeDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
Failonga2026-08-0932.6n=20131.4n=20155.3n=20158.5n=20131.7n=20131.9 ± 0.631.4–32.6
Maitara North2026-08-0933.1n=17133.6n=17148.6n=17151.0n=17126.3n=17131.0 ± 4.126.3–33.6
AM Point 22026-08-0940.4n=22333.4n=22357.1n=22360.5n=22328.0n=22333.9 ± 6.228.0–40.4
Pretty Shallow2026-08-1032.3n=10532.0n=10553.7n=10555.6n=10530.2n=10531.5 ± 1.130.2–32.3
Kusu Drift2026-08-1026.1n=14327.6n=14351.3n=14356.2n=14333.9n=14329.2 ± 4.126.1–33.9
Pulau Sebatang2026-08-1041.1n=26533.5n=26565.3n=26566.7n=26536.2n=26536.9 ± 3.933.5–41.1
Tower Corner2026-08-1134.0n=24235.4n=24255.4n=24257.6n=24234.4n=24234.6 ± 0.734.0–35.4
Batu Mangga2026-08-1120.9n=23635.8n=23652.1n=23655.0n=23632.9n=23629.9 ± 7.920.9–35.8
Sali Bay-Batu Mangga2026-08-1129.2n=12529.1n=12551.5n=12551.2n=12529.0n=12529.1 ± 0.129.0–29.2
Tanjung Naraka2026-08-1229.6n=13630.1n=13653.5n=13654.6n=13635.3n=13631.7 ± 3.229.6–35.3
Tanjung Dehebra2026-08-1225.0n=10327.7n=10351.8n=10354.0n=10334.3n=10329.0 ± 4.825.0–34.3
Kusu Ridge2026-08-1245.2n=24942.7n=24963.6n=24966.0n=24927.7n=24938.5 ± 9.527.7–45.2
Jerot2026-08-1322.9n=17227.5n=17255.2n=17256.8n=17241.6n=17230.7 ± 9.722.9–41.6
Proco2026-08-1332.0n=22731.9n=22751.5n=22754.2n=22733.8n=22732.6 ± 1.131.9–33.8
Kusu Drift (round 2)2026-08-1343.0n=10531.8n=10560.6n=10565.0n=10537.9n=10537.6 ± 5.631.8–43.0
Merida Cavern2026-08-1427.3n=12524.3n=12537.8n=12540.6n=12531.4n=12527.7 ± 3.624.3–31.4
Batu Balong shallows-near anchorage2026-08-1456.5n=21333.3n=21370.2n=21373.8n=21338.1n=21342.6 ± 12.233.3–56.5
Batu Balong2026-08-1439.4n=12628.0n=12660.0n=12661.3n=12634.1n=12633.8 ± 5.728.0–39.4
Batu Mandi2026-08-1411.9n=2217.0n=2234.0n=2236.4n=2228.4n=2219.1 ± 8.411.9–28.4
Siko Tanjung2026-08-1541.3n=14330.5n=14357.6n=14359.3n=14332.3n=14334.7 ± 5.830.5–41.3
Siko Pinnacle2026-08-1511.6n=4818.8n=4833.5n=4837.0n=4826.6n=4819.0 ± 7.511.6–26.6
Reene's Rock (Pulau Gewang)2026-08-1513.1n=3918.4n=3932.4n=3937.0n=3922.9n=3918.1 ± 4.913.1–22.9
Lelay Utara (north)2026-08-1647.8n=17431.1n=17457.2n=17458.5n=17430.1n=17436.3 ± 9.930.1–47.8
All sites (unweighted mean)32.029.852.655.132.031.3 ± 1.329.8–32.0
Notes for Halmahera 2026 (11)
  • [a] No gold standard exists (no transect, no frames manually annotated in CoralNet).
  • [b] Not applicable: no CoralNet source / hand annotations exist for these photos.
  • [c] Value read from cv_method_tables.json (injector output, generated 2026-09-28 22:33); the scope text there gives denominator and frame set.
  • [d] 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).
  • [e] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not trusted (Anderson): shown, never averaged. Area fraction, not a point count.
  • [f] 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.
  • [g] Not run by ruling: this is an UNEDITED region (EDITED_UNEDITED_RULES.md).
  • [h] Claude = blind 7-call point count on the UNEDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), UNK kept in; mean across photos (arm ledgers, via cv_method_tables.json).
  • 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

Alor 2026

16 sites · 2,723 photos

No gold standard.

SiteDateClaudeDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
Goat Fence2026-07-2810.7n=7219.3n=7250.2n=7249.9n=7215.4n=7215.1 ± 4.310.7–19.3
Pasir Merah (Red Sand)2026-07-2916.1n=17425.9n=17439.0n=17443.4n=17416.3n=17419.4 ± 5.616.1–25.9
Bakalang Jetty2026-07-297.6n=5115.8n=5124.7n=5125.6n=5118.1n=5113.8 ± 5.57.6–18.1
Bama Wall2026-07-2910.9n=23327.4n=23323.9n=23327.2n=2339.0n=23315.8 ± 10.19.0–27.4
Beang Abang (Baeng Beach / Baeng Jetty)2026-07-30–313.5n=2612.4n=2613.5n=2618.5n=2610.9n=268.9 ± 4.83.5–12.4
Goa Wolang (Walang Cave)2026-07-307.7n=2819.1n=2814.8n=2817.9n=288.8n=2811.9 ± 6.37.7–19.1
Clown Valley2026-07-3041.0n=20931.6n=20956.8n=20958.7n=20922.4n=20931.7 ± 9.322.4–41.0
Motonglang (Cathedral)2026-07-305.8n=13227.8n=13219.5n=13224.6n=1327.6n=13213.7 ± 12.25.8–27.8
Limarahing (Black Sands)2026-07-3128.9n=16231.6n=16249.0n=16250.9n=16218.2n=16226.2 ± 7.118.2–31.6
Apuri Bay2026-07-3128.7n=20828.3n=20848.2n=20849.8n=20720.2n=20825.7 ± 4.820.2–28.7
Monaseli (Munaseli)2026-08-0119.9n=17821.0n=17837.7n=17838.9n=17815.9n=17818.9 ± 2.715.9–21.0
Babylon2026-08-01–0230.0n=41530.5n=41552.4n=41555.2n=41520.6n=41527.0 ± 5.620.6–30.5
The Arch2026-08-0222.7n=15523.2n=15548.9n=15550.6n=15518.9n=15521.6 ± 2.418.9–23.2
Umapura2026-08-0217.7n=19728.7n=19741.3n=18647.7n=18417.5n=19721.3 ± 6.417.5–28.7
Harilolong (School's Out)2026-08-0336.2n=27530.6n=27551.7n=27553.7n=27521.0n=27529.3 ± 7.721.0–36.2
Sulambali Wall2026-08-0322.3n=20825.1n=20844.2n=20847.9n=20816.3n=20821.2 ± 4.516.3–25.1
All sites (unweighted mean)19.424.938.541.316.120.1 ± 4.516.1–24.9
Notes for Alor 2026 (11)
  • [a] No gold standard exists (no transect, no frames manually annotated in CoralNet).
  • [b] Not applicable: no CoralNet source / hand annotations exist for these photos.
  • [c] Value read from cv_method_tables.json (injector output, generated 2026-09-28 22:33); the scope text there gives denominator and frame set.
  • [d] 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).
  • [e] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not trusted (Anderson): shown, never averaged. Area fraction, not a point count.
  • [f] 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.
  • [g] Not run by ruling: this is an UNEDITED region (EDITED_UNEDITED_RULES.md).
  • [h] Claude = blind 7-call point count on the UNEDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), UNK kept in; mean across photos (arm ledgers, via cv_method_tables.json).
  • 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

BVI Necker 2026 (snorkels)

5 sites · 324 photos

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

SiteDateClaudeGold standardDINOv3CoralSCOPMean ± SD of CVRange
Red Dock2026-07-018.1n=876.3n=3520.2n=9320.8n=9316.3 ± 7.28.0–20.8
Turtle Beach2026-07-013.3n=67—15.8n=7017.7n=7012.3 ± 7.83.3–17.7
Bali Hai to Turtle Beach2026-07-022.5n=89—13.8n=8912.1n=899.5 ± 6.12.5–13.8
Turtle Beach round 32026-07-021.9n=16—19.6n=2114.4n=2112.0 ± 9.11.9–19.6
Watersport Beach to Bali Hai2026-07-022.4n=51—11.5n=5112.3n=518.7 ± 5.52.4–12.3
All sites (unweighted mean)3.6—16.215.411.8 ± 7.03.6–16.2
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

Anguilla 2026

4 sites · 302 photos

No gold standard.

SiteDateClaudeDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
Deep South2026-02-286.6n=6313.3n=639.8n=6311.3n=6320.5n=6113.5 ± 7.06.6–20.5
Oosterdiep Wreck2026-02-283.4n=6616.2n=666.0n=666.1n=6615.2n=6011.6 ± 7.13.4–16.2
The Pipe2026-03-026.6n=6519.4n=6510.2n=6511.0n=6519.3n=5815.1 ± 7.46.6–19.4
Catheley H2026-03-026.2n=10814.7n=1089.4n=1089.7n=10816.5n=10712.5 ± 5.56.2–16.5
All sites (unweighted mean)5.715.98.89.517.913.2 ± 6.55.7–17.9
Notes for Anguilla 2026 (11)
  • [a] PENDING Anderson's annotations (not hand-annotated yet; 'I'll do them later', 2026-09-29).
  • [b] PENDING Anderson's annotations (not hand-annotated yet; 'I'll do them later', 2026-09-29).
  • [c] Value read from cv_method_tables.json (injector output, generated 2026-09-28 22:33); the scope text there gives denominator and frame set.
  • [d] 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).
  • [e] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not trusted (Anderson): shown, never averaged. Area fraction, not a point count.
  • [f] 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.
  • [g] Not applicable: no edited set exists (backlog 09-25: 0 edited frames).
  • [h] Claude = blind 7-call point count on the UNEDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), UNK kept in; mean across photos (arm ledgers, via cv_method_tables.json).
  • 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

St. Eustatius 2026

4 sites · 299 photos

No gold standard.

SiteDateClaudeDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
The Ledge to Hangover2026-02-242.9n=8817.0n=8816.6n=8721.5n=8826.0n=8615.3 ± 11.62.9–26.0
STENAPA Reef2026-02-244.6n=10318.0n=1037.2n=1027.5n=10221.2n=9814.6 ± 8.84.6–21.2
Double Wreck2026-02-255.6n=4615.6n=4612.6n=4614.3n=4626.5n=4515.9 ± 10.45.6–26.5
Chien Tong wreck2026-02-253.8n=6216.6n=626.0n=625.6n=6213.3n=5711.2 ± 6.73.8–16.6
All sites (unweighted mean)4.216.810.612.221.814.3 ± 9.04.2–21.8
Notes for St. Eustatius 2026 (11)
  • [a] PENDING Anderson's annotations (not hand-annotated yet; 'I'll do them later', 2026-09-29).
  • [b] PENDING Anderson's annotations (not hand-annotated yet; 'I'll do them later', 2026-09-29).
  • [c] Value read from cv_method_tables.json (injector output, generated 2026-09-28 22:33); the scope text there gives denominator and frame set.
  • [d] 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).
  • [e] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not trusted (Anderson): shown, never averaged. Area fraction, not a point count.
  • [f] 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.
  • [g] Not applicable: no edited set exists (backlog 09-25: 0 edited frames).
  • [h] Claude = blind 7-call point count on the UNEDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), UNK kept in; mean across photos (arm ledgers, via cv_method_tables.json).
  • 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

Saba 2026

4 sites · 342 photos

No gold standard.

SiteDateClaudeDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
Ladder Labyrinth2026-02-218.9n=8218.3n=8212.7n=8215.4n=8222.3n=7816.5 ± 6.98.9–22.3
Man-of-War Shoals2026-02-213.7n=10219.8n=10216.0n=10125.1n=10129.3n=9817.6 ± 12.93.7–29.3
Tent Drift2026-02-214.3n=7317.7n=7311.7n=7314.5n=7322.1n=6314.7 ± 9.34.3–22.1
Tent Shallow2026-02-223.0n=8515.0n=857.9n=8511.7n=8521.3n=7213.1 ± 9.33.0–21.3
All sites (unweighted mean)5.017.712.116.723.815.5 ± 9.65.0–23.8
Notes for Saba 2026 (11)
  • [a] PENDING Anderson's annotations (not hand-annotated yet; 'I'll do them later', 2026-09-29).
  • [b] PENDING Anderson's annotations (not hand-annotated yet; 'I'll do them later', 2026-09-29).
  • [c] Value read from cv_method_tables.json (injector output, generated 2026-09-28 22:33); the scope text there gives denominator and frame set.
  • [d] 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).
  • [e] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not trusted (Anderson): shown, never averaged. Area fraction, not a point count.
  • [f] 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.
  • [g] Not applicable: no edited set exists (backlog 09-25: 0 edited frames).
  • [h] Claude = blind 7-call point count on the UNEDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), UNK kept in; mean across photos (arm ledgers, via cv_method_tables.json).
  • 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

St. Maarten 2026

2 sites · 192 photos

No gold standard.

SiteDateClaudeDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
Mike's Maze2026-02-1913.5n=7819.4n=7828.1n=7733.4n=7725.2n=7519.4 ± 5.813.5–25.2
Proselyte Reef2026-02-197.9n=11418.4n=11420.7n=11422.4n=11423.9n=11216.7 ± 8.17.9–23.9
All sites (unweighted mean)10.718.924.427.924.618.1 ± 7.010.7–24.5
Notes for St. Maarten 2026 (11)
  • [a] PENDING Anderson's annotations (not hand-annotated yet; 'I'll do them later', 2026-09-29).
  • [b] PENDING Anderson's annotations (not hand-annotated yet; 'I'll do them later', 2026-09-29).
  • [c] Value read from cv_method_tables.json (injector output, generated 2026-09-28 22:33); the scope text there gives denominator and frame set.
  • [d] 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).
  • [e] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not trusted (Anderson): shown, never averaged. Area fraction, not a point count.
  • [f] 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.
  • [g] Not applicable: no edited set exists (backlog 09-25: 0 edited frames).
  • [h] Claude = blind 7-call point count on the UNEDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), UNK kept in; mean across photos (arm ledgers, via cv_method_tables.json).
  • 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

Mauritius 2025

7 sites · 344 photos

Gold standard = 344 manually scored (via CoralNet) photos (draft manuscript).

SiteDateClaudeCodexGold standardCoralNetDINOv3Coralscapes v1CoralSCOPMean ± SD of CVRange
Reef Garden2025-02-0126.7n=10519.3n=15partial 15/10526.3n=10527.8n=10544.6n=10538.4n=9722.2n=10530.3 ± 9.822.2–44.6
Ti Korail2025-02-017.4n=400.0n=4partial 4/407.0n=4010.6n=4021.9n=4013.0n=3820.5n=4015.1 ± 7.27.4–21.9
Bamboo2025-02-0413.8n=4017.1n=8partial 8/4017.2n=4011.8n=4019.9n=4013.1n=3916.9n=4015.6 ± 3.511.8–19.9
Aquarium-Big Rock2025-02-048.3n=408.9n=6partial 6/4014.0n=407.0n=4015.3n=408.2n=3513.0n=4010.9 ± 3.97.0–15.3
Carpenter2025-02-0912.3n=4013.8n=8partial 8/4014.6n=4010.9n=4021.2n=4013.6n=3721.0n=4016.4 ± 5.510.9–21.2
Lost Anchor2025-02-099.2n=3912.0n=5partial 5/3910.5n=3910.0n=3920.3n=3911.9n=3820.5n=3915.0 ± 6.29.2–20.5
Anemone Garden2025-02-102.9n=4020.0n=3partial 3/402.7n=407.7n=4033.3n=4017.9n=3930.5n=4018.6 ± 15.52.9–33.3
All sites (unweighted mean)11.513.013.212.325.216.620.717.4 ± 6.711.5–25.2
Notes for Mauritius 2025 (11)
  • [a] Value read from cv_method_tables.json (injector output, generated 2026-09-28 22:33); the scope text there gives denominator and frame set.
  • [b] Gold = Anderson's manual point count, draft Mauritius manuscript (unpublished, v42 Table 3), ALL-30 denominator (water and fish left in); every CV cell is on the SAME 344 frames.
  • [c] On external drive (not verified): Coralscapes v2 ran on 574 unedited frames 2026-09-18 (~/Documents/Claude/logs/coralscapesv2_mauritius_20260918.log, last lines: CSV -> /Volumes/OneDrive/Coral computer vision/regions/Mauritius_unedited_coralscapesv2_20260918/benthic_coverage.csv); no local copy, no site values anywhere on disk.
  • [d] Coralscapes v1 (EPFL-ECEO/coralscapes-vit-b-dpt) on UNEDITED frames: 10 live hard-coral classes, area / whole frame, frames with >50% background dropped (mauritius-coral-cv README rule); recomputed in model_comparison_audits/mauritius_cv_review_20260928/per_site.csv (filtered=True, mode=live) from mauritius-coral-cv/data/coralscapes_benthic_coverage_*.csv. The pins file this value under the v2 id ('coralscapes'); it is v1.
  • [e] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not trusted (Anderson): shown, never averaged. Area fraction, not a point count.
  • [f] 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.
  • [g] Not applicable: Mauritius has no edited copies (EDITED_UNEDITED_RULES.md: 347/347 unedited).
  • 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 (*).
  • Codex: OpenAI Codex run as a blind point caller (tests 1–3 and revisions 4–6, Codex CV/03–11), 30 CoralNet points per photo, hard-coral points / 30 (all points in the denominator, abstentions count as not coral), mean across photos. 49 photos in all (3–15 per site) out of the 344 reference photos, so each cell is partial (n/N) and thin. Values = the paper's 49-image set as filed on the pins, reproduced exactly from mauritius-coral-cv/data/codex_point_calls_49_images.csv. Shown, not averaged: the supplement (v4, Table S11) says neither language model was run over a site's full frame set, so neither gives a comparable site-level estimate.
  • 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

Red Sea - Saudi Arabia 2023-24

22 sites · 4,554 photos

Gold standard = published values, by area only (no per-site gold). 583 photos manually scored (via CoralNet) for the benchmark. Benchmark photos mostly unedited; CoralSCOP ran on edited photos.

SiteDateClaudeClaude full setManually annotated in CoralNetDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
Abu Ejeiah (Thuwal)2023-09-02–2024-10-0411.3n=4partial 0/36613.5n=430.6n=414.4n=36715.7n=36726.2n=145partial 145/36722.7 ± 10.111.3–30.6
Cement Reef (Thuwal)2023-09-02–2024-02-2327.5n=9526.9n=142partial 142/47922.1n=9526.4n=9519.2n=47220.0n=47247.0n=1partial 1/47133.7 ± 11.626.4–47.0
Cement Wreck (Thuwal)2023-09-08–2024-07-2620.1n=11partial 0/38718.1n=1133.7n=1110.4n=39211.0n=391partial 0/39226.9 ± 9.620.1–33.7
Rose Reef (Thuwal)2023-09-08–2024-10-0424.4n=3partial 0/43820.0n=346.7n=315.8n=45717.6n=45721.0n=68partial 68/45630.7 ± 13.921.0–46.7
Um Mankout (Yanbu)2023-09-1532.7n=2partial 0/22732.9n=267.2n=225.6n=22926.5n=229partial 0/22950.0 ± 24.432.7–67.2
Al Fasmah (Yanbu)2023-09-16—partial 0/116——24.4n=11625.1n=116partial 0/116——
Alrass Alakhdar (Yanbu)2023-09-16—partial 0/57——27.2n=5827.1n=58partial 0/58——
Thetis (Yanbu)2023-10-06–2024-08-0926.8n=5partial 0/16922.3n=523.5n=514.8n=17115.6n=17119.1n=23partial 23/17123.1 ± 3.819.2–26.8
Mr. Umluj (Yanbu)2023-10-07–2024-08-09—16.7n=139partial 139/326——22.0n=32623.1n=326partial 0/326——
Long Reef (Thuwal)2023-11-17—partial 0/88——18.4n=9119.7n=91partial 0/91——
Cave Reef (Thuwal)2024-02-0933.3n=2partial 0/6536.7n=236.7n=211.5n=7112.5n=71partial 0/7135.0 ± 2.433.3–36.7
Abu Galawa (Yanbu)2024-04-19—partial 0/135——22.3n=14024.0n=140partial 0/140——
Chicken Wreck (Jeddah)2024-08-02–1716.1n=125partial 0/17116.0n=12527.4n=12615.1n=21015.7n=210partial 0/15321.7 ± 8.016.1–27.4
Drogon Corals (Jeddah)2024-08-020.0n=1partial 0/650.0n=113.3n=111.9n=6612.4n=66partial 0/666.7 ± 9.40.0–13.3
Anderson's Reef-I (Jeddah)2024-08-0350.0n=1partial 0/14666.7n=176.7n=122.4n=14723.9n=147partial 0/14763.3 ± 18.950.0–76.7
Anderson's Reef-II (Jeddah)2024-08-0342.2n=318.3n=139partial 139/17539.8n=344.4n=318.3n=17419.3n=174partial 0/17443.3 ± 1.542.2–44.4
Coral Gardens (Jeddah)2024-08-10–1735.8n=156partial 0/27128.6n=15644.5n=15619.1n=25420.6n=25436.4n=2partial 2/25438.9 ± 4.935.8–44.5
north Rabigh (shore) (Rabigh)2024-08-1631.2n=175partial 0/15823.5n=17556.5n=17514.8n=12215.9n=122partial 0/12243.8 ± 17.931.2–56.5
unnamed site 1 (Al Qunfudah)2024-08-30—partial 0/123——18.1n=11918.9n=119partial 0/119——
unnamed site 2 (Al Qunfudah)2024-08-30—partial 0/86——13.3n=7014.7n=7027.5n=1partial 1/7027.527.5–27.5
Shark Reef (Thuwal)2024-08-31–09-14—partial 0/352——12.2n=33413.4n=334partial 0/334——
Bhadur Reef (Jeddah)2024-09-07—partial 0/174——16.6n=16816.9n=168partial 0/168——
All sites (unweighted mean)————17.618.6———
Notes for Red Sea - Saudi Arabia 2023-24 (16)
  • [a] No per-site gold standard: the published manual cover (Mayfield 2025, Environments 12:248, Table 4) is by AREA only - Thuwal 16.1% (n=2567); Jeddah 21.6% (n=1160); Rabigh 23.6% (n=173).
  • [b] Anderson's own CoralNet points on the benchmark frames (redsea_claude_arm, 746 frames), (30 − WATER − FISH − DVR), Unk kept (Red Sea paper §2.2), recomputed per frame by score_redsea_claude_20260919.py ('refb').
  • [c] Not built: CoralNet robot labels exist in the Red Sea export, but the Site field is empty on its 207,090 rows, so no site join has been made (build_region_cv_table.py --list).
  • [d] DINOv3 (priority6 fine-tune) point predictions at Anderson's CoralNet points on the benchmark frames (CRD AI/redsea_cv_2026-09/redsea_dinov3_point_predictions.csv); hard_coral / (points − WATER − FISH − DVR by reference label). No full-set DINOv3 run exists (runner written 2026-09-23, never run).
  • [e] THIN SAMPLE (under 30 frames at this site): indicative only.
  • [f] Coralscapes v2 on EDITED frames (data_reports/redsea_cv_20260923/coralscapes_edited_v2_by_pin.json; per-frame outputs on the external drive).
  • [g] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not trusted (Anderson): shown, never averaged. Area fraction, not a point count.
  • [h] Coralscapes v1 on UNEDITED frames (coralscapes_unedited_by_pin.json; the CSV also carries the unedited v2 value). Per-frame outputs on the external drive.
  • [i] 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.
  • [j] CoralSCOP full-set run, EDITED frames (maps/redsea_map_build/coralscop_survey_saudi = the CoralNet-annotated Saudi subset, and coralscop_survey_egypt_edited), IN PROGRESS.
  • [k] PARTIAL - run still in progress; value is on the frames scored so far (n shown; N = frames in the arm/manifest for this site, see header line).
  • [l] Claude benchmark arm (redsea_claude_arm, complete 746/746): blind HC/NOT/UNK calls at Anderson's 30 CoralNet points, same denominator as the manually annotated (CoralNet) cell ('clb'). Frames shown were mostly UNEDITED (645 of 747 index rows); basis per site in the CSV.
  • [m] Claude FULL-SET arm (redsea_full_claude_arm; EDITED, AI-processed where no edit exists; 20 pts, 7-call, (HC+BL)/(20 − WATER − FISH − OBST)), RUNNING: shown, not averaged until complete.
  • 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

Red Sea - Egypt 2024

11 sites · 1,368 photos

Gold standard = published values, by area only (no per-site gold). 159 photos manually scored (via CoralNet) for the benchmark. CoralSCOP ran on both edited and unedited photos (separate columns).

SiteDateClaudeClaude full setManually annotated in CoralNetDINOv3Coralscapes v1Coralscapes v2CoralSCOP uneditedCoralSCOP editedMean ± SD of CVRange
Marsa Nakari-North (Marsa Nakari)2024-09-20—partial 0/172——18.2n=916.9n=18628.4n=20527.1n=142partial 142/18827.7 ± 1.027.0–28.4
Nakari House Reef-South (Marsa Nakari)2024-09-2027.7n=30partial 0/10323.3n=3046.1n=3010.1n=4517.9n=10825.9n=11926.7n=101partial 101/10831.6 ± 9.725.9–46.1
Habili El Ghadeer (Marsa Nakari)2024-09-21—partial 0/147——10.5n=915.4n=14824.7n=154partial 154/168partial 0/14824.724.7–24.7
Om Halhala (Marsa Nakari)2024-09-21—27.4n=145partial 145/206——12.8n=1219.4n=20631.0n=99partial 99/221partial 0/20631.031.0–31.0
Gazella Reef (Wadi Lahami)2024-09-2229.8n=3017.1n=146partial 146/20423.0n=3046.4n=3012.2n=3611.2n=20518.0n=7partial 7/223partial 0/20831.4 ± 14.318.0–46.4
Habili Daisy (Wadi Lahami)2024-09-22—partial 0/50——12.3n=5816.2n=822.7n=49partial 49/58partial 0/822.722.7–22.7
Shabrour-north (Wadi Lahami)2024-09-2227.5n=99partial 0/9224.2n=9945.2n=9916.2n=10118.2n=9417.4n=9partial 9/108partial 0/9530.0 ± 14.117.4–45.2
Shabrour-south (Wadi Lahami)2024-09-22–23—partial 0/221——10.8n=3818.8n=12827.4n=102partial 102/242partial 0/12927.427.4–27.4
Abu Galawa-small (Wadi Lahami)2024-09-23—20.9n=149partial 149/188——15.0n=203—23.1n=203—23.123.1–23.1
Habili Gazella (Wadi Lahami)2024-09-23—17.6n=147partial 147/176——11.2n=191—21.9n=190partial 190/191—21.921.9–21.9
Salah (Wadi Lahami)2024-09-23—24.6n=146partial 146/212——18.7n=231—29.1n=231—29.129.1–29.1
All sites (unweighted mean)————13.4—24.5—24.524.5–24.5
Notes for Red Sea - Egypt 2024 (17)
  • [a] No per-site gold standard: the published manual cover (Mayfield 2025, Environments 12:248, Table 4) is by AREA only - Wadi Lahami 22.1% (n=1180); Marsa Nakari 21.4% (n=674).
  • [b] Not built: CoralNet robot labels exist in the Red Sea export, but the Site field is empty on its 207,090 rows, so no site join has been made (build_region_cv_table.py --list).
  • [c] Coralscapes v1 on UNEDITED frames (coralscapes_unedited_by_pin.json; the CSV also carries the unedited v2 value). Per-frame outputs on the external drive.
  • [d] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not trusted (Anderson): shown, never averaged. Area fraction, not a point count.
  • [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] CoralSCOP full-set run, Egypt 'unedited' group (coralscop_survey_egypt_unedited; for dives 58-62 these bins are the '-AI' processed folders), IN PROGRESS.
  • [g] Claude FULL-SET arm (redsea_full_claude_arm; EDITED, AI-processed where no edit exists; 20 pts, 7-call, (HC+BL)/(20 − WATER − FISH − OBST)), RUNNING: shown, not averaged until complete.
  • [h] PARTIAL - run still in progress; value is on the frames scored so far (n shown; N = frames in the arm/manifest for this site, see header line).
  • [i] Anderson's own CoralNet points on the benchmark frames (redsea_claude_arm, 746 frames), (30 − WATER − FISH − DVR), Unk kept (Red Sea paper §2.2), recomputed per frame by score_redsea_claude_20260919.py ('refb').
  • [j] DINOv3 (priority6 fine-tune) point predictions at Anderson's CoralNet points on the benchmark frames (CRD AI/redsea_cv_2026-09/redsea_dinov3_point_predictions.csv); hard_coral / (points − WATER − FISH − DVR by reference label). No full-set DINOv3 run exists (runner written 2026-09-23, never run).
  • [k] Coralscapes v2 on EDITED frames (data_reports/redsea_cv_20260923/coralscapes_edited_v2_by_pin.json; per-frame outputs on the external drive).
  • [l] THIN SAMPLE (under 30 frames at this site): indicative only.
  • [m] CoralSCOP full-set run, EDITED frames (maps/redsea_map_build/coralscop_survey_saudi = the CoralNet-annotated Saudi subset, and coralscop_survey_egypt_edited), IN PROGRESS.
  • [n] Claude benchmark arm (redsea_claude_arm, complete 746/746): blind HC/NOT/UNK calls at Anderson's 30 CoralNet points, same denominator as the manually annotated (CoralNet) cell ('clb'). Frames shown were mostly UNEDITED (645 of 747 index rows); basis per site in the CSV.
  • 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

Bahrain / Persian Gulf 2023-24

2 sites · 36 photos

No gold standard.

SiteDateCoralSCOPMean ± SD of CVRange
artificial_reef2023-09-2335.7n=2135.735.7–35.7
dilmunia_reef2023-09-2310.6n=1510.610.6–10.6
All sites (unweighted mean)23.123.123.1–23.1
Notes for Bahrain / Persian Gulf 2023-24 (7)
  • [a] No gold standard exists for this region.
  • [b] 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.
  • [c] Bahrain CoralSCOP: maps/bahrain_map_build/coralscop_survey (36 ok frames, 2 folders, both Bahrain 2023 dives; source folders '...-edited'). Bahrain 2024 (2 dives, 151 frames) has no CV output of any kind.
  • [d] 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

Oman 2024

7 sites · 660 photos

Gold standard = published values (Platax 2025, Table 1) from 660 CoralNet-annotated photos. CoralSCOP ran on both edited and unedited photos (separate columns).

SiteDateClaudeGold standardCoralNetDINOv3Coralscapes v1Coralscapes v2CoralSCOP uneditedCoralSCOP editedMean ± SD of CVRange
Cave Site2024-05-2069.9n=1551.4n=17048.7n=10631.2n=15150.3n=17150.3n=17149.0n=14950.5n=14949.9 ± 13.731.2–69.9
Novice Bay2024-05-2042.2n=1538.5n=135—23.0n=12046.8n=13547.5n=13545.9n=11147.3n=11139.6 ± 11.323.0–47.3
Cat Island2024-05-2136.1n=1524.9n=8124.9n=8124.9n=7134.1n=8034.2n=8033.3n=7036.5n=7032.7 ± 5.424.9–36.5
Jissah Point2024-05-2126.3n=1533.3n=6133.3n=6116.0n=5431.9n=6131.5n=6134.9n=5438.1n=5428.8 ± 9.916.0–38.1
Daymaniyat 12024-05-2551.2n=1541.7n=7941.7n=7934.4n=5626.6n=7929.5n=7923.2n=5624.6n=5633.4 ± 12.923.2–51.2
Daymaniyat 22024-05-2549.3n=1538.1n=10638.1n=10646.5n=8439.5n=10637.3n=10642.2n=8150.4n=8147.1 ± 3.742.2–50.4
Daymaniyat 32024-05-2533.8n=1333.0n=2833.0n=2834.2n=1820.3n=2827.0n=2817.6n=1913.3n=1924.7 ± 10.913.3–34.2
All sites (unweighted mean)44.137.3—30.035.636.835.237.236.6 ± 5.830.0–44.1
Notes for Oman 2024 (16)
  • [a] Gold standard = Mayfield AB (2025) Platax 22:27-44, doi:10.29926/platax.202512_22.0003, Table 1 (p. 31), 'Coral cover (%)', printed verbatim; n = # photos in Table 1. Denominator (Methods, p. 32): hard coral relative to the BENTHIC part of the image only - water and fish removed, Unk kept (coral / (30 - WATER - FISH)). The manuscript's own 'Oman photo by photo.csv' reproduces all seven means under this rule (and 37.3 +/- 8.3 across sites, p. 37).
  • [b] Gold standard here is NOT manually annotated in CoralNet: every frame at this site carries CoralNet robot labels (CoralNet export; only Novice Bay 135 + 65 Cave Site frames are Anderson's own). The CoralNet cell is therefore the SAME measurement and is excluded from the CV mean.
  • [c] CoralNet = CoralNet's own classifier labels only (annotator 'robot' in the point export), same denominator as the gold standard.
  • [d] DINOv3: site-level file of record only (oman_cover_by_site_20260807.csv); the per-image output no longer exists, so its denominator is unrecorded and the value cannot be audited. Anderson has retired DINOv3 from future runs. The Cat Island value equalling the gold standard (24.9) is a coincidence - the file predates the corrected gold standard.
  • [e] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt), run 2026-09-18 on UNEDITED originals: hard-coral AREA / WHOLE frame (water, fish, background left in the denominator - different from the gold standard). Not trusted (Anderson); shown, never averaged. Source: CORALSCAPES_V1_V2_UNEDITED_20260918.md.
  • [f] Cat Island Coralscapes n = 80: one annotated frame could not be assigned a site in that run.
  • [g] CoralSCOP = sum of colony-mask AREA fractions per frame (whole frame in the denominator; not a point count - denominator differs from the gold standard). Edited and unedited are PAIRED: only frames scored in both runs.
  • [h] Claude = blind point calls on the EDITED photos (sheet-B display), 30 points per frame, same denominator as the gold standard (Unk in, water/fish out); mean of per-frame cover (the gold standard's aggregation). Only 15 frames per site (13 at Daymaniyat 3): thin. The CSV also carries the pooled value, the gold standard recomputed on the same frames, and the UNEDITED sheet-B fair test re-scored on the gold-standard denominator.
  • [i] Table 1 prints n = 170; 171 frames are on disk and in the per-photo file (reproduces 51.4 exactly).
  • [j] Gold standard partly manually annotated in CoralNet: 65 of the site's frames are Anderson's own points, the other 106 are CoralNet robot labels; the CoralNet cell uses only those 106 robot frames, so it is a real (partly independent) comparison.
  • [k] Table 1 prints the Daymaniyat 3 SD as '4.' (typo). The per-photo file gives 23.7 (23.75); earlier tables carried 23.7 as if published.
  • [l] Blank: every Novice Bay frame was manually annotated in CoralNet by Anderson, so CoralNet produced no robot label there. The old 38.6 (n=114) has no traceable source.
  • [m] Novice Bay CoralSCOP changed from 38.2/41.0 (66 paired frames, 2026-08-24, edited run then incomplete) to 45.9/47.3 (111 paired frames, finished edited run).
  • 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

Kenya 2024

20 sites · 2,317 photos

Gold standard = published values (Environments 12:276, Appendix A) from 2,300 CoralNet-annotated photos. 703 photos also manually scored (via CoralNet).

SiteDateClaudeGold standardManually annotated in CoralNetCoralNetDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
Sandy Patch/Point (Mombasa)2024-03-119.2n=829.3n=11710.8n=828.8n=11740.0n=11737.6n=11737.2n=11725.9n=11025.0 ± 15.49.2–40.0
Brain Coral (Mombasa)2024-03-11—27.4n=136—27.4n=13642.7n=13737.7n=13738.9n=13723.7n=12533.2 ± 13.423.7–42.7
MMP snorkel site (Mombasa)2024-03-14—39.6n=54—39.6n=5454.2n=5433.4n=5435.9n=5423.3n=4138.8 ± 21.923.3–54.2
Diani Beach (Diani)2024-03-16—34.9n=50—34.2n=5038.7n=5329.9n=5330.2n=5322.3n=3530.5 ± 11.622.3–38.7
Reefolution nursery (Shimoni-Wasini)2024-03-20—40.2n=258—39.6n=25843.8n=26432.0n=26433.2n=26416.3n=14130.1 ± 19.416.3–43.8
Coral Garden (Kisite National Park)2024-03-2143.8n=2755.0n=17851.4n=2746.4n=17846.0n=17944.3n=17943.7n=17920.9n=15036.9 ± 13.920.9–46.0
Outer Kisite (Kisite National Park)2024-03-21—36.6n=85—36.0n=8342.6n=8344.8n=8343.7n=8317.3n=6829.9 ± 17.917.3–42.6
Dolphin Point (Kisite National Park)2024-03-21—39.9n=108—39.8n=10842.5n=10835.2n=10834.4n=10827.0n=9634.8 ± 11.027.0–42.5
Kisite National Park-site 4 (Kisite National Park)2024-03-21—45.5n=71—45.5n=7136.1n=7133.0n=7134.0n=7117.9n=5027.0 ± 12.917.9–36.1
Lower Mpunguti-site 1 (Kisite National Park)2024-03-2116.7n=118.1n=3620.0n=117.1n=3637.3n=3617.2n=3616.2n=3620.5n=2724.8 ± 11.016.7–37.3
Lower Mpunguti-site 2 (Kisite National Park)2024-03-21—34.2n=27—34.2n=2742.6n=2718.9n=2717.4n=2716.9n=2729.8 ± 18.216.9–42.6
Kilifi Bay (Kilifi)2024-03-26—25.1n=134—25.6n=13431.4n=13622.3n=13621.4n=13618.5n=12324.9 ± 9.118.5–31.4
Bofa (Kilifi)2024-03-26—32.4n=97—34.0n=9747.6n=10429.5n=10430.4n=10426.6n=8737.1 ± 14.826.6–47.6
Round Reef (Watamu)2024-03-2712.8n=17517.8n=17515.6n=17513.7n=17527.7n=17511.8n=17514.4n=17515.7n=1618.7 ± 7.912.8–27.7
Uyombo (Watamu)2024-03-2823.0n=37424.8n=37522.8n=37424.5n=37536.7n=37419.3n=37421.9n=37429.1n=2429.6 ± 6.923.0–36.7
North Canyon (Watamu)2024-03-29—21.6n=83—21.6n=8320.3n=8215.3n=8214.7n=8219.1n=7919.7 ± 0.819.1–20.3
Trumpet Area (Watamu)2024-03-29—20.8n=193—20.8n=19335.2n=19321.3n=19322.3n=19320.7n=17627.9 ± 10.220.7–35.2
Chipopo (Pate (Lamu))2024-04-037.2n=4012.8n=408.6n=4017.3n=4021.5n=406.5n=407.3n=4011.8n=3513.5 ± 7.37.2–21.5
Pilot Nursery (Pate (Lamu))2024-04-0314.1n=2114.0n=2112.4n=2119.0n=2132.6n=2116.2n=2118.5n=2111.2n=1619.3 ± 11.611.2–32.6
Ocean Trust Nursery (Pate (Lamu))2024-04-0323.3n=5728.9n=6222.9n=5728.1n=6240.3n=6319.8n=6318.1n=6310.9n=4124.8 ± 14.810.9–40.3
All sites (unweighted mean)—29.9—29.738.026.326.719.828.9 ± 12.919.8–38.0
Notes for Kenya 2024 (16)
  • [a] Gold = published cover, Mayfield 2025 Environments 12:276, Appendix A (CoralNet point annotation; denominator coral / (30 − WATER − FISH), Unk kept); n = photos in the published set. NB the CoralNet export holds 22,016 abmayfield points and 47,644 robot points.
  • [b] Anderson's own labels (abmayfield only) on exactly the Claude frames, all-30 (like-for-like with the Claude cell).
  • [c] CoralNet = the source's Machine suggestion 1 on ALL unedited frames. The classifier was trained on Anderson's own labels of these same images, so it is CIRCULAR with the gold standard (cv_method_tables.json: 'CIRCULAR, not an independent check').
  • [d] Same measurement as (or circular with) the reference: shown, not averaged.
  • [e] DINOv3 on ALL unedited photos at the site, coral / (30 − WATER − FISH), Unk kept (pin coral_cover.methods; source kenya_1450_dinov3 + pins).
  • [f] Coralscapes run 2026-09-18 on UNEDITED originals; per-frame outputs are on the external drive (/Volumes/OneDrive/.../Kenya_unedited_20260918, ..._coralscapesv2_20260918; logs/coralscapes*_unedited_20260918.log); site values read from the pins (not re-verified).
  • [g] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not trusted (Anderson): shown, never averaged. Area fraction, not a point count.
  • [h] 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.
  • [i] CoralSCOP full run on unedited originals (coral_size_coralscop/kenya_originals_2026-09/coralscop_full, 1,444 ok of 1,450); site values read from the pins.
  • [j] Not applicable by ruling: Kenya edited files are CROPPED; annotations are on the originals (Anderson 2026-09-17: 'for kenya, it's only on unedited').
  • [k] Claude = blind point calls, 30 CoralNet points per frame, on UNEDITED originals (verified per batch in KENYA_CLAUDE_ARM_20260917.md), cover = H/30 (all-30, Unk in the denominator - the arm's canonical make_fig3_methods.score() rule, NOT the gold standard's denominator); batches 1-8, batch 9 caller C only, batch 10 (103 of 109 called); recomputed here from the predictions_kenya_batch*.json files.
  • [l] THIN SAMPLE (under 30 frames at this site): indicative only.
  • [m] GENUINELY NOT RUN at this site: none of the 703 Claude-scored Kenya frames is from here (Claude batches 1-10 cover 8 of 20 sites).
  • 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

East Timor (Atauro) 2024

7 sites · 990 photos

No gold standard.

SiteDateClaudeDINOv3CoralSCOPMean ± SD of CVRange
Raiketa2024-01-2946.4n=20039.9n=20044.4n=20043.6 ± 3.339.9–46.4
Inner Reef2024-01-3044.6n=19338.4n=19352.6n=19345.2 ± 7.138.4–52.6
Haruine2024-01-30–3141.1n=18335.6n=18338.7n=18338.5 ± 2.835.6–41.1
Biqueli Wall2024-01-3027.4n=14730.2n=14738.4n=14732.0 ± 5.727.4–38.4
Compass Reef2024-01-31–02-0138.0n=9930.5n=9948.9n=9939.1 ± 9.230.5–48.9
Cabbage Patch2024-01-3137.2n=11530.7n=11537.6n=11535.2 ± 3.930.7–37.6
Pinnacle2024-01-3119.7n=5326.1n=5338.6n=5328.1 ± 9.619.7–38.6
All sites (unweighted mean)36.333.142.737.4 ± 4.933.1–42.7
Notes for East Timor (Atauro) 2024 (12)
  • [a] No gold standard exists (no transect, no frames manually annotated in CoralNet).
  • [b] Not applicable: no CoralNet source / hand annotations exist for these photos.
  • [c] Value read from cv_method_tables.json (injector output, generated 2026-09-28 22:33); the scope text there gives denominator and frame set.
  • [d] 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).
  • [e] GENUINELY NOT RUN: no Coralscapes v2 output for Atauro (backlog 0/6; not in _from_external_20260925).
  • [f] GENUINELY NOT RUN: no Coralscapes v1 output for Atauro.
  • [g] Not run: EDITED region by ruling; unedited originals exist but were never staged for CoralSCOP.
  • [h] 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.
  • [i] Claude = blind 7-call point count on the EDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), UNK kept in; mean across photos (arm ledgers, via cv_method_tables.json).
  • 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

Komodo 2019

17 sites · 1,381 photos

No gold standard.

SiteDateClaudeDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
The Cauldron / Shotgun2009-08-14–2019-01-2126.0n=6220.9n=6250.8n=6250.8n=6233.7n=6226.9 ± 6.420.9–33.7
Tatawa Besar2009-08-16–2019-01-2722.0n=15523.9n=15552.6n=15554.6n=15534.2n=15526.7 ± 6.622.0–34.2
Siaba Kecil2009-08-16–2019-01-2625.4n=7230.0n=7263.8n=7263.7n=7239.3n=7231.6 ± 7.125.4–39.3
Pengah Kecil2009-08-18–2019-01-2529.9n=3730.5n=3761.8n=3760.2n=3736.2n=3732.2 ± 3.529.9–36.2
Siaba Besar2019-01-19–2652.9n=12131.6n=12164.0n=12165.8n=12141.5n=12142.0 ± 10.731.6–52.9
Mawan2019-01-19–2614.2n=15020.3n=15049.4n=14852.9n=14929.9n=14821.5 ± 7.914.2–29.9
Hot Rocks (Sangeang)2019-01-209.1n=12025.8n=12056.5n=12054.4n=12038.1n=12024.3 ± 14.69.1–38.1
Deep Purple2019-01-2026.4n=6029.6n=6058.2n=6054.6n=6039.6n=6031.9 ± 6.926.4–39.6
Makassar Reef / Manta Point2019-01-223.3n=5910.1n=5933.1n=5231.3n=489.1n=577.5 ± 3.73.3–10.1
Bonsai Rock2019-01-2217.5n=6626.5n=6652.1n=6653.2n=6637.7n=6627.2 ± 10.117.5–37.7
Padar Island (Padar Kecil)2019-01-238.1n=5624.0n=5651.3n=5651.1n=5645.8n=5626.0 ± 18.98.1–45.8
Secret Garden2019-01-2312.8n=3924.9n=3951.1n=3955.1n=3938.0n=3925.2 ± 12.612.8–38.0
Three Sisters (Tiga Dara)2019-01-238.3n=3419.7n=3428.6n=3429.7n=3428.5n=3418.8 ± 10.18.3–28.5
Pink Beach (Pantai Merah)2019-01-2424.8n=7130.0n=7166.1n=7166.8n=7137.3n=7130.7 ± 6.324.8–37.3
Loh Sera2019-01-2429.7n=11834.6n=11872.6n=11873.3n=11845.5n=11736.6 ± 8.129.7–45.5
Police Corner2019-01-2521.6n=7830.0n=7865.3n=7867.9n=7851.8n=7834.5 ± 15.621.6–51.8
Tatawa Kecil2019-01-2732.1n=8328.8n=8356.6n=8356.8n=8337.9n=8332.9 ± 4.628.8–37.9
All sites (unweighted mean)21.425.954.955.436.728.0 ± 7.921.4–36.7
Notes for Komodo 2019 (11)
  • [a] No gold standard exists (no transect, no frames manually annotated in CoralNet).
  • [b] Not applicable: no CoralNet source / hand annotations exist for these photos.
  • [c] Value read from cv_method_tables.json (injector output, generated 2026-09-28 22:33); the scope text there gives denominator and frame set.
  • [d] 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).
  • [e] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not trusted (Anderson): shown, never averaged. Area fraction, not a point count.
  • [f] Not run: EDITED region by ruling; unedited originals exist but were never staged for CoralSCOP.
  • [g] 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.
  • [h] Claude = blind 7-call point count on the EDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), UNK kept in; mean across photos (arm ledgers, via cv_method_tables.json).
  • 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

Raja Ampat 2018 (incl. Triton Bay)

16 sites · 1,094 photos

No gold standard.

SiteDateClaudeDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
Seruanus (Triton Bay)2018-10-149.1n=6618.7n=6643.7n=6642.0n=6630.6n=6619.5 ± 10.89.1–30.6
Little Komodo (Triton Bay)2018-10-1411.7n=7322.9n=7543.6n=7344.7n=7231.2n=7521.9 ± 9.811.7–31.2
Last Frontier (Triton Bay)2018-10-157.8n=1410.2n=1516.9n=1414.7n=1416.0n=1511.3 ± 4.27.8–16.0
Bo Rainbow (Triton Bay)2018-10-1512.7n=4722.0n=4759.6n=4761.6n=4742.1n=4725.6 ± 15.012.7–42.1
Karang Paradise2018-10-1644.4n=13228.7n=13666.4n=12767.5n=12731.4n=13434.8 ± 8.428.7–44.4
Pulau Koon dive 22018-10-1724.7n=5219.4n=5348.8n=5146.7n=5133.0n=5325.7 ± 6.819.4–33.0
Pulau Koon dive 32018-10-1729.6n=9831.1n=9875.4n=9876.0n=9846.9n=9835.9 ± 9.629.6–46.9
Shadow Reef2018-10-18–2016.9n=8517.6n=10744.1n=8346.9n=8118.2n=10617.6 ± 0.716.9–18.2
Nudi Rock2018-10-1929.7n=9725.9n=10061.4n=9861.9n=9832.4n=10029.3 ± 3.325.9–32.4
Little Juliet2018-10-1926.2n=1323.8n=1348.1n=1251.2n=1227.8n=1325.9 ± 2.023.8–27.8
Anchovy2018-10-2032.4n=6425.8n=6455.1n=6455.1n=6438.0n=6432.1 ± 6.125.8–38.0
Two Tree Island2018-10-207.2n=7718.6n=8831.2n=8231.0n=7518.8n=8814.9 ± 6.67.2–18.8
My Reef2018-10-2133.2n=5431.8n=5455.2n=5451.6n=5439.2n=5434.7 ± 3.931.8–39.2
Melissa's Garden (Fam/Penemu)2018-10-2140.9n=8935.1n=8971.3n=8970.0n=8940.7n=8938.9 ± 3.335.1–40.9
Blue Magic2018-10-22–2311.1n=4019.3n=4229.0n=4035.9n=4027.4n=4219.3 ± 8.211.1–27.4
Mioskon2018-10-2211.9n=4723.8n=4743.4n=4745.0n=4732.7n=4722.8 ± 10.411.9–32.7
All sites (unweighted mean)21.823.449.650.131.625.6 ± 5.321.8–31.6
Notes for Raja Ampat 2018 (incl. Triton Bay) (11)
  • [a] No gold standard exists (no transect, no frames manually annotated in CoralNet).
  • [b] Not applicable: no CoralNet source / hand annotations exist for these photos.
  • [c] Value read from cv_method_tables.json (injector output, generated 2026-09-28 22:33); the scope text there gives denominator and frame set.
  • [d] 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).
  • [e] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not trusted (Anderson): shown, never averaged. Area fraction, not a point count.
  • [f] Not run: EDITED region by ruling; unedited originals exist but were never staged for CoralSCOP.
  • [g] 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.
  • [h] Claude = blind 7-call point count on the EDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), UNK kept in; mean across photos (arm ledgers, via cv_method_tables.json).
  • 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

Maluku 2018 - Banda

7 sites · 885 photos

No gold standard.

SiteDateClaudeDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
Lighthouse2018-04-25–2628.9n=13323.8n=5636.5n=5638.5n=5636.3n=13229.7 ± 6.323.8–36.3
Lava Flow (Gunung Api)2018-04-2554.6n=19035.3n=19075.6n=19076.5n=19037.0n=18942.3 ± 10.735.3–54.6
Tanjung Pisang (Banana Island)2018-04-2536.4n=18030.9n=18065.2n=18067.3n=18039.8n=18035.7 ± 4.530.9–39.8
Airport (Banda Ara)2018-04-2638.2n=7235.6n=7271.2n=7272.7n=7245.3n=7239.7 ± 5.035.6–45.3
Goat Hall (Pulau Ai)2018-04-2736.5n=14432.5n=14461.6n=14463.2n=14436.9n=14435.3 ± 2.432.5–36.9
Shark Point (Pulau Ai)2018-04-2733.4n=8830.1n=8865.6n=8863.7n=8834.0n=8832.5 ± 2.130.1–34.0
Batu Kapal2018-04-2816.4n=7824.3n=7840.8n=7843.2n=7826.4n=7822.4 ± 5.316.4–26.4
All sites (unweighted mean)34.930.459.560.736.533.9 ± 3.230.4–36.5
Notes for Maluku 2018 - Banda (11)
  • [a] No gold standard exists (no transect, no frames manually annotated in CoralNet).
  • [b] Not applicable: no CoralNet source / hand annotations exist for these photos.
  • [c] Value read from cv_method_tables.json (injector output, generated 2026-09-28 22:33); the scope text there gives denominator and frame set.
  • [d] 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).
  • [e] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not trusted (Anderson): shown, never averaged. Area fraction, not a point count.
  • [f] Not run: EDITED region by ruling; unedited originals exist but were never staged for CoralSCOP.
  • [g] 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.
  • [h] Claude = blind 7-call point count on the EDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), UNK kept in; mean across photos (arm ledgers, via cv_method_tables.json).
  • 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

Maldives #2 2017

24 sites · 919 photos

No gold standard.

SiteDateDINOv3CoralSCOPMean ± SD of CVRange
checkout [Checkout dive (Hulhumale outer reef)]2017-05-1315.4n=1022.1n=1018.8 ± 4.715.4–22.1
back_faru [Back Faru]2017-05-14–2113.9n=7114.6n=2014.2 ± 0.413.9–14.6
bodu_hithi_thila [Bodu Hithi Thila]2017-05-14–1514.8n=5216.5n=2015.6 ± 1.214.8–16.5
feydoo_wall [Feydoo Wall]2017-05-1523.6n=1320.5n=1322.1 ± 2.220.5–23.6
georgias_reef [Georgia's Reef]2017-05-1516.8n=3323.1n=2019.9 ± 4.516.7–23.2
miyaro_faru [Miyaro Faru]2017-05-1617.0n=1823.4n=1820.2 ± 4.517.0–23.4
vaadhoo_caves [Vaadhoo Caves]2017-05-167.2n=2114.4n=2010.8 ± 5.17.3–14.4
veligandu [Veligandu]2017-05-1613.0n=1722.4n=1717.7 ± 6.713.0–22.4
ghuli_faru [Ghuli Faru]2017-05-1710.3n=2412.7n=1811.5 ± 1.710.3–12.7
miyaru_bay [Miyaru Bay]2017-05-17–1815.1n=2115.0n=2015.1 ± 0.015.0–15.1
big_sand_bank_2017 [Big Sand Bank]2017-05-1815.8n=2319.6n=2017.7 ± 2.715.8–19.6
kudagiri_wreck [Kuda Giri Wreck]2017-05-18–2218.2n=6321.0n=2019.6 ± 2.018.2–21.0
hodeli_bodu_giri [Hodeli Bodu Giri]2017-05-1921.0n=2825.4n=2023.2 ± 3.121.0–25.4
kandooma_thila [Kandooma Thila]2017-05-19–2213.2n=507.3n=1910.3 ± 4.27.3–13.2
fish_tank [Fish Tank]2017-05-2111.2n=194.9n=178.1 ± 4.44.9–11.2
lankan_faru [Lankan Faru]2017-05-21–229.2n=8522.3n=1815.8 ± 9.39.2–22.3
camel_point [Camel Point]2017-05-2322.9n=7425.5n=2024.2 ± 1.822.9–25.5
lohosfushi_corner [Lohosfushi Corner]2017-05-2314.2n=2516.3n=2015.2 ± 1.514.1–16.3
ari_south_tip [Ari Atoll south tip]2017-05-2418.5n=8418.7n=2018.6 ± 0.118.5–18.7
dhigga_thila [Dhigga Thila]2017-05-2421.6n=2428.6n=2025.1 ± 5.021.6–28.5
dhega_thila [Dhega Thila]2017-05-2525.2n=6120.8n=2023.0 ± 3.220.8–25.2
fesdhoo_wreck [Fesdhoo wreck]2017-05-2515.1n=3213.8n=2014.4 ± 0.913.8–15.1
moofushi_kandu [Moofushi Kandu]2017-05-2520.8n=3410.9n=2015.8 ± 7.110.9–20.8
fish_house [Fish House (Fish Head)]2017-05-2612.5n=376.3n=209.4 ± 4.46.3–12.5
All sites (unweighted mean)16.117.816.9 ± 1.216.1–17.8
Notes for Maldives #2 2017 (12)
  • [a] No gold standard (no transect or frames manually annotated in CoralNet for the 2017 trip).
  • [b] Not applicable: no CoralNet source / hand annotations of Anderson's Maldives photos exist (wide search 2026-09-29; the only Maldives CoralNet source found is an outside project).
  • [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 the Maldives on disk or recorded on the external drive (wide search 2026-09-29).
  • [e] Not applicable: both Maldives trips are EDITED sets (EDITED_UNEDITED_RULES.md); no unedited originals staged.
  • [f] 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.
  • [g] CoralSCOP ran on a staged subset (maps/maldives_map_build/coralscop_survey: 878 ok frames over 47 folders, up to ~20 frames per folder at the larger sites), not every photo.
  • [h] THIN SAMPLE (under 30 frames at this site): indicative only.
  • [i] GENUINELY NOT RUN: no Claude arm for the Maldives (model_comparison_audits/ has none).
  • 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

Philippines 2017 (Palawan)

6 sites · 396 photos

No gold standard.

SiteDateDINOv3CoralSCOPMean ± SD of CVRange
reef_1_snorkel2017-01-2029.1n=847.8n=838.5 ± 13.229.1–47.8
bacuit_archipelago2017-01-2419.7n=30515.0n=2017.3 ± 3.315.0–19.7
akikushima_1_27_172017-01-2719.9n=1216.3n=1118.1 ± 2.616.3–19.9
coral_garden_1_27_172017-01-2735.1n=2525.6n=2030.4 ± 6.725.6–35.1
taiei_maru_1_27_172017-01-2711.6n=1717.8n=1714.7 ± 4.411.6–17.8
shipwreck_dive_1—20.9n=2937.7n=1929.3 ± 11.920.9–37.7
All sites (unweighted mean)22.726.724.7 ± 2.822.7–26.7
Notes for Philippines 2017 (Palawan) (10)
  • [a] No gold standard exists for this region.
  • [b] Not applicable: no CoralNet annotations of these photos.
  • [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] THIN SAMPLE (under 30 frames at this site): indicative only.
  • [e] GENUINELY NOT RUN: no Coralscapes output for this trip.
  • [f] 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.
  • [g] GENUINELY NOT RUN: no Claude arm for this trip.
  • 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

Philippines 2016 (Anilao)

7 sites · 609 photos

No gold standard.

SiteDateDINOv3CoralSCOPMean ± SD of CVRange
dive_1_lipgpo_island2016-09-3021.0n=9434.2n=9427.6 ± 9.421.0–34.2
dive_2_dive_and_trek2016-09-3027.3n=8931.1n=8929.2 ± 2.627.3–31.0
dive_3_monte_carlo_hotel2016-09-3022.3n=8222.0n=8222.1 ± 0.222.0–22.3
fundiveanilaoanderson_sept30_oct12016-09-30–10-0123.1n=14321.6n=3022.4 ± 1.121.6–23.1
dive_5_kirby_s_rock2016-10-0123.3n=7327.7n=7325.5 ± 3.123.3–27.7
dive_6_beatriz_island2016-10-0118.1n=7619.7n=7618.9 ± 1.118.1–19.7
dive_7_sombrero_island2016-10-0125.1n=5227.2n=5226.2 ± 1.525.1–27.2
All sites (unweighted mean)22.926.224.6 ± 2.422.9–26.2
Notes for Philippines 2016 (Anilao) (9)
  • [a] No gold standard exists for this region.
  • [b] Not applicable: no CoralNet annotations of these photos.
  • [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 output for this trip.
  • [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] GENUINELY NOT RUN: no Claude arm for this trip.
  • 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

Maldives #1 2016

18 sites · 2,934 photos

No gold standard.

SiteDateDINOv3CoralSCOPMean ± SD of CVRange
mal01_small_sand_bank [Small Sand Bank]2016-01-14–1923.1n=32717.1n=1620.1 ± 4.217.1–23.1
mal02_raaeundi [Raaeundi Reef]2016-01-1524.2n=29119.3n=2021.8 ± 3.419.3–24.2
mal03_veli [Veli]2016-01-1630.8n=20038.6n=2034.7 ± 5.530.8–38.6
big_sand_bank_2016 [Big Sand Bank]2016-01-1723.6n=5633.8n=2028.7 ± 7.223.6–33.8
mal04_stage [Stage]2016-01-1723.0n=20837.4n=2030.2 ± 10.123.0–37.3
anantara_dhigu [Anantara Dhigu house reef]2016-01-1833.3n=8132.1n=2032.7 ± 0.832.1–33.3
hr_dhigu_0118 [Anantara Dhigu house reef]2016-01-1824.4n=4436.2n=2030.3 ± 8.424.4–36.2
mal05_maafushi [Maafushi]2016-01-1821.2n=21040.5n=2030.9 ± 13.621.2–40.5
anantara_kihava [Anantara Kihavah house reef]2016-01-21–2627.5n=13635.2n=1931.3 ± 5.527.5–35.2
mal06_baa1 [Gaagandufaruhuraa]2016-01-2126.7n=24344.5n=1935.6 ± 12.626.7–44.5
hr_drift_0122 [Anantara Kihavah house reef]2016-01-2214.7n=4519.3n=2017.0 ± 3.214.7–19.3
mal07_baa2 [Kashidhoo Faru]2016-01-22–2325.8n=30826.2n=1926.0 ± 0.325.8–26.2
four_seasons_0124 [Four Seasons Landaa Giraavaru]2016-01-2429.5n=14644.7n=2037.1 ± 10.729.5–44.7
aquarium_restaurant_0125 [Anantara Kihavah house reef]2016-01-2519.0n=3524.9n=2021.9 ± 4.119.0–24.9
golden_wall_0125 [Golden Wall]2016-01-2513.8n=1727.2n=1720.5 ± 9.413.8–27.2
mal08_baa3 [Baa Atoll (Hanifaru)]—21.0n=29025.1n=1923.1 ± 2.921.0–25.1
mal09_baa4 [Baa Atoll (Hanifaru)]—18.0n=8925.2n=2021.6 ± 5.118.0–25.2
mal10_baa5 [Baa Atoll (Hanifaru)]—20.7n=20830.0n=2025.3 ± 6.620.7–30.0
All sites (unweighted mean)23.431.027.1 ± 5.423.3–31.0
Notes for Maldives #1 2016 (12)
  • [a] No frame- or site-matched gold standard: the 2016 bleaching sheets (maps/maldives_map_build/bleaching_stats_2016.json) are April/July 2016 transects at other Baa/Raa/Gaafu Alifu sites, not these January 2016 dives (backlog: 'partial: no Jan-2016 benthic transect data found').
  • [b] Not applicable: no CoralNet source / hand annotations of Anderson's Maldives photos exist (wide search 2026-09-29; the only Maldives CoralNet source found is an outside project).
  • [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 the Maldives on disk or recorded on the external drive (wide search 2026-09-29).
  • [e] Not applicable: both Maldives trips are EDITED sets (EDITED_UNEDITED_RULES.md); no unedited originals staged.
  • [f] 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.
  • [g] CoralSCOP ran on a staged subset (maps/maldives_map_build/coralscop_survey: 878 ok frames over 47 folders, up to ~20 frames per folder at the larger sites), not every photo.
  • [h] THIN SAMPLE (under 30 frames at this site): indicative only.
  • [i] GENUINELY NOT RUN: no Claude arm for the Maldives (model_comparison_audits/ has none).
  • 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

Chagos #2 2015

49 sites · 1,176 photos

Gold standard = LOF diver transects (no photos). 480 photos manually scored (via CoralNet).

SiteDateClaudeGold standardManually annotated in CoralNetCoralNetDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
BIPB602015-04-1429.5n=3842.238.8n=38—33.4n=3855.5n=3857.2n=3841.9n=3834.9 ± 6.329.5–41.9
BIPB612015-04-1417.0n=1420.017.9n=14—22.1n=1441.0n=1443.0n=1432.5n=1423.9 ± 7.917.0–32.5
BIPB622015-04-1433.6n=1959.648.0n=19—35.3n=1957.6n=1857.5n=1831.6n=1933.5 ± 1.931.6–35.3
BIPB632015-04-1535.0n=2956.046.9n=29—33.3n=2966.2n=2966.7n=2938.4n=2935.6 ± 2.633.3–38.4
BIPB642015-04-1522.8n=2131.222.7n=21—26.2n=2134.3n=2141.4n=2133.3n=2127.4 ± 5.422.8–33.3
BIPB652015-04-1530.3n=1941.432.9n=19—32.3n=1952.3n=1954.9n=1927.5n=1930.0 ± 2.427.5–32.3
BIPB662015-04-162.3n=67.012.5n=6—26.1n=641.9n=640.9n=628.8n=619.1 ± 14.62.3–28.8
BIPB672015-04-1626.1n=4141.936.4n=41—33.8n=4149.3n=4152.2n=4129.5n=4129.8 ± 3.926.1–33.8
BIPB682015-04-1636.8n=4355.044.9n=43—38.5n=4371.9n=4373.8n=4339.2n=4338.2 ± 1.236.8–39.2
BIPB692015-04-16–05-0239.1n=11664.754.2n=11056.1n=647.4n=11674.4n=11675.4n=11642.6n=11646.3 ± 7.439.1–56.1
BIPB702015-04-17–05-0134.9n=5560.345.4n=2845.1n=2740.2n=5556.6n=5559.5n=5535.7n=5539.0 ± 4.734.9–45.1
BIPB722015-04-18–05-0144.3n=7572.450.5n=4156.1n=3449.2n=7564.4n=7569.0n=7539.9n=7547.4 ± 7.039.9–56.1
BIPB732015-04-1837.6n=4550.1—45.2n=4545.3n=4548.0n=4550.6n=4546.9n=4543.8 ± 4.237.6–46.9
BIPB742015-04-18–05-0151.1n=2162.439.3n=856.0n=1348.9n=2164.5n=2165.9n=2147.1n=2150.8 ± 3.947.1–56.0
BIPB752015-04-1932.4n=943.9—39.2n=942.2n=951.0n=956.5n=943.7n=939.4 ± 5.032.4–43.7
BIPB762015-04-1920.9n=2035.9—29.9n=2031.5n=2034.7n=2039.8n=2029.0n=2027.8 ± 4.720.9–31.5
BINI772015-04-2025.7n=1143.0—29.8n=1126.1n=1138.2n=1141.1n=1133.5n=1128.8 ± 3.625.7–33.5
BINI782015-04-2024.8n=1619.1—24.4n=1629.0n=1626.9n=1628.6n=1628.4n=1626.6 ± 2.424.4–29.0
BINI792015-04-2017.9n=1531.5—19.4n=1521.6n=1527.2n=1525.0n=1521.2n=1520.0 ± 1.717.9–21.6
BINI802015-04-2129.4n=2442.837.6n=1026.8n=1436.9n=2442.4n=2445.9n=2434.9n=2432.0 ± 4.726.8–36.9
BIVB812015-04-2156.7n=551.7—37.8n=544.7n=560.9n=563.4n=549.2n=547.1 ± 7.937.8–56.7
BIBL822015-04-2228.3n=2237.917.5n=427.6n=1826.1n=2242.0n=2250.2n=2235.8n=2229.4 ± 4.326.1–35.8
BIBL832015-04-2223.8n=1533.120.8n=925.6n=624.7n=1542.2n=1545.2n=1540.7n=1528.7 ± 8.023.8–40.7
BISA842015-04-2217.5n=748.825.8n=216.7n=528.6n=735.3n=734.3n=740.1n=725.7 ± 11.016.7–40.1
BIBL862015-04-2334.4n=2747.2—31.9n=2736.0n=2750.5n=2749.4n=2725.1n=2731.9 ± 4.825.1–36.0
BIBL872015-04-2330.3n=3540.6—27.9n=3534.0n=3541.5n=3548.0n=3529.5n=3530.4 ± 2.627.9–34.0
BISA882015-04-2342.7n=1356.7—45.8n=1354.1n=1355.3n=1359.9n=1343.4n=1346.5 ± 5.242.7–54.1
BISA912015-04-2440.3n=1352.4—46.7n=1355.6n=1358.0n=1364.8n=1344.0n=1346.6 ± 6.540.3–55.6
BISP892015-04-2419.9n=1333.2—17.2n=1321.5n=1328.6n=1329.9n=1327.6n=1321.6 ± 4.417.2–27.6
BISP902015-04-2424.7n=839.8—27.6n=830.8n=841.6n=843.2n=837.9n=830.2 ± 5.724.7–37.9
BISA922015-04-2518.8n=3924.5—29.4n=3925.6n=3933.1n=3934.3n=3929.0n=3925.7 ± 4.918.8–29.4
BISA932015-04-2533.0n=2739.1—38.7n=2726.4n=2858.3n=2661.0n=2726.4n=2831.1 ± 5.926.4–38.7
BISA942015-04-2523.1n=240.3—47.4n=256.7n=243.5n=263.2n=221.4n=237.1 ± 17.621.4–56.7
BISA972015-04-2651.1n=1254.4—51.1n=1249.7n=1267.8n=1269.3n=1229.4n=1245.3 ± 10.629.4–51.1
BISP952015-04-2617.8n=2656.9—21.9n=2620.4n=2628.0n=2627.6n=2623.1n=2520.8 ± 2.317.8–23.1
BISP962015-04-2651.5n=1886.0—55.6n=1849.3n=1869.0n=1872.7n=1848.2n=1851.1 ± 3.348.2–55.6
BIVB992015-04-2735.4n=1155.6—20.8n=1124.5n=1164.2n=1162.9n=1118.5n=1124.8 ± 7.518.5–35.4
BISA1022015-04-2821.5n=2730.9—31.1n=2730.1n=2733.2n=2734.7n=2734.4n=2729.3 ± 5.521.5–34.4
BISP1002015-04-2831.9n=753.9—21.7n=734.8n=742.2n=743.2n=727.0n=728.9 ± 5.821.7–34.8
BISP1012015-04-2823.0n=1350.0—39.1n=1332.1n=1342.2n=1343.5n=1332.6n=1331.7 ± 6.623.0–39.1
BISA1032015-04-2930.1n=1932.5—28.4n=1937.2n=1940.5n=1944.5n=1936.1n=1933.0 ± 4.328.4–37.2
BISA1042015-04-2935.7n=1636.2—44.2n=1645.6n=1646.0n=1650.6n=1641.0n=1641.6 ± 4.435.7–45.6
BISA1052015-04-2927.8n=1950.3—23.8n=1936.0n=1941.8n=1947.0n=1923.4n=1927.8 ± 5.823.4–36.0
BIPB1102015-04-3029.5n=4539.3—36.6n=4536.6n=4544.3n=4349.6n=4336.7n=4434.9 ± 3.629.5–36.7
BISA1072015-04-3018.5n=1224.6—18.0n=1220.6n=1232.1n=1241.2n=1221.7n=1219.7 ± 1.818.0–21.7
BIVB1082015-04-3011.7n=1145.740.0n=125.1n=1023.9n=1123.2n=1018.1n=1128.2n=1122.2 ± 7.211.7–28.2
BIVB1092015-04-3035.0n=3334.736.7n=135.0n=3239.0n=3367.3n=3366.1n=3324.9n=3333.5 ± 6.024.9–39.0
BIPB1122015-05-0228.7n=1445.043.6n=14—44.3n=1450.6n=1455.0n=1437.7n=1436.9 ± 7.828.7–44.3
BIPB1142015-05-02–0320.0n=2939.629.0n=2234.4n=731.7n=2937.0n=2938.5n=2935.2n=2930.3 ± 7.020.0–35.2
All sites (unweighted mean)29.744.1——35.147.350.133.832.8 ± 2.829.7–35.1
Notes for Chagos #2 2015 (12)
  • [a] Value read from cv_method_tables.json (injector output, generated 2026-09-28 22:33); the scope text there gives denominator and frame set.
  • [b] Gold = LOF in-water diver transect (Mayfield & Dempsey 2026, Oceans 7:1, suppl. S1): a different instrument, site-level only, never within 200 m of the frames.
  • [c] Anderson's own CoralNet points on the ground-truth frames at this site (30 − WATER − FISH, Unk kept).
  • [d] CoralNet robot labels (the cover set), same denominator.
  • [e] DINOv3 point predictions at the same CoralNet points (CRD AI/chagos_cv_2026-09/chagos_dinov3_point_predictions.csv, 2,500 frames).
  • [f] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not trusted (Anderson): shown, never averaged. Area fraction, not a point count.
  • [g] GENUINELY NOT RUN: only the EDITED website photos were staged for CoralSCOP (Chagos_coralscop_20260920, 2,498 ok); no unedited Chagos CoralSCOP run exists (wide search 2026-09-29).
  • [h] 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.
  • [i] Claude arm chagos_claude_arm COMPLETE (2,500/2,500 frames; 2,495 assigned to these 104 site rows): blind calls at 30 CoralNet points on the EDITED website photos, HC+BL / (30 − WATER − FISH), Unk kept.
  • 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

Chagos #1 2015

55 sites · 1,324 photos

Gold standard = LOF diver transects (no photos). 120 photos manually scored (via CoralNet).

SiteDateClaudeGold standardManually annotated in CoralNetCoralNetDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
BICA012015-03-090.0n=14.60.0n=1—3.3n=11.3n=10.4n=10.3n=11.2 ± 1.80.0–3.3
BIEG022015-03-1018.2n=2131.825.5n=21—23.7n=2131.1n=2134.5n=2123.7n=2121.9 ± 3.218.2–23.7
BIEG032015-03-1016.6n=3637.323.9n=36—18.4n=3631.5n=3634.0n=3625.7n=3620.2 ± 4.816.6–25.7
BIEG042015-03-1025.9n=1929.233.9n=19—35.1n=1936.4n=1938.7n=1928.2n=1929.7 ± 4.825.9–35.1
BIEG052015-03-1139.1n=5946.740.7n=3557.8n=2440.2n=5961.2n=5963.6n=5941.4n=5944.6 ± 8.839.1–57.8
BIEG062015-03-1131.6n=3043.5—44.2n=3034.0n=3046.0n=3046.5n=3022.6n=3033.1 ± 8.922.6–44.2
BIEG072015-03-1128.5n=847.540.6n=8—56.7n=830.9n=833.5n=826.8n=837.3 ± 16.826.8–56.7
BIDI082015-03-1235.2n=3559.0—36.4n=3529.1n=3550.2n=3553.1n=3538.9n=3534.9 ± 4.229.1–38.9
BIDI092015-03-1233.6n=3037.7—38.0n=3023.6n=3056.2n=3061.0n=3032.2n=3031.9 ± 6.023.6–38.0
BIDI102015-03-1222.5n=1323.0—27.1n=1316.2n=1344.2n=1349.4n=1332.9n=1324.7 ± 7.116.2–32.9
BIDI122015-03-132.9n=144.2—10.4n=1411.0n=148.2n=1410.0n=1417.9n=1410.6 ± 6.12.9–17.9
BIDI132015-03-1317.8n=117.8—20.5n=1118.2n=1122.0n=1123.5n=1124.9n=1120.4 ± 3.317.8–24.9
BIDI142015-03-1319.2n=915.2—23.5n=921.9n=926.7n=926.2n=930.3n=923.7 ± 4.719.2–30.3
BIDI152015-03-1436.4n=920.1—24.6n=937.0n=945.3n=948.3n=920.6n=929.6 ± 8.320.6–37.0
BIDI162015-03-1424.4n=1330.4—25.6n=1322.3n=1332.2n=1335.6n=1330.9n=1325.8 ± 3.722.3–30.9
BIDI16.52015-03-1426.3n=1145.6n=13—38.7n=1141.8n=1137.2n=1139.4n=1129.7n=1134.1 ± 7.326.3–41.8
BIEA182015-03-1517.7n=3520.7—22.3n=3514.6n=3532.9n=3532.4n=3428.1n=3520.7 ± 5.914.6–28.1
BIEA192015-03-1530.8n=1427.5—27.2n=1425.5n=1441.9n=1444.8n=1439.2n=1430.7 ± 6.125.5–39.2
BIEA202015-03-1522.2n=1910.8—24.6n=1916.0n=1930.8n=1935.4n=1927.1n=1922.5 ± 4.816.0–27.1
BIEA222015-03-1630.8n=2321.6—49.2n=2344.8n=2344.0n=2346.1n=2350.0n=2343.7 ± 8.930.8–50.0
BIEA232015-03-1628.5n=2424.6—35.7n=2428.8n=2443.8n=2449.2n=2439.5n=2433.1 ± 5.428.5–39.5
BIEA242015-03-1637.4n=1112.7—40.6n=1141.2n=1148.5n=1149.3n=1139.2n=1139.6 ± 1.737.4–41.2
BIEA24.52015-03-1615.2n=5——19.6n=520.7n=520.3n=522.0n=521.2n=519.2 ± 2.715.2–21.2
BIEA252015-03-1716.4n=1915.7—29.5n=1925.6n=1931.1n=1932.8n=1929.8n=1925.3 ± 6.216.4–29.8
BIEA262015-03-1714.5n=3021.0—21.0n=3014.1n=3023.3n=3028.4n=3027.2n=3019.2 ± 6.214.1–27.2
BITB272015-03-1723.8n=1636.6—34.5n=1624.4n=1645.9n=1647.8n=1637.5n=1630.1 ± 7.023.8–37.5
BITB282015-03-1818.5n=4328.5—26.9n=4320.2n=4333.5n=4337.5n=4329.0n=4323.6 ± 5.118.5–29.0
BITB292015-03-1818.5n=4930.2—29.6n=4923.7n=4939.5n=4942.1n=4925.9n=4924.4 ± 4.618.5–29.6
BITB302015-03-1814.0n=1242.5—29.9n=1218.3n=1233.9n=1238.2n=1230.3n=1223.1 ± 8.214.0–30.3
BITB322015-03-1911.7n=1930.3—16.7n=1911.8n=1917.9n=1919.8n=1919.9n=1915.0 ± 4.011.7–19.9
BITB332015-03-197.9n=2718.5—20.3n=2711.5n=2718.4n=2617.5n=2622.3n=2715.5 ± 6.97.9–22.3
BITB342015-03-1937.0n=1161.1—41.4n=1134.5n=1156.5n=1166.7n=1046.6n=1139.9 ± 5.334.5–46.6
BITB34.52015-03-1921.9n=3——47.2n=353.3n=329.2n=328.6n=333.4n=339.0 ± 14.121.9–53.3
BICC352015-03-206.7n=1513.8—17.0n=1513.1n=1521.0n=1523.9n=1520.5n=1514.3 ± 5.96.7–20.5
BICC362015-03-206.4n=155.8—19.4n=1518.2n=1521.0n=1526.5n=1524.7n=1517.2 ± 7.76.4–24.7
BITB372015-03-2029.0n=5638.8—37.3n=5629.6n=5751.2n=5554.5n=5542.4n=5734.6 ± 6.429.0–42.4
BITB382015-03-2126.3n=4456.0—33.7n=4428.1n=4454.1n=4456.3n=4431.5n=4429.9 ± 3.326.3–33.7
BITB392015-03-2136.7n=6050.3—40.1n=6031.7n=6153.3n=5957.5n=5944.4n=6138.2 ± 5.431.7–44.4
BITB402015-03-2123.6n=544.9—28.5n=522.7n=539.7n=545.7n=530.8n=526.4 ± 3.922.7–30.8
BITB412015-03-2224.9n=1341.9—31.8n=1331.8n=1340.3n=1342.0n=1339.5n=1332.0 ± 6.024.9–39.5
BITB422015-03-2222.6n=4936.2—31.1n=4924.4n=5139.5n=5042.8n=5135.4n=5128.4 ± 5.922.6–35.4
BITB432015-03-2222.3n=1228.7—35.9n=1214.7n=1246.7n=1250.3n=1236.6n=1227.4 ± 10.714.7–36.6
BITB442015-03-2326.9n=3640.5—33.2n=3629.4n=3644.2n=3647.0n=3638.7n=3632.0 ± 5.126.9–38.7
BITB452015-03-2333.3n=1345.6—44.5n=1339.2n=1350.5n=1351.9n=1336.1n=1338.3 ± 4.833.3–44.5
BITB462015-03-2318.4n=1125.8—27.5n=1115.2n=1134.5n=1037.7n=1031.3n=1123.1 ± 7.515.2–31.3
BITB472015-03-2425.5n=253.7—43.0n=235.0n=246.0n=245.2n=233.2n=234.2 ± 7.225.5–43.0
BICC502015-03-2515.7n=1112.1—29.4n=1130.0n=1136.9n=1148.0n=1128.8n=1126.0 ± 6.915.7–30.0
BICC512015-03-2520.7n=1715.4—21.9n=1724.1n=1726.9n=1728.4n=1727.5n=1723.6 ± 3.020.7–27.5
BIEA522015-03-2534.5n=2234.1—38.1n=2221.8n=2249.6n=2250.0n=2239.1n=2233.4 ± 8.021.8–39.1
BITB532015-03-2623.4n=2236.7—35.2n=2230.8n=2240.7n=2242.0n=2227.4n=2229.2 ± 5.023.4–35.2
BITB542015-03-2629.8n=9949.5—34.3n=9920.8n=9950.0n=9753.9n=9730.9n=9928.9 ± 5.820.8–34.3
BITB552015-03-2633.2n=1447.6—43.3n=1433.8n=1447.1n=1451.0n=1438.3n=1437.1 ± 4.733.2–43.3
BITB562015-03-2719.4n=1321.3—19.9n=1320.3n=1325.0n=1326.2n=1318.0n=1319.4 ± 1.018.0–20.3
BITB572015-03-27–2843.4n=11140.1—44.0n=11133.2n=11165.6n=11167.8n=11136.4n=11139.2 ± 5.333.2–44.0
BITB582015-03-2831.0n=167.0—31.0n=140.0n=128.1n=132.6n=129.9n=133.0 ± 4.729.9–40.0
All sites (unweighted mean)23.6———26.437.540.330.826.9 ± 3.623.6–30.8
Notes for Chagos #1 2015 (14)
  • [a] Value read from cv_method_tables.json (injector output, generated 2026-09-28 22:33); the scope text there gives denominator and frame set.
  • [b] Gold = LOF in-water diver transect (Mayfield & Dempsey 2026, Oceans 7:1, suppl. S1): a different instrument, site-level only, never within 200 m of the frames.
  • [c] CoralNet robot labels (the cover set), same denominator.
  • [d] DINOv3 point predictions at the same CoralNet points (CRD AI/chagos_cv_2026-09/chagos_dinov3_point_predictions.csv, 2,500 frames).
  • [e] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not trusted (Anderson): shown, never averaged. Area fraction, not a point count.
  • [f] GENUINELY NOT RUN: only the EDITED website photos were staged for CoralSCOP (Chagos_coralscop_20260920, 2,498 ok); no unedited Chagos CoralSCOP run exists (wide search 2026-09-29).
  • [g] 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.
  • [h] Claude arm chagos_claude_arm COMPLETE (2,500/2,500 frames; 2,495 assigned to these 104 site rows): blind calls at 30 CoralNet points on the EDITED website photos, HC+BL / (30 − WATER − FISH), Unk kept.
  • [i] Anderson's own CoralNet points on the ground-truth frames at this site (30 − WATER − FISH, Unk kept).
  • [j] No LOF transect at this site: the gold cell is Anderson's own CoralNet points (frames manually annotated in CoralNet) instead.
  • [k] No gold standard exists for this region.
  • 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

All trips: what has run, photo basis and photo counts

TripDivesPhotosReference / transectsManual pointsCoralNetClaudeDINOv3Coralscapes v2Coralscapes v1CoralSCOPTable
Halmahera 2026233,647 (1,687 iCloud-dataless on 09-25)———3,593/3,593 UNEDITED - full-set arm, complete3,593 UNEDITED3,593 UNEDITED3,593 UNEDITED3,733 ok UNEDITED (map build)included
Alor 2026172,618———2,723/2,723 UNEDITED - full-set arm, complete2,723 UNEDITED2,709 UNEDITED2,712 UNEDITED3,375 ok UNEDITED (alor_2026_full)included
West Timor 2026 (not on list)135————————excluded - nothing run
BVI Necker 2026 (snorkels; not on list)5324—35 frames / 1,050 pts (Anderson, Red Dock only)275 frames robot-labelled but label set has no hard-coral class (unusable)310/310 UNEDITED - all CoralNet frames, complete324 UNEDITED——324 ok UNEDITEDincluded
Anguilla 20264348—pending Anderson's annotationspending Anderson's annotations302/302 UNEDITED - full-set arm, complete302 UNEDITED302 UNEDITED302 UNEDITED299 ok UNEDITEDincluded
St. Eustatius 20264364—pending Anderson's annotationspending Anderson's annotations299 UNEDITED - full-set arm (saba_statia_sxm, 833/833 complete)299 UNEDITED298 UNEDITED297 UNEDITED293 ok UNEDITEDincluded
Saba 20264379—pending Anderson's annotationspending Anderson's annotations342 UNEDITED - full-set arm (saba_statia_sxm, complete)342 UNEDITED341 UNEDITED341 UNEDITED325 ok UNEDITEDincluded
St. Maarten 20262206—pending Anderson's annotationspending Anderson's annotations192 UNEDITED - full-set arm (saba_statia_sxm, complete)192 UNEDITED191 UNEDITED191 UNEDITED190 ok UNEDITEDincluded
Florida 2019-2025311,070NOAA AOML FKNMS / Cheeca / Urban-corals transect sites - 'likely', data file not located locally———unverified (backlog 09-25: 1/31 - path points into maldives_map_build, a mis-match)——653 ok (florida_map_build) + 171 ok (NOAA urban corals) + 197 ok (Cheeca Rocks all photos); mixed edited/uneditedexcluded - CoralSCOP only
Mauritius 20257615draft manuscript manual (unpublished)344 frames x 30 pts (TableS2)344 frames (CoralNet classifier)344 UNEDITED - manuscript context protocol (not fully blind); + blind ctx batches / Codex subsets344 UNEDITED574 run UNEDITED - on external drive (not verified)574 run UNEDITED (323 in site table after >50% background filter)344 ok UNEDITED (coralscop_manual323)included
Red Sea Saudi 2023-2445 (44 in the full Claude arm; KAUST 63-72 not built)4,796 (Saudi 2023 1,409 + 2024 3,387; backlog)Mayfield 2025 Environments 12:248 Table 4 - by AREA only (Thuwal/Jeddah/Rabigh etc., 3,900 photos)751 frames / 22,529 pts abmayfield (Saudi+Egypt export)6,330 frames / 189,871 robot pts (Saudi+Egypt; not site-joined)benchmark 587 frames (of 746, mostly UNEDITED; complete) + full-set EDITED arm RUNNING (Saudi 406 of 4,574 when read)584 (benchmark CoralNet frames only)4,553 EDITED (by pin)4,554 UNEDITED (by pin)full-set EDITED partial 240/4,495 (running) + benchmark 746 (both countries)included (partial)
Egypt 2024 (Red Sea)121,823Table 4 by area (Marsa Nakari / Wadi Lahami, 1,854 photos)(in Red Sea export above)(in Red Sea export above)benchmark 159 frames (09-20 and 09-22 dives only) + full-set arm RUNNING (Egypt 702 of 1,771 when read)159 (benchmark frames only)1,083 EDITED (by pin)933 UNEDITED (by pin)edited partial 243/1,090 + unedited/AI partial 1,367/1,969 (running)included (partial)
South Africa 20242191————————excluded - nothing run
Bahrain / Persian Gulf 2023-244191———————36 ok EDITED (2023 dives only; 2024 not run)included (Anderson's exception; cover very low)
Oman 2024 (not on list)7703Mayfield 2025 Platax Table 1 (660 photos)200 frames / 6,000 pts abmayfield (Novice Bay 135 + Cave Site 65)461 frames robot103 EDITED sheet-B - SAMPLE (15/site); + 103 UNEDITED sheet-B fair test554 - site-level file only (per-frame on external drive)661 UNEDITED (external drive; same)661 UNEDITED (external drive; site values in CORALSCAPES_V1_V2_UNEDITED_20260918.md)unedited 650 ok/737 + edited 670 ok/677included
Kenya 2024192,201Mayfield 2025 Environments 12:276 App. A (2,300 photos, CoralNet)2,322 images; 22,016 abmayfield pts (+47,644 robot pts in the same export)2,298 frames Machine-suggestion-1 (CIRCULAR with the reference)703 UNEDITED (batches 1-10; 8 of 20 sites) + kenya_claude_arm 99-frame benchmark arms2,317 UNEDITED (pins; kenya_1450_dinov3 = 1,450)2,319 UNEDITED (external drive; site values on pins)2,319 UNEDITED (external drive; site values on pins)1,444 ok UNEDITED (kenya_originals, 1,450) + 50 (batch 1)included
East Timor (Atauro) 20246948———990/990 EDITED - full-set arm, complete990 EDITED——1,075 ok EDITED (atauro_edited_2026-09)included
Bahamas 20222194———————192 okexcluded - CoralSCOP only
Turks & Caicos 2022221,130———————1,071 okexcluded - CoralSCOP only
St. Lucia 20222247———————326 okexcluded - CoralSCOP only
Dominica 20223364———————1,049 ok UNEDITED + edited variants (57 edited, 499 LR-edited, 211 paired-edited, 27+26 Toucari)excluded - CoralSCOP only
Dry Tortugas 202111243———————272 ok + 233 ok (v2)excluded - CoralSCOP only
Barbados 20202148———————147 ok (UPDATED: backlog 09-25 said 0/2)excluded - CoralSCOP only
SVG 201915438————————excluded - nothing run
FGB (TX) 20195162———————156 ok (noaa_map_build)excluded - CoralSCOP only
Komodo 2019231,470———1,411/1,411 EDITED - full-set arm, complete1,381 in site table (1,456 selected)1,411 EDITED1,411 EDITED1,485 ok EDITEDincluded
Raja Ampat 2018 (+Triton Bay)201,144———1,110/1,110 EDITED - full-set arm, complete1,094 in site table (1,241 selected)1,243 EDITED1,110 EDITED1,238 ok EDITED (unedited never run)included
Spermonde 2018 (not on list)2142———————139 okexcluded - CoralSCOP only
Maluku 2018 (Ambon + Banda)111,092———Banda 885/885 EDITED - full-set arm, complete; Ambon noneBanda 808Banda 808 EDITEDBanda 808 EDITEDBanda 939 ok + Ambon 179 okBanda included; Ambon excluded (CoralSCOP only)
Maldives #2 201733925————4,067 photos both trips (919 in #2 table)——878 ok both trips, staged subset (450 in #2 table)included
Bonaire 201712616———————610 ok EDITEDexcluded - CoralSCOP only
Philippines 2017 (Palawan)7224————605 (396 excl. topside)——175 ok + 424 ok (v2 flat run)included
Philippines 2016 (Anilao)6466————609——496 ok (+60 ok Verde Island Passage build)included
Dongsha 20165241———————193 okexcluded - CoralSCOP only
Maldives #1 2016261,2112016 bleaching sheets - Apr/Jul 2016 transects at other sites (partial; not these Jan dives)———(see #2; 2,934 in #1 table)——(see #2; 349 in #1 table)included
Chagos #2 2015571,373LOF GRE in-water transects (suppl. S1)600 frames / 18,000 pts both trips (480 frames at #2 sites)1,900 frames robot both trips (695 at #2 sites)2,500/2,500 EDITED both trips (1,175 at #2 sites) - complete2,500 EDITED both trips2,500 EDITED both trips2,595 EDITED both trips2,498 ok EDITED both trips (unedited never run)included
Chagos #1 2015601,751 (101 dataless)LOF GRE in-water transects(120 frames at #1 sites)(1,200 at #1 sites)(1,320 at #1 sites)(1,324 at #1 sites)(1,314 at #1 sites)(1,315 at #1 sites)(1,324 at #1 sites)included
Palau 2015842,643LOF GRE933 transects (93 dives)———————excluded - nothing run
Solomons 2014722,064 (804 dataless)LOF GRE933 transects (74 dives)———————excluded - nothing run
GBR 2014771,369LOF GRE933 transects (73 dives)———————excluded - nothing run
New Caledonia 2013591,572LOF GRE933 transects (61 dives)———————excluded - nothing run
Tonga 201342599LOF GRE933 + Tonga-Benthic-KSLOF.xlsx (50 dives)———————excluded - nothing run
Fiji 201350461LOF GRE933 transects (55 dives)———————excluded - nothing run
Australs-Cooks 201349321LOF GRE933 + cook islands compiled benthic.xlsx (61 dives)———————excluded - nothing run
Philippines 201300 (topside only)————————excluded - no underwater photos
Taiwan00 resolvable (backlog)———————none (Catlin Seaview CoralSCOP blocked: images on external drive)excluded