Reef cover by method

Hard-coral cover (%) at each site from every method run so far, one table per region, newest trip first (36 regions, 427 sites). Built from the method output files.

Shaded value = in the CV mean (shade deepens with cover) Gray italic = shown for comparison but left out of the CV mean: the gold standard (the reference value) and the manually annotated in CoralNet CoralNet benchmark annotations (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

22 sites · 3,593 photos

No gold standard.

SiteDateHeadline (site page)ClaudeDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
Failonga2026-08-0933%Claude33a,en=20131a,bn=20155a,cn=20158a,cn=20132a,dn=20132 ± 0.631–33
Maitara North2026-08-0933%Claude33a,en=17134a,bn=17149a,cn=17151a,cn=17126a,dn=17131 ± 4.126–34
AM Point 22026-08-0940%Claude40a,en=22333a,bn=22357a,cn=22361a,cn=22328a,dn=22334 ± 6.228–40
Pretty Shallow2026-08-1032%Claude32a,en=10532a,bn=10554a,cn=10556a,cn=10530a,dn=10531 ± 1.130–32
Kusu Drift2026-08-1026%Claude26a,en=14328a,bn=14351a,cn=14356a,cn=14334a,dn=14329 ± 4.126–34
Pulau Sebatang2026-08-1041%Claude41a,en=26534a,bn=26565a,cn=26567a,cn=26536a,dn=26537 ± 3.934–41
Tower Corner2026-08-1134%Claude34a,en=24235a,bn=24255a,cn=24258a,cn=24234a,dn=24235 ± 0.734–35
Sali Bay-Batu Mangga2026-08-1125%Claude25a,en=36132a,bn=36152a,cn=36153a,cn=36131a,dn=36129 ± 3.925–32
Tanjung Naraka2026-08-1230%Claude30a,en=13630a,bn=13654a,cn=13655a,cn=13635a,dn=13632 ± 3.230–35
Tanjung Dehebra2026-08-1225%Claude25a,en=10328a,bn=10352a,cn=10354a,cn=10334a,dn=10329 ± 4.825–34
Kusu Ridge2026-08-1245%Claude45a,en=24943a,bn=24964a,cn=24966a,cn=24928a,dn=24939 ± 9.528–45
Jerot2026-08-1323%Claude23a,en=17228a,bn=17255a,cn=17257a,cn=17242a,dn=17231 ± 9.723–42
Proco2026-08-1332%Claude32a,en=22732a,bn=22751a,cn=22754a,cn=22734a,dn=22733 ± 1.132–34
Kusu Drift (round 2)2026-08-1343%Claude43a,en=10532a,bn=10561a,cn=10565a,cn=10538a,dn=10538 ± 5.632–43
Merida Cavern2026-08-1427%Claude27a,en=12524a,bn=12538a,cn=12541a,cn=12531a,dn=12528 ± 3.624–31
Batu Balong shallows-near anchorage2026-08-1457%Claude57a,en=21333a,bn=21370a,cn=21374a,cn=21338a,dn=21343 ± 12.233–57
Batu Balong2026-08-1439%Claude39a,en=12628a,bn=12660a,cn=12661a,cn=12634a,dn=12634 ± 5.728–39
Batu Mandi2026-08-1412%Claude12a,en=2217a,bn=2234a,cn=2236a,cn=2228a,dn=2219 ± 8.412–28
Siko Tanjung2026-08-1541%Claude41a,en=14331a,bn=14358a,cn=14359a,cn=14332a,dn=14335 ± 5.831–41
Siko Pinnacle2026-08-1512%Claude12a,en=4819a,bn=4833a,cn=4837a,cn=4827a,dn=4819 ± 7.512–27
Reene's Rock (Pulau Gewang)2026-08-1513%Claude13a,en=3918a,bn=3932a,cn=3937a,cn=3923a,dn=3918 ± 4.913–23
Lelay Utara (north)2026-08-1648%Claude48a,en=17431a,bn=17457a,cn=17458a,cn=17430a,dn=17436 ± 9.930–48
All sites (unweighted mean)—323053553231 ± 1.530–32
Notes for Halmahera 2026 (9)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [b] 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's site cover table).
  • [c] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not considered reliable for these photos: shown, never averaged. Area fraction, not a point count.
  • [d] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [e] Claude = blind 7-call point count on the UNEDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator; mean across photos.
  • 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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.

SiteDateHeadline (site page)ClaudeDINOv3Coralscapes v1Coralscapes v2CoralSCOP uneditedCoralSCOP editedMean ± SD of CVRange
Goat Fence2026-07-2811%Claude11a,in=7219a,bn=7250a,cn=7250a,cn=72—29a,d,hn=7220 ± 9.411–29
Pasir Merah (Red Sand)2026-07-2916%Claude16a,fn=17426a,bn=17439a,cn=17443a,cn=17416a,dn=174—e19 ± 5.616–26
Bakalang Jetty2026-07-297.6%Claude7.6a,fn=5116a,bn=5125a,cn=5126a,cn=5118a,dn=51—e14 ± 5.57.6–18
Bama Wall2026-07-2911%Claude11a,fn=23327a,bn=23324a,cn=23327a,cn=2339.0a,dn=233—e16 ± 10.19.0–27
Beang Abang (Baeng Beach / Baeng Jetty)2026-07-30–31—3.5a,f,gn=2612a,bn=2614a,cn=2619a,cn=2611a,dn=26—e——
Goa Wolang (Walang Cave)2026-07-307.7%Claude7.7a,fn=2819a,bn=2815a,cn=2818a,cn=288.8a,dn=28—e12 ± 6.37.7–19
Clown Valley2026-07-3041%Claude41a,fn=20932a,bn=20957a,cn=20959a,cn=20922a,dn=209—e32 ± 9.322–41
Motonglang (Cathedral)2026-07-305.8%Claude5.8a,fn=13228a,bn=13220a,cn=13225a,cn=1327.6a,dn=132—e14 ± 12.25.8–28
Limarahing (Black Sands)2026-07-3129%Claude29a,fn=16232a,bn=16249a,cn=16251a,cn=16218a,dn=162—e26 ± 7.118–32
Apuri Bay2026-07-3129%Claude29a,fn=20828a,bn=20848a,cn=20850a,cn=20720a,dn=208—e26 ± 4.820–29
Monaseli (Munaseli)2026-08-0120%Claude20a,fn=17821a,bn=17838a,cn=17839a,cn=17816a,dn=178—e19 ± 2.716–21
Babylon2026-08-01–0230%Claude30a,fn=41530a,bn=41552a,cn=41555a,cn=41521a,dn=415—e27 ± 5.621–30
The Arch2026-08-0223%Claude23a,fn=15523a,bn=15549a,cn=15551a,cn=15519a,dn=155—e22 ± 2.419–23
Umapura2026-08-0218%Claude18a,fn=19729a,bn=19741a,cn=18648a,cn=18418a,dn=197—e21 ± 6.418–29
Harilolong (School's Out)2026-08-0336%Claude36a,fn=27531a,bn=27552a,cn=27554a,cn=27521a,dn=275—e29 ± 7.721–36
Sulambali Wall2026-08-0322%Claude22a,fn=20825a,bn=20844a,cn=20848a,cn=20816a,dn=208—e21 ± 4.516–25
All sites (unweighted mean)—20264043——23 ± 3.820–26
Notes for Alor 2026 (13)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [b] 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's site cover table).
  • [c] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not considered reliable for these photos: shown, never averaged. Area fraction, not a point count.
  • [d] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [e] Not run: this site's photos are unedited (Goat Fence alone was scored on edited photos).
  • [f] Claude = blind 7-call point count on the UNEDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator; mean across photos.
  • [g] No coral-cover estimate for this site: its photographs show no reef habitat (muck dive), so these values are a comparison only, left out of the row mean and the All-sites row.
  • [h] CoralSCOP here = the site page's frame-matched row: run on exactly the photos behind the Claude cell (per-frame values joined by file; capped at 100% per photo; zero-filled). Comparison only.
  • [i] Claude = blind 7-call point count on the EDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator; mean across photos.
  • 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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)

4 sites · 324 photos

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

SiteDateHeadline (site page)ClaudeGold standardDINOv3CoralSCOPMean ± SD of CVRange
Red Dock2026-07-017.4%manual, area8.0dn=876.3en=3520bn=9321cn=9316 ± 7.28.0–21
Turtle Beach2026-07-01–023.1%Claude, area3.1dn=83—a17bn=9117cn=9112 ± 7.93.1–17
Bali Hai to Turtle Beach2026-07-022.5%Claude, area2.5dn=89—a14bn=8912cn=899.5 ± 6.12.5–14
Watersport Beach to Bali Hai2026-07-022.4%Claude, area2.4dn=51—a11bn=5112cn=518.7 ± 5.52.4–12
All sites (unweighted mean)4.0%Claude, island4.0—161612 ± 6.74.0–16
Notes for BVI Necker 2026 (snorkels) (9)
  • [a] No hand-labeled frame at this area.
  • [b] 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's site cover table).
  • [c] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [d] 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 (the arm's results notes). 310/310 frames.
  • [e] Gold = the author's own CoralNet labels on the 35 Red Dock frames in which he labeled all 30 points, on the all-30 denominator used for Claude here. With water and fish points left out of the denominator, the same frames read 7.4%, the value quoted on the Red Dock page.
  • 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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.

SiteDateHeadline (site page)ClaudeDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
Deep South2026-02-286.6%Claude6.6a,fn=6313a,bn=639.8a,cn=6311a,cn=6320a,d,en=6313 ± 6.96.6–20
Oosterdiep Wreck2026-02-283.4%Claude3.4a,fn=6616a,bn=666.0a,cn=666.1a,cn=6615a,dn=6012 ± 7.13.4–16
The Pipe2026-03-026.6%Claude6.6a,fn=6519a,bn=6510a,cn=6511a,cn=6519a,d,en=6515 ± 7.36.6–19
Catheley H2026-03-026.2%Claude6.2a,fn=10815a,bn=1089.4a,cn=1089.7a,cn=10816a,d,en=10812 ± 5.56.2–16
All sites (unweighted mean)—5.7168.99.51813 ± 6.55.7–18
Notes for Anguilla 2026 (10)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [b] 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's site cover table).
  • [c] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not considered reliable for these photos: shown, never averaged. Area fraction, not a point count.
  • [d] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [e] CoralSCOP here = the site page's frame-matched row: run on exactly the photos behind the Claude cell (per-frame values joined by file; capped at 100% per photo; zero-filled). Comparison only.
  • [f] Claude = blind 7-call point count on the UNEDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator; mean across photos.
  • 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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.

SiteDateHeadline (site page)ClaudeDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
The Ledge to Hangover2026-02-242.9%Claude2.9a,fn=8817a,bn=8817a,cn=8721a,cn=8826a,d,en=8815 ± 11.62.9–26
STENAPA Reef2026-02-244.6%Claude4.6a,fn=10318a,bn=1037.2a,cn=1027.5a,cn=10220a,d,en=10314 ± 8.54.6–20
Double Wreck2026-02-255.6%Claude5.6a,fn=4616a,bn=4613a,cn=4614a,cn=4626a,d,en=4616 ± 10.35.6–26
Chien Tong wreck2026-02-253.8%Claude3.8a,fn=6217a,bn=626.0a,cn=625.6a,cn=6213a,d,en=6211 ± 6.53.8–17
All sites (unweighted mean)—4.21711122114 ± 8.84.2–21
Notes for St. Eustatius 2026 (10)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [b] 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's site cover table).
  • [c] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not considered reliable for these photos: shown, never averaged. Area fraction, not a point count.
  • [d] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [e] CoralSCOP here = the site page's frame-matched row: run on exactly the photos behind the Claude cell (per-frame values joined by file; capped at 100% per photo; zero-filled). Comparison only.
  • [f] Claude = blind 7-call point count on the UNEDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator; mean across photos.
  • 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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.

SiteDateHeadline (site page)ClaudeDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
Ladder Labyrinth2026-02-218.9%Claude8.9a,en=8218a,bn=8213a,cn=8215a,cn=8223a,d,fn=8217 ± 7.28.9–23
Man-of-War Shoals2026-02-213.7%Claude3.7a,en=10220a,bn=10216a,cn=10125a,cn=10129a,d,fn=10218 ± 13.03.7–29
Tent Drift2026-02-214.3%Claude4.3a,en=7318a,bn=7312a,cn=7314a,cn=7321a,d,fn=7314 ± 8.94.3–21
Tent Shallow2026-02-223.0%Claude3.0a,en=8515a,bn=857.9a,cn=8512a,cn=8521a,dn=7213 ± 9.33.0–21
All sites (unweighted mean)—5.01812172415 ± 9.65.0–24
Notes for Saba 2026 (10)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [b] 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's site cover table).
  • [c] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not considered reliable for these photos: shown, never averaged. Area fraction, not a point count.
  • [d] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [e] Claude = blind 7-call point count on the UNEDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator; mean across photos.
  • [f] CoralSCOP here = the site page's frame-matched row: run on exactly the photos behind the Claude cell (per-frame values joined by file; capped at 100% per photo; zero-filled). Comparison 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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.

SiteDateHeadline (site page)ClaudeDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
Mike's Maze2026-02-1914%Claude14a,en=7819a,bn=7828a,cn=7733a,cn=7725a,dn=7519 ± 5.914–25
Proselyte Reef2026-02-197.9%Claude7.9a,en=11418a,bn=11421a,cn=11422a,cn=11424a,d,fn=11417 ± 8.07.9–24
All sites (unweighted mean)—111924282418 ± 6.911–24
Notes for St. Maarten 2026 (10)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [b] 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's site cover table).
  • [c] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not considered reliable for these photos: shown, never averaged. Area fraction, not a point count.
  • [d] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [e] Claude = blind 7-call point count on the UNEDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator; mean across photos.
  • [f] CoralSCOP here = the site page's frame-matched row: run on exactly the photos behind the Claude cell (per-frame values joined by file; capped at 100% per photo; zero-filled). Comparison 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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

Florida 2019-2025

16 sites · 742 photos

Survey column = the NOAA NCRMP / AOML 2016 photo-transect value, the headline at Cheeca Rocks, The Rocks, Little Conch, and Crocker Reef. Claude and CoralSCOP ran on the author's photos.

SiteDateHeadline (site page)ClaudeSurvey (headline)CoralSCOP uneditedCoralSCOP edited
Emerald Reef (Key Biscayne)2019-05-15–2021-05-2712%Claude12dn=50—a—f34b,in=57
Cheeca Rocks (FKNMS)2019-05-31–2021-05-0632%transect40d,hn=13632en=40—f47b,gn=136
MacArthur Causeway South (Urban Corals), Miami2019-06-2619%*Claude19d,jn=13—a—f—
Crocker Reef (FKNMS)2019-07-10–12-041.5%published—1.5en=40—f—
Little Conch (FKNMS)2019-07-10–12-046.9%published—6.9en=40—f—
The Rocks (FKNMS)2019-07-10–12-049.3%published—9.3en=40—f—
Looe Key2021-12-0920%Claude20dn=108—a—f25b,in=87
Tenneco Towers†2022-02-262.6%Claude2.6dn=45—a13b,gn=45—
Rabbit Ears (Pompano Beach)2022-04-084.4%*Claude4.4d,jn=17—a—f27b,in=18
United Caribbean (wreck, Fort Lauderdale)2022-04-083.8%*Claude3.8d,jn=14—a—f19b,in=14
Neptune Memorial Reef2022-04-161.6%Claude1.6dn=35—a—f—
Orion shipwreck (Key Biscayne)2022-04-164.2%Claude4.2dn=65—a—f15b,in=78
Key Largo Dry Rocks (Christ of the Abyss)2023-04-158.3%Claude8.3d,jn=23—a—f22b,in=39
North Dry Rocks (Key Largo)2023-04-1520%Claude20dn=113—a—f27b,in=118
Belcher Barge (Key Biscayne)†2025-11-221.4%Claude1.4dn=74—a14bn=76—c
Flamingo Reef (Key Biscayne)†2025-11-2221%Claude21dn=49—a31bn=58—c
Notes for Florida 2019-2025 (16)
  • [a] No survey at this site (no transect, no frames manually annotated in CoralNet).
  • [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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [c] Not applicable: the 2025 Key Biscayne dives have no edited set.
  • [d] Claude = blind 7-call point count (20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator); value = the site's headline where Claude is the headline.
  • [e] NOAA NCRMP / NOAA AOML photo-transect survey, August 2016 (Manzello et al. 2019): the site headline. n = frames of the survey, not the author's photos.
  • [f] Not run on these photos.
  • [g] CoralSCOP here = the site page's frame-matched row: run on exactly the photos behind the Claude cell (per-frame values joined by file; capped at 100% per photo; zero-filled). Comparison only.
  • [h] Comparison only at a survey site: the NOAA NCRMP / AOML 2016 photo-transect value is the headline.
  • [i] CoralSCOP (mask-sum area index) = the value the map published before the Claude arm (folder-level run, not frame-matched to the Claude photos). Comparison only, never averaged.
  • [j] THIN SAMPLE (under 30 frames at this site; marked in the headline column only): 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 in gray italic but never averaged.
  • * The site's headline coral cover rests on only 10–19 reef-habitat photos; this estimate is less certain.
  • † Includes unedited camera originals: the 2025 Key Biscayne dives have no edited set, and Tenneco Towers (26 February 2022) was scored on 7 edited photos and 38 unedited originals.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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).

SiteDateHeadline (site page)ClaudeGold standardCoralNetDINOv3Coralscapes v1CoralSCOPMean ± SD of CVRange
Reef Garden2025-02-0126%manual27an=10526an=10528an=10545an=10538a,b,cn=9722a,dn=10530 ± 9.822–45
Ti Korail2025-02-017.0%manual7.4an=407.0an=4011an=4022an=4013a,b,cn=3821a,dn=4015 ± 7.27.4–22
Bamboo2025-02-0417%manual14an=4017an=4012an=4020an=4013a,b,cn=3917a,dn=4016 ± 3.612–20
Aquarium-Big Rock2025-02-0414%manual8.3an=4014an=407.0an=4015an=408.2a,b,cn=3113a,dn=4011 ± 3.97.0–15
Carpenter2025-02-0915%manual12an=4015an=4011an=4021an=4013a,b,cn=3121a,dn=4016 ± 5.511–21
Lost Anchor2025-02-0911%manual9.2an=3911an=3910an=3920an=3912a,b,cn=3221a,dn=3915 ± 6.29.2–21
Anemone Garden2025-02-102.7%manual2.9an=402.7an=407.7an=4033an=4017a,b,cn=3331a,dn=4019 ± 15.52.9–33
All sites (six reef sites, Table 3 final row; Anemone Garden not counted)—13151324161917 ± 5.313–24
Notes for Mauritius 2025 (8)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [b] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not considered reliable for these photos: shown, never averaged. Area fraction, not a point count.
  • [c] Coralscapes v1 (ver. 1; Sauder et al., 2025) = Table 3 of the Mauritius manuscript (Mayfield et al. 2026 Platax): pixel-area segmentation on the 301 fully human-labeled frames (Table 3 note); never averaged.
  • [d] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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 · 5,148 photos

Gold standard = each site's own published CoralNet point count (Environments 12:248, Appendix A; most frames annotated by CoralNet's AI, trained and checked by the author). 583 photos manually scored (via CoralNet) for the benchmark. Benchmark photos mostly unedited; CoralSCOP ran on edited photos.

SiteDateHeadline (site page)ClaudeGold standardBenchmark subset (manual CoralNet; not the gold standard)DINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
Abu Ejeiah (Thuwal)2023-09-02–2024-10-0415%published point count17jn=36615an=417too few photos (n=4)b,ctoo few photos (n=4)d,c14g,fn=36716e,fn=36724h,in=36721 ± 4.817–24
Cement Reef (Thuwal)2023-09-02–2024-11-2922%published point count24jn=47822an=46922bn=9526dn=9519g,fn=47220e,fn=47229h,in=47127 ± 2.624–29
Cement Wreck (Thuwal)2023-09-08–2024-11-0913%published point count12jn=38713an=38318bn=1134dn=1110g,fn=39211e,fn=39121h,i,kn=39122 ± 10.712–34
Rose Reef (Thuwal)2023-09-08–2024-11-2914%published point count13jn=43814an=720too few photos (n=3)b,ctoo few photos (n=3)d,c16g,fn=45718e,fn=45724h,i,kn=45519 ± 7.513–24
Um Mankout (Yanbu)2023-09-1526%published point count29jn=22726an=228too few photos (n=2)b,ctoo few photos (n=2)d,c26g,fn=22927e,fn=22932h,in=22930 ± 2.129–32
Al Fasmah (Yanbu)2023-09-1622%published point count27jn=11622an=116——24g,fn=11625e,fn=11631h,in=11629 ± 3.127–31
Alrass Alakhdar (Yanbu)2023-09-1630%published point count39jn=5730an=59——27g,fn=5827e,fn=5836h,in=5838 ± 2.336–39
Thetis (Yanbu)2023-10-06–2024-08-0916%published point count12jn=16916an=170too few photos (n=5)b,ctoo few photos (n=5)d,c15g,fn=17116e,fn=17121h,in=17116 ± 5.812–21
Mr. Umluj (Yanbu)2023-10-07–2024-11-0817%published point count17jn=32617an=326——22g,fn=32623e,fn=32622h,in=32620 ± 3.717–22
Long Reef (Thuwal)2023-11-1714%published point count12jn=8814an=90——18g,fn=9120e,fn=9124h,in=9118 ± 9.012–24
Cave Reef (Thuwal)2024-02-0916%published point count16jn=6516an=68too few photos (n=2)b,ctoo few photos (n=2)d,c12g,fn=7113e,fn=7122h,i,kn=7019 ± 4.316–22
Abu Galawa (Yanbu)2024-04-19–11-0818%published point count22jn=13518an=138——22g,fn=14024e,fn=14031h,i,kn=13826 ± 6.622–31
Chicken Wreck (Jeddah)2024-08-02–1717%published point count10jn=17117an=25116bn=12527dn=12615g,fn=21016e,fn=21024h,in=15321 ± 9.210–27
Drogon Corals (Jeddah)2024-08-0214%published point count10jn=6514an=67too few photos (n=1)b,ctoo few photos (n=1)d,c12g,fn=6612e,fn=6622h,in=6616 ± 8.110–22
Anderson's Reef-I (Jeddah)2024-08-0323%published point count20jn=14623an=146too few photos (n=1)b,ctoo few photos (n=1)d,c22g,fn=14724e,fn=14729h,in=14725 ± 6.620–29
Anderson's Reef-II (Jeddah)2024-08-0321%published point count19jn=17521an=174too few photos (n=3)b,ctoo few photos (n=3)d,c18g,fn=17419e,fn=17425h,in=17422 ± 4.119–25
Coral Gardens (Jeddah)2024-08-10–1728%published point count25jn=27128an=30029bn=15644dn=15619g,fn=25421e,fn=25429h,in=25433 ± 10.125–44
north Rabigh (shore) (Rabigh)2024-08-1624%published point count23jn=15824an=17323bn=17557dn=17515g,fn=12216e,fn=12230h,in=12236 ± 17.823–57
unnamed site 1 (Al Qunfudah)2024-08-3016%published point count18jn=12316an=130——18g,fn=11919e,fn=11931h,in=11924 ± 9.718–31
unnamed site 2 (Al Qunfudah)2024-08-3017%published point count11jn=8617an=80——13g,fn=7015e,fn=7024h,in=7018 ± 9.311–24
Shark Reef (Thuwal)2024-08-31–09-1416%published point count14jn=35116an=421——12g,fn=33413e,fn=33422h,i,kn=33018 ± 5.814–22
Bhadur Reef (Jeddah)2024-09-0721%published point count18jn=17421an=222——17g,fn=16817e,fn=16827h,i,kn=16723 ± 6.318–27
All sites (unweighted mean)—1919——18192622 ± 5.519–26
Notes for Red Sea - Saudi Arabia 2023-24 (15)
  • [a] Gold = this site's own published CoralNet point count by the author (Mayfield 2025, Environments 12:248, Appendix A; n = photos; most frames annotated by CoralNet's AI, trained and checked by the author), the value the site page, report card, and map pop-up print. The paper's Table 4 also gives AREA means: Thuwal 16% (n=2567); Jeddah 22% (n=1160); Rabigh 24% (n=173); Yanbu 20% (n=1036); Qunfudah 17% (n=210).
  • [b] Separate benchmark, not a second gold value: the author's own CoralNet points on the benchmark frames (746 frames), (30 − WATER − FISH − DVR), unknown points kept (Red Sea paper Section 2.2), recomputed per frame.
  • [c] Too few photos (under 10) at this site for a value: not shown and not averaged (photo-count rule: 10 or more photos for a value, 20 or more for a confident one).
  • [d] DINOv3 (fine-tuned) point predictions at the author's CoralNet points on the benchmark frames; hard coral / (points − WATER − FISH − DVR by reference label). No full-set DINOv3 run exists.
  • [e] Coralscapes v2 on EDITED frames.
  • [f] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not considered reliable for these photos: shown, never averaged. Area fraction, not a point count.
  • [g] Coralscapes v1 on UNEDITED frames.
  • [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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [i] CoralSCOP full-set run, EDITED frames (the CoralNet-annotated Saudi subset), finished (4,485 of 4,495 frames scored; 10 could not be scored: failed, unreadable or skipped at the memory cap).
  • [j] Claude = the full-set blind point count (EDITED, AI-processed where no edit exists; 20 pts, 7-call, (HC+BL)/(20 − WATER − FISH − OBST)), the value the site pages show.
  • [k] Run finished at this site; n leaves out the frame(s) CoralSCOP could not score (mask generation failed, unreadable file, or skipped at the runner's memory cap), so n is below the site's frame count. Not partial.
  • 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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 · 1,966 photos

Gold standard = each site's own published CoralNet point count (Environments 12:248, Appendix A; most frames annotated by CoralNet's AI, trained and checked by the author). 159 photos manually scored (via CoralNet) for the benchmark. CoralSCOP ran on both edited and unedited photos (separate columns).

SiteDateHeadline (site page)ClaudeGold standardBenchmark subset (manual CoralNet; not the gold standard)DINOv3Coralscapes v1Coralscapes v2CoralSCOP uneditedCoralSCOP editedMean ± SD of CVRange
Marsa Nakari-North (Marsa Nakari)2024-09-2023%published point count23fn=17223an=195——too few photos (n=9)b,c,l17i,cn=18628d,en=20529d,kn=18827 ± 3.523–29
Nakari House Reef-South (Marsa Nakari)2024-09-2025%published point count24fn=10325an=10223gn=3046hn=3010b,cn=4518i,cn=10826d,en=11928d,k,jn=10731 ± 10.224–46
Habili El Ghadeer (Marsa Nakari)2024-09-2113%published point count16fn=14713an=161——too few photos (n=9)b,c,l15i,cn=14824d,en=16824d,kn=14821 ± 4.316–24
Om Halhala (Marsa Nakari)2024-09-2124%published point count29fn=20624an=216——13b,cn=1219i,cn=20630d,en=22130d,k,jn=20030 ± 0.729–30
Gazella Reef (Wadi Lahami)2024-09-2217%published point count18fn=20417an=21423gn=3046hn=3012b,cn=3611i,cn=20522d,e,jn=22222d,kn=20827 ± 13.018–46
Habili Daisy (Wadi Lahami)2024-09-2223%published point count18fn=5023an=52——12b,cn=58too few photos (n=8)i,c,l24d,en=58too few photos (n=8)d,k,l21 ± 4.018–24
Shabrour-north (Wadi Lahami)2024-09-2224%published point count21fn=9224an=10024gn=9945hn=9916b,cn=10118i,cn=9428d,en=10830d,k,jn=7431 ± 10.121–45
Shabrour-south (Wadi Lahami)2024-09-22–2325%published point count24fn=22125an=223——11b,cn=3819i,cn=12827d,e,jn=24129d,kn=12927 ± 2.724–29
Abu Galawa-small (Wadi Lahami)2024-09-2321%published point count26fn=18821an=192——15b,cn=203—23d,en=203—24 ± 1.923–26
Habili Gazella (Wadi Lahami)2024-09-2318%published point count19fn=17618an=181——11b,cn=191—22d,e,jn=190—21 ± 1.819–22
Salah (Wadi Lahami)2024-09-2328%published point count28fn=21228an=218——19b,cn=231—29d,en=231—29 ± 0.628–29
All sites (unweighted mean)—2222————26—24 ± 2.422–26
Notes for Red Sea - Egypt 2024 (16)
  • [a] Gold = this site's own published CoralNet point count by the author (Mayfield 2025, Environments 12:248, Appendix A; n = photos; most frames annotated by CoralNet's AI, trained and checked by the author), the value the site page, report card, and map pop-up print. The paper's Table 4 also gives AREA means: Wadi Lahami 22% (n=1180); Marsa Nakari 21% (n=674).
  • [b] Coralscapes v1 on UNEDITED frames.
  • [c] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not considered reliable for these photos: shown, never averaged. Area fraction, not a point count.
  • [d] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [e] CoralSCOP full-set run, Egypt unedited frames (for dives 58-62, the AI-processed photos), finished (1,966 of 1,969 frames scored; 3 could not be scored: failed, unreadable or skipped at the memory cap).
  • [f] Claude = the full-set blind point count (EDITED, AI-processed where no edit exists; 20 pts, 7-call, (HC+BL)/(20 − WATER − FISH − OBST)), the value the site pages show.
  • [g] Separate benchmark, not a second gold value: the author's own CoralNet points on the benchmark frames (746 frames), (30 − WATER − FISH − DVR), unknown points kept (Red Sea paper Section 2.2), recomputed per frame.
  • [h] DINOv3 (fine-tuned) point predictions at the author's CoralNet points on the benchmark frames; hard coral / (points − WATER − FISH − DVR by reference label). No full-set DINOv3 run exists.
  • [i] Coralscapes v2 on EDITED frames.
  • [j] Run finished at this site; n leaves out the frame(s) CoralSCOP could not score (mask generation failed, unreadable file, or skipped at the runner's memory cap), so n is below the site's frame count. Not partial.
  • [k] CoralSCOP full-set run, EDITED frames, finished (1,062 of 1,090 frames scored; 28 could not be scored: failed, unreadable or skipped at the memory cap).
  • [l] Too few photos (under 10) at this site for a value: not shown and not averaged (photo-count rule: 10 or more photos for a value, 20 or more for a confident one).
  • 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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

4 sites · 129 photos

No gold standard.

SiteDateHeadline (site page)ClaudeCoralSCOP uneditedCoralSCOP editedMean ± SD of CVRange
Dilmunia Rock (Bahrain)2023-09-230.5%*Claude0.5a,cn=10—5.0a,bn=102.8 ± 3.20.5–5.0
Artificial Reef (Bahrain)2023-09-230.9%*Claude0.9a,cn=16—40a,bn=1621 ± 27.80.9–40
Hayr Bul Thamah (Bahrain)2024-08-239.9%Claude9.9a,cn=72————
Al-Boom Wreck (Bahrain)2024-08-230.0%Claude0.0a,cn=319.7a,b,dn=31—4.9 ± 6.90.0–9.7
All sites (unweighted mean)—2.8————
Notes for Bahrain / Persian Gulf 2023-24 (9)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [c] Claude = blind 7-call point count on the EDITED (2023) AND UNEDITED (2024) photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator; mean across photos.
  • [d] CoralSCOP here = the site page's frame-matched row: run on exactly the photos behind the Claude cell (per-frame values joined by file; capped at 100% per photo; zero-filled). Comparison 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 in gray italic but never averaged.
  • * The site's headline coral cover rests on only 10–19 reef-habitat photos; this estimate is less certain.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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).

SiteDateHeadline (site page)ClaudeGold standardCoralNetDINOv3Coralscapes v1Coralscapes v2CoralSCOP uneditedCoralSCOP editedMean ± SD of CVRange
Cave Site2024-05-2051%published point count70hn=1551a,i,jn=17049c,jn=10631dn=15150en=17150en=17149gn=14951gn=14950 ± 13.731–70
Novice Bay2024-05-2039%published point count42hn=1539an=135—c,l23dn=12047en=13548en=13546g,mn=11147g,mn=11140 ± 11.323–47
Cat Island2024-05-2125%published point count36hn=1525a,bn=8125c,bn=8125dn=7134e,fn=8034e,fn=8033gn=7036gn=7033 ± 5.425–36
Jissah Point2024-05-2133%published point count26hn=1533a,bn=6133c,bn=6116dn=5432en=6132en=6135gn=5438gn=5429 ± 9.916–38
Daymaniyat 12024-05-2542%published point count51hn=1542a,bn=7942c,bn=7934dn=5627en=7930en=7923gn=5625gn=5633 ± 12.923–51
Daymaniyat 22024-05-2538%published point count49hn=1538a,bn=10638c,bn=10647dn=8440en=10637en=10642gn=8150gn=8147 ± 3.742–50
Daymaniyat 32024-05-2533%published point count34hn=1333a,k,bn=2833c,bn=2834dn=1820en=2827en=2818gn=1913gn=1925 ± 10.913–34
All sites (unweighted mean)—4437—303637353737 ± 5.830–44
Notes for Oman 2024 (17)
  • [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, unknown points kept (coral / (30 - WATER - FISH)). The manuscript's per-photo file 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 the author's own 135 Novice Bay and 65 Cave Site frames are manually annotated in CoralNet). 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; the per-image output no longer exists, so its denominator is unrecorded and the value cannot be audited. DINOv3 is not used in 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 considered reliable for these photos; shown, never averaged. Source: the Coralscapes v1/v2 run notes.
  • [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, 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.
  • [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 carry the author'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).
  • [l] Blank: every Novice Bay frame was manually annotated in CoralNet by the author, so CoralNet produced no robot label there.
  • [m] Novice Bay CoralSCOP = the finished edited run (111 paired frames).
  • 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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).

SiteDateHeadline (site page)ClaudeGold standardManually annotated in CoralNetCoralNetDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
Sandy Patch/Point (Mombasa)2024-03-1129%manualtoo few photos (n=8)k,c29an=117too few photos (n=8)b,c29d,en=11740fn=11738g,hn=11737g,hn=11726i,jn=11033 ± 10.026–40
Brain Coral (Mombasa)2024-03-1127%manual—l27an=136—27d,en=13643fn=13738g,hn=13739g,hn=13724i,jn=12533 ± 13.424–43
MMP snorkel site (Mombasa)2024-03-1440%manual—l40an=54—40d,en=5454fn=5433g,hn=5436g,hn=5423i,jn=4139 ± 21.823–54
Diani Beach (Diani)2024-03-1635%manual—l35an=50—34d,en=5039fn=5330g,hn=5330g,hn=5322i,jn=3530 ± 11.622–39
Reefolution nursery (Shimoni-Wasini)2024-03-2040%manual—l40an=258—40d,en=25844fn=26432g,hn=26433g,hn=26416i,jn=14130 ± 19.416–44
Coral Garden (Kisite National Park)2024-03-2155%manual44kn=2755an=17851bn=2746d,en=17846fn=17944g,hn=17944g,hn=17921i,jn=15037 ± 13.921–46
Outer Kisite (Kisite National Park)2024-03-2137%manual—l37an=85—36d,en=8343fn=8345g,hn=8344g,hn=8317i,jn=6830 ± 17.917–43
Dolphin Point (Kisite National Park)2024-03-2140%manual—l40an=108—40d,en=10842fn=10835g,hn=10834g,hn=10827i,jn=9635 ± 10.927–42
Kisite National Park-site 4 (Kisite National Park)2024-03-2146%manual—l46an=71—46d,en=7136fn=7133g,hn=7134g,hn=7118i,jn=5027 ± 12.918–36
Lower Mpunguti-site 1 (Kisite National Park)2024-03-2118%manualtoo few photos (n=1)k,c18an=36too few photos (n=1)b,c17d,en=3637fn=3617g,hn=3616g,hn=3621i,jn=2729 ± 11.921–37
Lower Mpunguti-site 2 (Kisite National Park)2024-03-2134%manual—l34an=27—34d,en=2743fn=2719g,hn=2717g,hn=2717i,jn=2730 ± 18.217–43
Kilifi Bay (Kilifi)2024-03-2625%manual—l25an=134—26d,en=13431fn=13622g,hn=13621g,hn=13619i,jn=12325 ± 9.119–31
Bofa (Kilifi)2024-03-2632%manual—l32an=97—34d,en=9748fn=10429g,hn=10430g,hn=10427i,jn=8737 ± 14.827–48
Round Reef (Watamu)2024-03-2718%manual13kn=17518an=17516bn=17514d,en=17528fn=17512g,hn=17514g,hn=17516i,jn=1619 ± 7.913–28
Uyombo (Watamu)2024-03-2825%manual23kn=37425an=37523bn=37425d,en=37537fn=37419g,hn=37422g,hn=37429i,jn=2430 ± 6.923–37
North Canyon (Watamu)2024-03-2922%manual—l22an=83—22d,en=8320fn=8215g,hn=8215g,hn=8219i,jn=7920 ± 0.819–20
Trumpet Area (Watamu)2024-03-2921%manual—l21an=193—21d,en=19335fn=19321g,hn=19322g,hn=19321i,jn=17628 ± 10.321–35
Chipopo (Pate (Lamu))2024-04-0313%manual7.3kn=4013an=408.6bn=4017d,en=4022fn=406.5g,hn=407.3g,hn=4012i,jn=3514 ± 7.37.3–22
Pilot Nursery (Pate (Lamu))2024-04-0314%manual14kn=2114an=2112bn=2119d,en=2133fn=2116g,hn=2118g,hn=2111i,jn=1619 ± 11.611–33
Ocean Trust Nursery (Pate (Lamu))2024-04-0329%manual23kn=5729an=6223bn=5728d,en=6240fn=6320g,hn=6318g,hn=6311i,jn=4125 ± 14.811–40
All sites (unweighted mean)——30—303826272029 ± 12.920–38
Notes for Kenya 2024 (16)
  • [a] Gold = published cover, Mayfield 2025 Environments 12:276, Appendix A (CoralNet point annotation; denominator coral / (30 − WATER − FISH), unknown points kept); n = photos in the published set. The CoralNet export holds 22,016 of the author's points and 47,644 robot points.
  • [b] The author's own labels on exactly the Claude frames, all-30 (like-for-like with the Claude cell).
  • [c] Too few photos (under 10) at this site for a value: not shown and not averaged (photo-count rule: 10 or more photos for a value, 20 or more for a confident one).
  • [d] CoralNet = the source's Machine suggestion 1 on ALL unedited frames. The classifier was trained on the author's own labels of these same images, so it is circular with the gold standard (not an independent check).
  • [e] Same measurement as (or circular with) the reference: shown, not averaged.
  • [f] DINOv3 on ALL unedited photos at the site, coral / (30 − WATER − FISH), unknown points kept.
  • [g] Coralscapes run 2026-09-18 on UNEDITED originals; site values as on the site pages.
  • [h] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not considered reliable for these photos: shown, never averaged. Area fraction, not a point count.
  • [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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [j] CoralSCOP full run on the unedited originals (1,444 of 1,450 frames scored); site values as on the site pages.
  • [k] Claude = blind point calls, 30 CoralNet points per frame, on UNEDITED originals cover = H/30 (all-30, Unk in the denominator; not the gold standard's denominator); batches 1-8, batch 9 caller C only, batch 10 (103 of 109 called).
  • [l] 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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.

SiteDateHeadline (site page)ClaudeDINOv3CoralSCOPMean ± SD of CVRange
Raiketa2024-01-2946%Claude46a,dn=20040a,bn=20044a,cn=20044 ± 3.340–46
Inner Reef2024-01-3045%Claude45a,dn=19338a,bn=19353a,cn=19345 ± 7.138–53
Haruine2024-01-3041%Claude41a,dn=18336a,bn=18339a,cn=18338 ± 2.836–41
Biqueli Wall2024-01-3027%Claude27a,dn=14730a,bn=14738a,cn=14732 ± 5.727–38
Compass Reef2024-01-31–02-0138%Claude38a,dn=9931a,bn=9949a,cn=9939 ± 9.331–49
Cabbage Patch2024-01-3137%Claude37a,dn=11531a,bn=11538a,cn=11535 ± 3.931–38
Pinnacle2024-01-3120%Claude20a,dn=5326a,bn=5339a,cn=5328 ± 9.620–39
All sites (unweighted mean)—36334337 ± 4.933–43
Notes for East Timor (Atauro) 2024 (8)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [b] 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's site cover table).
  • [c] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [d] Claude = blind 7-call point count on the EDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator; mean across photos.
  • 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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

Bahamas 2022 (New Providence)

2 sites · 177 photos

No gold standard.

SiteDateHeadline (site page)ClaudeCoralSCOPMean ± SD of CVRange
King Reef (New Providence, west end)2022-10-139.1%Claude9.1a,cn=6015a,bn=6012 ± 4.59.1–15
Goulding Cay (New Providence)2022-10-137.9%Claude7.9a,cn=11720a,bn=11714 ± 8.27.9–20
All sites (unweighted mean)—8.51713 ± 6.38.5–17
Notes for Bahamas 2022 (New Providence) (7)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [c] Claude = blind 7-call point count on the EDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator; mean across photos.
  • 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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

Turks & Caicos 2022

11 sites · 791 photos

No gold standard.

SiteDateHeadline (site page)ClaudeCoralSCOP
Eel Garden†2022-07-3120%Claude20c,d,en=2137a,b
Shark Hotel2022-07-3115%Claude15c,dn=4419a,b
The Amphitheatre2022-08-0112%Claude12c,dn=5524a,b
Thunderdome2022-08-018.6%Claude8.6c,dn=5417a,bn=68
Brandywine2022-08-0211%Claude11c,dn=9420a,b
Coral Stairway2022-08-0214%Claude14c,dn=15024a,b
G Spot2022-08-034.6%Claude4.6c,dn=4716a,b
Rock and Roll2022-08-035.9%Claude5.9c,dn=9019a,b
Boat Cove2022-08-0411%Claude11c,dn=7425a,b
Spanish Anchor2022-08-048.1%Claude8.1c,dn=8820a,b
The Gully2022-08-057.1%Claude7.1c,dn=7418a,b
All sites (unweighted mean)—1122
Notes for Turks & Caicos 2022 (10)
  • [a] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [b] CoralSCOP (mask-sum area index) = the value the map published before the Claude arm, from the earlier folder-level CoralSCOP run; NOT frame-matched to the Claude photos (n = that run's frames where recorded). Comparison only, never averaged.
  • [c] Claude = blind 7-call point count on the EDITED photos where they exist (Eel Garden: edited and unedited photos, †) (20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator); value = the site's headline as printed on its page.
  • [d] Turks and Caicos: the site value is the mean of its two dive means; n = photos over both dives.
  • [e] THIN SAMPLE (under 30 frames at this site; marked in the headline column only): 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 in gray italic but never averaged.
  • † Includes unedited photographs: edited versions of these photos do not exist.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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

Dominica 2022

4 sites · 532 photos

No gold standard.

SiteDateHeadline (site page)ClaudeDINOv3CoralSCOP uneditedCoralSCOP edited
Champagne Reef2022-06-2229%Claude29dn=153—a33bn=19133b,c
Coral Gardens2022-06-2211%Claude11dn=145—a—e27b,c
Scott's Head Pinnacle2022-06-2215%Claude15dn=19226f—e45b,c
Toucari Bay2022-07-0328%Claude28dn=42—a—e33b,gn=42
All sites (unweighted mean)—21——34
Notes for Dominica 2022 (11)
  • [a] Not run on these photos.
  • [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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [c] CoralSCOP (mask-sum area index) = the value the map published before the Claude arm, from the earlier folder-level CoralSCOP run; NOT frame-matched to the Claude photos (n = that run's frames where recorded). Comparison only, never averaged.
  • [d] Claude = blind 7-call point count on the EDITED photos (20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator); value = the site's headline as printed on its page.
  • [e] Not applicable: only the edited photos are used for this region.
  • [f] DINOv3 = the value behind the earlier two-method headline; comparison only.
  • [g] CoralSCOP here = the site page's frame-matched row: run on exactly the photos behind the Claude cell (per-frame values joined by file; capped at 100% per photo; zero-filled). Comparison 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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. Lucia 2022

4 sites · 312 photos

No gold standard.

SiteDateHeadline (site page)ClaudeCoralSCOP
Anse Chastanet2022-06-25—7.8a,bn=36—
Soufriere Pinnacles2022-06-2611%Claude11an=12930c,dn=136
Superman's Flight2022-06-269.0%Claude9.0an=10631c,dn=110
Pigeon Island2022-06-3012%Claude12an=4128c,dn=42
All sites (unweighted mean)—1130
Notes for St. Lucia 2022 (8)
  • [a] Claude = blind 7-call point count on the EDITED photos (20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator); value = the site's headline as printed on its page.
  • [b] No coral-cover estimate is published for this site (its photographs show jetty pilings, not reef habitat), so these values are a comparison only, left out of the row mean and the All-sites row.
  • [c] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [d] CoralSCOP (mask-sum area index) = the value the map published before the Claude arm, from the earlier folder-level CoralSCOP run; NOT frame-matched to the Claude photos (n = that run's frames where recorded). Comparison only, never averaged.
  • 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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

Dry Tortugas 2021

4 sites · 229 photos

Agency survey column = each site's headline: NOAA NCRMP ReefBudget 2024 at Bird Key Reef (6 fixed transects), FWC DTCREMP 2021 permanent stations elsewhere. Claude and CoralSCOP ran on the author's June 2021 photos (comparison only).

SiteDateHeadline (site page)ClaudeAgency survey (headline)CoralSCOP uneditedCoralSCOP edited
Bird Key Reef, Dry Tortugas2021-06-25–285.1%transect28e,fn=1235.1an=6—b42c,dn=114
White Shoal, Dry Tortugas2021-06-263.3%transect26e,fn=553.3an=422c,dn=35—g
Black Coral Rock, Dry Tortugas2021-06-2812%transect18e,fn=2512an=435c,dn=15—g
Texas Reef, Dry Tortugas2021-06-286.8%transect21e,fn=266.8an=4—b39c,hn=26
All sites (unweighted mean)—236.9——
Notes for Dry Tortugas 2021 (12)
  • [a] Agency survey, the site headline: Bird Key Reef = NOAA NCRMP ReefBudget 2024, 6 fixed transects (NCEI 0311367); other sites = FWC DTCREMP 2021 permanent stations. n = transects / stations, not photos.
  • [b] Not run on unedited photos at this site.
  • [c] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [d] CoralSCOP v2 (mask-sum), a folder-level run on Anderson's June 2021 dive folders: not frame-matched to the Claude photos (n = that run's frames). Comparison only.
  • [e] Claude = blind 7-call point count on the EDITED photos (20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator); value = the site's headline as printed on its page.
  • [f] At Dry Tortugas the Claude value is a comparison only (survey headline); the high values reflect colony-focused photo selection.
  • [g] Not run on edited photos at this site.
  • [h] CoralSCOP here = the site page's frame-matched row: run on exactly the photos behind the Claude cell (per-frame values joined by file; capped at 100% per photo; zero-filled). Comparison 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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

Barbados 2020

2 sites · 152 photos

No gold standard.

SiteDateHeadline (site page)ClaudeCoralSCOP
Close Encounters2020-01-1819%Claude19cn=6437a,bn=62
G Spot2020-01-1829%Claude29cn=8831a,bn=85
All sites (unweighted mean)—2434
Notes for Barbados 2020 (7)
  • [a] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [b] CoralSCOP (mask-sum area index) = the value the map published before the Claude arm, from the earlier folder-level CoralSCOP run; NOT frame-matched to the Claude photos (n = that run's frames where recorded). Comparison only, never averaged.
  • [c] Claude = blind 7-call point count on the EDITED photos (20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator); value = the site's headline as printed on its page.
  • 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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. Vincent and the Grenadines 2019 (Mayreau)

4 sites · 148 photos

No gold standard.

SiteDateHeadline (site page)ClaudeCoralSCOP
Mayreau Gardens (CO2 vent site)2019-09-26–10-0123%Claude23a,dn=6638a,b,cn=66
Mayreau Gardens2019-09-26–307.0%Claude7.0a,dn=3321a,b,cn=33
Mayreau Gardens (second vent)2019-09-2630%Claude30a,dn=2139a,b,cn=21
Mayreau Shipwreck2019-10-014.9%Claude4.9a,dn=2818a,b,cn=28
All sites (unweighted mean)—1629
Notes for St. Vincent and the Grenadines 2019 (Mayreau) (8)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [c] CoralSCOP here = the site page's frame-matched row: run on exactly the photos behind the Claude cell (per-frame values joined by file; capped at 100% per photo; zero-filled). Comparison only.
  • [d] Claude = blind 7-call point count on the EDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator; mean across photos.
  • 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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

Flower Garden Banks 2019

1 site · 148 photos

Agency survey column = NOAA FGBNMS 2024 East Flower Garden Bank study-site transects, the site headline. Claude and CoralSCOP ran on the author's June 2019 photos (comparison only).

SiteDateHeadline (site page)ClaudeAgency survey (headline)CoralSCOP
East Flower Garden Bank, Buoy 32019-06-03–0450%transect51d,en=14850an=1658b,cn=148
Notes for Flower Garden Banks 2019 (9)
  • [a] NOAA FGBNMS long-term monitoring, EFGB study-site random photo transects (CPCe, 500 points), 16 July 2024, NCEI 0308032: the site headline. n = transects.
  • [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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [c] CoralSCOP here = the site page's frame-matched row: run on exactly the photos behind the Claude cell (per-frame values joined by file; capped at 100% per photo; zero-filled). Comparison only.
  • [d] Claude = blind 7-call point count on the EDITED photos (20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator); value = the site's headline where Claude is the headline.
  • [e] Comparison only at a survey site: the author's June 2019 photos; the NOAA FGBNMS 2024 transect value is the headline.
  • 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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.

SiteDateHeadline (site page)ClaudeDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
Siaba Besar2019-01-19–2653%Claude53a,en=12132a,bn=12164a,cn=12166a,cn=12142a,dn=12142 ± 10.732–53
Mawan2019-01-19–2614%Claude14a,en=15020a,bn=15049a,cn=14853a,cn=14930a,dn=14821 ± 7.914–30
Siaba Kecil2019-01-19–2625%Claude25a,en=7230a,bn=7264a,cn=7264a,cn=7239a,dn=7232 ± 7.125–39
Hot Rocks (Sangeang)2019-01-209.1%Claude9.1a,en=12026a,bn=12057a,cn=12054a,cn=12038a,dn=12024 ± 14.69.1–38
Deep Purple2019-01-2026%Claude26a,en=6030a,bn=6058a,cn=6055a,cn=6039a,dn=6032 ± 6.826–39
The Cauldron / Shotgun2019-01-2126%Claude26a,en=6221a,bn=6251a,cn=6251a,cn=6234a,dn=6227 ± 6.421–34
Tatawa Besar2019-01-22–2722%Claude22a,en=15524a,bn=15553a,cn=15555a,cn=15534a,dn=15527 ± 6.622–34
Makassar Reef / Manta Point2019-01-223.3%Claude3.3a,en=5910a,bn=5933a,cn=5231a,cn=489.1a,dn=577.5 ± 3.73.3–10
Bonsai Rock2019-01-2218%Claude18a,en=6627a,bn=6652a,cn=6653a,cn=6638a,dn=6627 ± 10.118–38
Padar Island (Padar Kecil)2019-01-238.1%Claude8.1a,en=5624a,bn=5651a,cn=5651a,cn=5646a,dn=5626 ± 18.98.1–46
Secret Garden2019-01-2313%Claude13a,en=3925a,bn=3951a,cn=3955a,cn=3938a,dn=3925 ± 12.613–38
Three Sisters (Tiga Dara)2019-01-238.3%Claude8.3a,en=3420a,bn=3429a,cn=3430a,cn=3428a,dn=3419 ± 10.18.3–28
Pink Beach (Pantai Merah)2019-01-2425%Claude25a,en=7130a,bn=7166a,cn=7167a,cn=7137a,dn=7131 ± 6.325–37
Loh Sera2019-01-2430%Claude30a,en=11835a,bn=11873a,cn=11873a,cn=11846a,dn=11737 ± 8.130–46
Pengah Kecil2019-01-2530%Claude30a,en=3731a,bn=3762a,cn=3760a,cn=3736a,dn=3732 ± 3.530–36
Police Corner2019-01-2522%Claude22a,en=7830a,bn=7865a,cn=7868a,cn=7852a,dn=7834 ± 15.622–52
Tatawa Kecil2019-01-2732%Claude32a,en=8329a,bn=8357a,cn=8357a,cn=8338a,dn=8333 ± 4.629–38
All sites (unweighted mean)—212655553728 ± 7.821–37
Notes for Komodo 2019 (9)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [b] 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's site cover table).
  • [c] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not considered reliable for these photos: shown, never averaged. Area fraction, not a point count.
  • [d] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [e] Claude = blind 7-call point count on the EDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator; mean across photos.
  • 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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.

SiteDateHeadline (site page)ClaudeDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
Seruanus (Triton Bay)2018-10-149.1%Claude9.1a,en=6619a,bn=6644a,cn=6642a,cn=6631a,dn=6619 ± 10.89.1–31
Little Komodo (Triton Bay)2018-10-1412%Claude12a,en=7323a,bn=7544a,cn=7345a,cn=7231a,dn=7522 ± 9.812–31
Last Frontier (Triton Bay)2018-10-157.8%*Claude7.8a,en=1410a,bn=1517a,cn=1415a,cn=1416a,dn=1511 ± 4.27.8–16
Bo Rainbow (Triton Bay)2018-10-1513%Claude13a,en=4722a,bn=4760a,cn=4762a,cn=4742a,dn=4726 ± 15.013–42
Karang Paradise2018-10-1644%Claude44a,en=13229a,bn=13666a,cn=12768a,cn=12731a,dn=13435 ± 8.429–44
Pulau Koon dive 22018-10-1725%Claude25a,en=5219a,bn=5349a,cn=5147a,cn=5133a,dn=5326 ± 6.619–33
Pulau Koon dive 32018-10-1730%Claude30a,en=9831a,bn=9875a,cn=9876a,cn=9846a,dn=9836 ± 9.330–46
Shadow Reef2018-10-18–2017%Claude17a,en=8518a,bn=10744a,cn=8347a,cn=8118a,dn=10618 ± 0.717–18
Nudi Rock2018-10-1930%Claude30a,en=9726a,bn=10061a,cn=9862a,cn=9832a,dn=10029 ± 3.326–32
Little Juliet2018-10-1926%*Claude26a,en=1324a,bn=1348a,cn=1251a,cn=1228a,dn=1326 ± 2.024–28
Anchovy2018-10-2032%Claude32a,en=6426a,bn=6455a,cn=6455a,cn=6438a,dn=6432 ± 6.126–38
Two Tree Island2018-10-207.2%Claude7.2a,en=7719a,bn=8831a,cn=8231a,cn=7519a,dn=8815 ± 6.67.2–19
My Reef2018-10-2133%Claude33a,en=5432a,bn=5455a,cn=5452a,cn=5439a,dn=5435 ± 3.932–39
Melissa's Garden (Fam/Penemu)2018-10-2141%Claude41a,en=8935a,bn=8971a,cn=8970a,cn=8941a,dn=8939 ± 3.335–41
Blue Magic2018-10-22–2311%Claude11a,en=4019a,bn=4229a,cn=4036a,cn=4027a,dn=4219 ± 8.211–27
Mioskon2018-10-2212%Claude12a,en=4724a,bn=4743a,cn=4745a,cn=4733a,dn=4723 ± 10.412–33
All sites (unweighted mean)—222350503226 ± 5.222–32
Notes for Raja Ampat 2018 (incl. Triton Bay) (10)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [b] 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's site cover table).
  • [c] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not considered reliable for these photos: shown, never averaged. Area fraction, not a point count.
  • [d] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [e] Claude = blind 7-call point count on the EDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator; mean across photos.
  • 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 in gray italic but never averaged.
  • * The site's headline coral cover rests on only 10–19 reef-habitat photos; this estimate is less certain.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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.

SiteDateHeadline (site page)ClaudeDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
Lighthouse2018-04-25–2629%Claude29a,en=13324a,bn=5636a,cn=5639a,cn=5636a,dn=13230 ± 6.324–36
Lava Flow (Gunung Api)2018-04-2555%Claude55a,en=19035a,bn=19076a,cn=19076a,cn=19037a,dn=18942 ± 10.735–55
Tanjung Pisang (Banana Island)2018-04-2536%Claude36a,en=18031a,bn=18065a,cn=18067a,cn=18040a,dn=18036 ± 4.531–40
Airport (Banda Ara)2018-04-2638%Claude38a,en=7236a,bn=7271a,cn=7273a,cn=7245a,dn=7240 ± 5.036–45
Goat Hall (Pulau Ai)2018-04-2736%Claude36a,en=14433a,bn=14462a,cn=14463a,cn=14437a,dn=14435 ± 2.433–37
Shark Point (Pulau Ai)2018-04-2733%Claude33a,en=8830a,bn=8866a,cn=8864a,cn=8834a,dn=8833 ± 2.130–34
Batu Kapal2018-04-2816%Claude16a,en=7824a,bn=7841a,cn=7843a,cn=7826a,dn=7822 ± 5.316–26
All sites (unweighted mean)—353059613734 ± 3.230–37
Notes for Maluku 2018 - Banda (9)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [b] 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's site cover table).
  • [c] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not considered reliable for these photos: shown, never averaged. Area fraction, not a point count.
  • [d] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [e] Claude = blind 7-call point count on the EDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator; mean across photos.
  • 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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

25 sites · 684 photos

No gold standard.

SiteDateHeadline (site page)ClaudeDINOv3CoralSCOP
Kurumba Faru (checkout dive)2017-05-13——15a22b,a
Back Faru2017-05-14–216.4%Claude6.4cn=7414a15b,a
Bodu Hithi Thila2017-05-14–158.7%Claude8.7cn=3015a16b,a
Feydoo Wall2017-05-15——24a21b,a
Georgia's Reef2017-05-1522%Claude22cn=3417a23b,a
Miyaro Faru2017-05-1624%Claude24c,dn=2017a23b,a
Vaadhoo Caves2017-05-164.7%*Claude4.7c,dn=187.3a15b,a
Veligandu2017-05-1617%*Claude17c,dn=1813a22b,a
Ghuli Faru2017-05-175.4%*Claude5.4c,dn=1710a13b,a
Miyaru Bay2017-05-17–1819%*Claude19c,dn=1915a15b,a
Big Sand Bank2017-05-1810%Claude10c,dn=2116a20b,a
Kuda Giri Wreck2017-05-18–2217%Claude17cn=5618a21b,a
Hodeli Bodu Giri2017-05-1934%Claude34cn=3421a25b,a
Kandooma Thila2017-05-19–226.3%*Claude6.3c,dn=1813a7.3b,a
Fish Tank2017-05-21——11a4.9b,a
Lankan Faru2017-05-21–2213%Claude13c,dn=209.2a22b,a
Camel Point2017-05-2321%Claude21cn=7623a26b,a
Lohosfushi Corner2017-05-2312%Claude12c,dn=2414a16b,a
Dhidhoo Beyru2017-05-2420%Claude20cn=42——
Dhigga Thila2017-05-248.4%Claude8.4c,dn=2522a29b,a
Lux Beyru2017-05-2421%Claude21c,dn=26——
Dhega Thila2017-05-2520%Claude20cn=4525a21b,a
Fesdhoo wreck2017-05-252.1%Claude2.1c,dn=2015a14b,a
Moofushi Kandu2017-05-257.4%Claude7.4c,dn=2421a11b,a
Fish House (Fish Head)2017-05-263.0%Claude3.0c,dn=2312a6.3b,a
Notes for Maldives #2 2017 (10)
  • [a] DINOv3 / CoralSCOP = the folder-level runs behind the earlier two-method map headline (staged subset, not frame-matched to the Claude photos); comparison only, never averaged.
  • [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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [c] Claude = blind 7-call point count on the EDITED photos (20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator); value = the site's headline where Claude is the headline.
  • [d] THIN SAMPLE (under 30 frames at this site; marked in the headline column only): 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 in gray italic but never averaged.
  • Site pages: one row per map pin, with the values its site page prints. Big Sand Bank was also surveyed in January 2016: this row is the May 2017 survey, and its page also gives the January 2016 value. Checkout dive, Feydoo Wall, and Fish Tank have too few photos (under 10) for a Claude value.
  • * The site's headline coral cover rests on only 10–19 reef-habitat photos; this estimate is less certain.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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

Bonaire 2017

12 sites · 609 photos

No gold standard.

SiteDateHeadline (site page)ClaudeCoralSCOP
Between Pink Beach and White Slave2017-02-208.6%Claude8.6cn=3422a,bn=38
The Cliff2017-02-2031%Claude31cn=5828a,bn=61
Karpata2017-02-2136%Claude36cn=4132a,bn=49
Oil Slick Leap2017-02-2126%Claude26c,dn=2928a,bn=33
Vista Blue2017-02-2111%Claude11cn=6030a,bn=68
Alice in Wonderland2017-02-2219%Claude19cn=4920a,bn=53
La Dania's Leap2017-02-2225%Claude25cn=5627a,bn=67
The Chairs2017-02-2222%Claude22cn=5328a,bn=51
Chogogo2017-02-2410%Claude10c,dn=2424a,bn=32
Hilma Hooker2017-02-2419%Claude19cn=3119a,bn=53
Wayaka II2017-02-2424%Claude24cn=5224a,bn=58
Weber's Joy2017-02-2421%Claude21cn=3925a,bn=46
All sites (unweighted mean)—2126
Notes for Bonaire 2017 (8)
  • [a] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [b] DINOv3 / CoralSCOP = the folder-level runs behind the earlier two-method map headline (staged subset, not frame-matched to the Claude photos); comparison only, never averaged.
  • [c] Claude = blind 7-call point count on the EDITED photos (20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator); value = the site's headline where Claude is the headline.
  • [d] THIN SAMPLE (under 30 frames at this site; marked in the headline column only): 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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)

10 sites · 286 photos

No gold standard.

SiteDateHeadline (site page)ClaudeDINOv3Coralscapes v2CoralSCOP
Alberguin wreck2017-01-2020%Claude20a,fn=3220a,bn=2334a,cn=3231a,d,en=32
Reef 1 (bay snorkel)2017-01-2029%*Claude29a,fn=10too few photos (n=8)a,b,h33a,cn=1047a,d,en=10
island-hopping stop 1 (San Miguel area)2017-01-2167%Claude67a,fn=4140a,bn=3791a,cn=4140a,d,en=41
island-hopping stop 22017-01-2139%Claude39a,fn=2031a,bn=2060a,cn=2043a,dn=20
Nat Nat (Bacuit Archipelago)2017-01-2414%*Claude14a,fn=1320a,bn=1221a,cn=1232a,d,en=13
Abdens (Bacuit Archipelago)2017-01-2442%Claude42a,fn=2023a,bn=2067a,cn=2034a,dn=20
Paglugaban Rocks (Bacuit Archipelago)2017-01-2434%Claude34a,fn=9827a,bn=9855a,cn=9831a,dn=98
Taiei Maru (wreck), Busuanga2017-01-27—3.4a,f,gn=1412a,bn=1412a,cn=1420a,dn=14
Coral Gardens, Busuanga2017-01-2741%Claude41a,fn=2535a,bn=2567a,cn=2526a,dn=25
Twin Lagoon (island-hopping snorkel)2017-01-288.4%*Claude8.4a,fn=1315a,bn=1320a,cn=1319a,dn=13
All sites (unweighted mean)—33—5034
Notes for Philippines 2017 (Palawan) (14)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [b] 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's site cover table).
  • [c] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not considered reliable for these photos: shown, never averaged. Area fraction, not a point count.
  • [d] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [e] CoralSCOP here = the site page's frame-matched row: run on exactly the photos behind the Claude cell (per-frame values joined by file; capped at 100% per photo; zero-filled). Comparison only.
  • [f] Claude = blind 7-call point count on the EDITED photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator; mean across photos.
  • [g] No coral-cover estimate is published for this site (a wreck, not reef habitat), so these values are a comparison only, left out of the row mean and the All-sites row.
  • [h] Too few photos (under 10) at this site for a value: not shown and not averaged (photo-count rule: 10 or more photos for a value, 20 or more for a confident one).
  • 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 in gray italic but never averaged.
  • Headline: Claude is the only coral-cover figure on the site pages, report cards and map pop-ups; DINOv3, Coralscapes v2 and CoralSCOP are comparison columns only, and no mean across methods is given. Akitsushima and Green Lagoon (4 usable photos each) have no headline: too few images to estimate coral cover. The Taiei Maru wreck has no headline (wreck habitat; coral cover not estimated); its row is a comparison only.
  • * The site's headline coral cover rests on only 10–19 reef-habitat photos; this estimate is less certain.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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)

6 sites · 357 photos

No gold standard.

SiteDateHeadline (site page)ClaudeDINOv3Coralscapes v2CoralSCOP uneditedCoralSCOP edited
Ligpo Island2016-09-3013%Claude13a,en=8021a,bn=7461a,cn=80—33a,dn=74
Dive and Trek2016-09-3024%Claude24a,en=7327a,bn=6540a,cn=73—34a,dn=65
Monte Carlo Hotel2016-09-3029%Claude29a,en=5724a,bn=5234a,cn=57—20a,d,fn=57
Kirby's Rock2016-10-018.6%Claude8.6a,en=3922a,bn=3740a,cn=39—27a,d,fn=39
Beatriz Island (Beatrice Rock)†2016-10-014.6%Claude4.6a,en=6517a,bn=6348a,cn=6519a,d,fn=65—
Sombrero Island†2016-10-0117%Claude17a,en=4324a,bn=3746a,cn=4327a,d,fn=43—
All sites (unweighted mean)—162345——
Notes for Philippines 2016 (Anilao) (12)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [b] 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's site cover table).
  • [c] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not considered reliable for these photos: shown, never averaged. Area fraction, not a point count.
  • [d] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [e] Claude = blind 7-call point count on the EDITED photos where they exist (Beatriz Island and Sombrero Island: edited and unedited photos, †), 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator; mean across photos.
  • [f] CoralSCOP here = the site page's frame-matched row: run on exactly the photos behind the Claude cell (per-frame values joined by file; capped at 100% per photo; zero-filled). Comparison 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 in gray italic but never averaged.
  • † Beatriz Island and Sombrero Island include unedited photographs (6 edited + 59 unedited originals at Beatriz, 14 + 29 at Sombrero): edited versions of most of those dives' photos do not exist. Every method in the row ran on the same photos.
  • Headline: Claude is the only coral-cover figure on the site pages, report cards and map pop-ups; DINOv3, Coralscapes v2 and CoralSCOP are comparison columns only, and no mean across methods is given.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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

Dongsha 2016

3 sites · 168 photos

No gold standard.

SiteDateHeadline (site page)ClaudeCoralSCOP uneditedCoralSCOP edited
Site 13†2016-04-0454%Claude54cn=3735a,bn=37—
Site 6 (branching)†2016-04-0557%Claude57cn=115—d39a,en=88
Site 7 (branching)†2016-04-0530%*Claude30c,fn=1625a,bn=16—
All sites (unweighted mean)—47——
Notes for Dongsha 2016 (12)
  • [a] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [b] CoralSCOP here = the site page's frame-matched row: run on exactly the photos behind the Claude cell (per-frame values joined by file; capped at 100% per photo; zero-filled). Comparison only.
  • [c] Claude = blind 7-call point count on the EDITED photos where they exist (Sites 13 and 7: unedited photos, †) (20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator); value = the site's headline as printed on its page.
  • [d] Not run on unedited photos at this site.
  • [e] CoralSCOP (mask-sum area index) = the value the map published before the Claude arm, from the earlier folder-level CoralSCOP run; NOT frame-matched to the Claude photos (n = that run's frames where recorded). Comparison only, never averaged.
  • [f] THIN SAMPLE (under 30 frames at this site; marked in the headline column only): 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 in gray italic but never averaged.
  • * The site's headline coral cover rests on only 10–19 reef-habitat photos; this estimate is less certain.
  • † Includes unedited photographs: edited versions of these photos do not exist.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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

13 sites · 230 photos

Transect column = the author's July 2016 PIT transects (after the 2016 bleaching), the headline at the seven transect sites. Elsewhere the headline is Claude or an earlier computer-vision value.

SiteDateHeadline (site page)ClaudeTransect survey (headline)DINOv3CoralSCOP
Little Sandy Bank2016-01-14–1931%transect21d,hn=4931gn=1023b17c,b
Raaeundi Reef2016-01-158.0%transect—8.0gn=524b19c,b
Veli2016-01-16fore reef 16% · lagoon 5.3%transect44d,hn=1216gn=831b39c,b
Stage2016-01-1728%transect—28gn=2023b38c,b
Anantara Dhigu house reef2016-01-1822%Claude22dn=44—a24b38c,b
Maafushi2016-01-1816%transect—16gn=621b40c,b
Gaagandufaruhuraa2016-01-2138%transect—38gn=17—e45cn=19
Anantara Kihavah house reef & Golden Wall2016-01-22–2517%CV mean——a15b19c,b
Kashidhoo Faru2016-01-2227%transect—27gn=13—e26cn=19
Baa Atoll site 82016-01-2345%Claude45dn=49—a—e—e
Baa Atoll site 92016-01-2439%Claude39d,fn=20—a—e—e
Four Seasons Landaa Giraavaru2016-01-2462%Claude62dn=38—a30b45c,b
Baa Atoll site 102016-01-2537%*Claude37d,fn=18—a—e—e
Notes for Maldives #1 2016 (14)
  • [a] No July 2016 transect at this site.
  • [b] DINOv3 / CoralSCOP = the folder-level runs behind the earlier two-method map headline (staged subset, not frame-matched to the Claude photos); comparison only, never averaged.
  • [c] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [d] Claude = blind 7-call point count on the EDITED photos (20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator); value = the site's headline where Claude is the headline.
  • [e] Not recorded for this site.
  • [f] THIN SAMPLE (under 30 frames at this site; marked in the headline column only): indicative only.
  • [g] The author's HARP / Coral Reef CPR long-term monitoring PIT transects, July 2016 (after the bleaching): the site headline. n = transects, not photos. Veli: fore reef (cell) and lagoon shown separately.
  • [h] Comparison only at a transect site: January 2016 photos, BEFORE the 2016 bleaching - compare with the April 2016 transects on the site page, not the July headline.
  • 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 in gray italic but never averaged.
  • Site pages: one row per map pin, with the values its site page prints. Gaagandufaruhuraa and Kashidhoo Faru carry Anderson's July 2016 transect headline; Anantara Kihavah has no transect and no Claude photos and keeps its earlier computer-vision headline.
  • * The site's headline coral cover rests on only 10–19 reef-habitat photos; this estimate is less certain.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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 · 1,132 photos

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

SiteDateHeadline (site page)ClaudeGold standardManually annotated in CoralNetCoralNetDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
BIPB602015-04-1442%transect30a,in=3842a,b39a,cn=38—33a,fn=3855a,gn=3857a,gn=3842a,hn=3835 ± 6.330–42
BIPB612015-04-1420%transect17a,in=1420a,b18a,cn=14—22a,fn=1441a,gn=1443a,gn=1432a,hn=1424 ± 7.917–32
BIPB622015-04-1459%transect34a,in=1959a,b48a,cn=19—35a,fn=1958a,gn=1858a,gn=1832a,hn=1933 ± 1.832–35
BIPB632015-04-1556%transect35a,in=2956a,b47a,cn=29—33a,fn=2966a,gn=2967a,gn=2938a,hn=2936 ± 2.633–38
BIPB642015-04-1533%transect23a,in=2133a,b23a,cn=21—26a,fn=2134a,gn=2141a,gn=2133a,hn=2127 ± 5.423–33
BIPB652015-04-1547%transect30a,in=1947a,b33a,cn=19—32a,fn=1952a,gn=1955a,gn=1928a,hn=1930 ± 2.428–32
BIPB662015-04-167.8%transecttoo few photos (n=6)a,i,d7.8a,btoo few photos (n=6)a,c,d—too few photos (n=6)a,f,dtoo few photos (n=6)a,g,dtoo few photos (n=6)a,g,dtoo few photos (n=6)a,h,d——
BIPB672015-04-1646%transect26a,in=4146a,b36a,cn=41—34a,fn=4149a,gn=4152a,gn=4129a,hn=4130 ± 3.926–34
BIPB682015-04-1659%transect37a,in=4359a,b45a,cn=43—39a,fn=4372a,gn=4374a,gn=4339a,hn=4338 ± 1.237–39
BIPB692015-04-16–05-0273%transect39a,in=11673a,b54a,cn=110too few photos (n=6)a,e,d47a,fn=11674a,gn=11675a,gn=11643a,hn=11643 ± 4.239–47
BIPB702015-04-17–05-0170%transect35a,in=5570a,b45a,cn=2845a,en=2740a,fn=5557a,gn=5559a,gn=5536a,hn=5539 ± 4.735–45
BIPB722015-04-18–05-0177%transect44a,in=7577a,b51a,cn=4156a,en=3449a,fn=7564a,gn=7569a,gn=7540a,hn=7547 ± 6.940–56
BIPB732015-04-1850%transect38a,in=4550a,b—45a,en=4545a,fn=4548a,gn=4551a,gn=4547a,hn=4544 ± 4.238–47
BIPB742015-04-18–05-0163%transect51a,in=2163a,btoo few photos (n=8)a,c,d56a,en=1349a,fn=2164a,gn=2166a,gn=2147a,hn=2151 ± 3.847–56
BIPB752015-04-1946%transecttoo few photos (n=9)a,i,d46a,b—too few photos (n=9)a,e,dtoo few photos (n=9)a,f,dtoo few photos (n=9)a,g,dtoo few photos (n=9)a,g,dtoo few photos (n=9)a,h,d——
BIPB762015-04-1940%transect21a,in=2040a,b—30a,en=2032a,fn=2035a,gn=2040a,gn=2029a,hn=2028 ± 4.721–32
BINI772015-04-2042%transect26a,in=1142a,b—30a,en=1126a,fn=1138a,gn=1141a,gn=1134a,hn=1129 ± 3.726–34
BINI782015-04-2022%transect25a,in=1622a,b—24a,en=1629a,fn=1627a,gn=1629a,gn=1628a,hn=1627 ± 2.424–29
BINI792015-04-2034%transect18a,in=1534a,b—19a,en=1522a,fn=1527a,gn=1525a,gn=1521a,hn=1520 ± 1.718–22
BINI802015-04-2149%transect29a,in=2449a,b38a,cn=1027a,en=1437a,fn=2442a,gn=2446a,gn=2435a,hn=2432 ± 4.727–37
BIVB812015-04-2143%transecttoo few photos (n=5)a,i,d43a,b—too few photos (n=5)a,e,dtoo few photos (n=5)a,f,dtoo few photos (n=5)a,g,dtoo few photos (n=5)a,g,dtoo few photos (n=5)a,h,d——
BIBL822015-04-2242%transect28a,in=2242a,btoo few photos (n=4)a,c,d28a,en=1826a,fn=2242a,gn=2250a,gn=2236a,hn=2229 ± 4.326–36
BIBL832015-04-2233%transect24a,in=1533a,btoo few photos (n=9)a,c,dtoo few photos (n=6)a,e,d25a,fn=1542a,gn=1545a,gn=1541a,hn=1530 ± 9.524–41
BISA842015-04-2249%transecttoo few photos (n=7)a,i,d49a,btoo few photos (n=2)a,c,dtoo few photos (n=5)a,e,dtoo few photos (n=7)a,f,dtoo few photos (n=7)a,g,dtoo few photos (n=7)a,g,dtoo few photos (n=7)a,h,d——
BIBL862015-04-2343%transect34a,in=2743a,b—32a,en=2736a,fn=2751a,gn=2749a,gn=2725a,hn=2732 ± 4.825–36
BIBL872015-04-2343%transect30a,in=3543a,b—28a,en=3534a,fn=3542a,gn=3548a,gn=3529a,hn=3530 ± 2.628–34
BISA882015-04-2360%transect43a,in=1360a,b—46a,en=1354a,fn=1355a,gn=1360a,gn=1343a,hn=1346 ± 5.243–54
BISA912015-04-2457%transect40a,in=1357a,b—47a,en=1356a,fn=1358a,gn=1365a,gn=1344a,hn=1347 ± 6.540–56
BISP892015-04-2434%transect20a,in=1334a,b—17a,en=1322a,fn=1329a,gn=1330a,gn=1328a,hn=1322 ± 4.417–28
BISP902015-04-2444%transecttoo few photos (n=8)a,i,d44a,b—too few photos (n=8)a,e,dtoo few photos (n=8)a,f,dtoo few photos (n=8)a,g,dtoo few photos (n=8)a,g,dtoo few photos (n=8)a,h,d——
BISA922015-04-2526%transect19a,in=3926a,b—29a,en=3926a,fn=3933a,gn=3934a,gn=3929a,hn=3926 ± 4.919–29
BISA932015-04-2541%transect33a,in=2741a,b—39a,en=2726a,fn=2858a,gn=2661a,gn=2726a,hn=2831 ± 5.926–39
BISA942015-04-2552%transecttoo few photos (n=2)a,i,d52a,b—too few photos (n=2)a,e,dtoo few photos (n=2)a,f,dtoo few photos (n=2)a,g,dtoo few photos (n=2)a,g,dtoo few photos (n=2)a,h,d——
BISA972015-04-2655%transect51a,in=1255a,b—51a,en=1250a,fn=1268a,gn=1269a,gn=1229a,hn=1245 ± 10.629–51
BISP952015-04-2661%transect18a,in=2661a,b—22a,en=2620a,fn=2628a,gn=2628a,gn=2623a,hn=2521 ± 2.318–23
BISP962015-04-2682%transect51a,in=1882a,b—56a,en=1849a,fn=1869a,gn=1873a,gn=1848a,hn=1851 ± 3.348–56
BIVB992015-04-2760%transect35a,in=1160a,b—21a,en=1125a,fn=1164a,gn=1163a,gn=1119a,hn=1125 ± 7.519–35
BISA1022015-04-2832%transect22a,in=2732a,b—31a,en=2730a,fn=2733a,gn=2735a,gn=2734a,hn=2729 ± 5.522–34
BISP1002015-04-2841%transecttoo few photos (n=7)a,i,d41a,b—too few photos (n=7)a,e,dtoo few photos (n=7)a,f,dtoo few photos (n=7)a,g,dtoo few photos (n=7)a,g,dtoo few photos (n=7)a,h,d——
BISP1012015-04-2852%transect23a,in=1352a,b—39a,en=1332a,fn=1342a,gn=1344a,gn=1333a,hn=1332 ± 6.623–39
BISA1032015-04-2933%transect30a,in=1933a,b—28a,en=1937a,fn=1941a,gn=1944a,gn=1936a,hn=1933 ± 4.428–37
BISA1042015-04-2936%transect36a,in=1636a,b—44a,en=1646a,fn=1646a,gn=1651a,gn=1641a,hn=1642 ± 4.436–46
BISA1052015-04-2952%transect28a,in=1952a,b—24a,en=1936a,fn=1942a,gn=1947a,gn=1923a,hn=1928 ± 5.823–36
BIPB1102015-04-3042%transect29a,in=4542a,b—37a,en=4537a,fn=4544a,gn=4350a,gn=4337a,hn=4435 ± 3.629–37
BISA1072015-04-3025%transect18a,in=1225a,b—18a,en=1221a,fn=1232a,gn=1241a,gn=1222a,hn=1220 ± 1.818–22
BIVB1082015-04-3047%transect12a,in=1147a,btoo few photos (n=1)a,c,d25a,en=1024a,fn=1123a,gn=1018a,gn=1128a,hn=1122 ± 7.212–28
BIVB1092015-04-3048%transect35a,in=3348a,btoo few photos (n=1)a,c,d35a,en=3239a,fn=3367a,gn=3366a,gn=3325a,hn=3333 ± 6.025–39
BIPB1122015-05-0246%transect29a,in=1446a,b44a,cn=14—44a,fn=1451a,gn=1455a,gn=1438a,hn=1437 ± 7.829–44
BIPB1142015-05-02–0343%transect20a,in=2943a,b29a,cn=22too few photos (n=7)a,e,d32a,fn=2937a,gn=2938a,gn=2935a,hn=2929 ± 8.020–35
All sites (unweighted mean)——46————————
Notes for Chagos #2 2015 (13)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [b] Gold = LOF in-water diver transect, the mean of all of the station's transects: a different instrument, site-level only, never within 200 m of the frames.
  • [c] The author's own CoralNet points on the ground-truth frames at this site (30 − WATER − FISH, unknown points kept).
  • [d] Too few photos (under 10) at this site for a value: not shown and not averaged (photo-count rule: 10 or more photos for a value, 20 or more for a confident one).
  • [e] CoralNet robot labels (the cover set), same denominator.
  • [f] DINOv3 point predictions at the same CoralNet points (2,500 frames).
  • [g] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not considered reliable for these photos: 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [i] Claude run complete (2,500 of 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), unknown points 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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 · 1,281 photos

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

SiteDateHeadline (site page)ClaudeGold standardManually annotated in CoralNetCoralNetDINOv3Coralscapes v1Coralscapes v2CoralSCOPMean ± SD of CVRange
BICA012015-03-094.6%transecttoo few photos (n=1)a,g,h4.6a,btoo few photos (n=1)a,i,h—too few photos (n=1)a,d,htoo few photos (n=1)a,e,htoo few photos (n=1)a,e,htoo few photos (n=1)a,f,h——
BIEG022015-03-1037%transect18a,gn=2137a,b25a,in=21—24a,dn=2131a,en=2135a,en=2124a,fn=2122 ± 3.218–24
BIEG032015-03-1040%transect17a,gn=3640a,b24a,in=36—18a,dn=3632a,en=3634a,en=3626a,fn=3620 ± 4.817–26
BIEG042015-03-1033%transect26a,gn=1933a,b34a,in=19—35a,dn=1936a,en=1939a,en=1928a,fn=1930 ± 4.826–35
BIEG052015-03-1144%transect39a,gn=5944a,b41a,in=3558a,cn=2440a,dn=5961a,en=5964a,en=5941a,fn=5945 ± 8.839–58
BIEG062015-03-1149%transect32a,gn=3049a,b—44a,cn=3034a,dn=3046a,en=3046a,en=3023a,fn=3033 ± 8.923–44
BIEG072015-03-1148%transecttoo few photos (n=8)a,g,h48a,btoo few photos (n=8)a,i,h—too few photos (n=8)a,d,htoo few photos (n=8)a,e,htoo few photos (n=8)a,e,htoo few photos (n=8)a,f,h——
BIDI082015-03-1261%transect35a,gn=3561a,b—36a,cn=3529a,dn=3550a,en=3553a,en=3539a,fn=3535 ± 4.229–39
BIDI092015-03-1238%transect34a,gn=3038a,b—38a,cn=3024a,dn=3056a,en=3061a,en=3032a,fn=3032 ± 6.024–38
BIDI102015-03-1223%transect23a,gn=1323a,b—27a,cn=1316a,dn=1344a,en=1349a,en=1333a,fn=1325 ± 7.116–33
BIDI122015-03-134.2%transect2.9a,gn=144.2a,b—10a,cn=1411a,dn=148.2a,en=1410a,en=1418a,fn=1411 ± 6.12.9–18
BIDI132015-03-137.2%transect18a,gn=117.2a,b—20a,cn=1118a,dn=1122a,en=1124a,en=1125a,fn=1120 ± 3.318–25
BIDI142015-03-1317%transecttoo few photos (n=9)a,g,h17a,b—too few photos (n=9)a,c,htoo few photos (n=9)a,d,htoo few photos (n=9)a,e,htoo few photos (n=9)a,e,htoo few photos (n=9)a,f,h——
BIDI152015-03-1420%transecttoo few photos (n=9)a,g,h20a,b—too few photos (n=9)a,c,htoo few photos (n=9)a,d,htoo few photos (n=9)a,e,htoo few photos (n=9)a,e,htoo few photos (n=9)a,f,h——
BIDI162015-03-1427%transect24a,gn=1327a,b—26a,cn=1322a,dn=1332a,en=1336a,en=1331a,fn=1326 ± 3.722–31
BIDI16.52015-03-1446%*manual26a,gn=1146a,jn=13—39a,cn=1142a,dn=1137a,en=1139a,en=1130a,fn=1134 ± 7.326–42
BIEA182015-03-1521%transect18a,gn=3521a,b—22a,cn=3515a,dn=3533a,en=3532a,en=3428a,fn=3521 ± 5.915–28
BIEA192015-03-1528%transect31a,gn=1428a,b—27a,cn=1425a,dn=1442a,en=1445a,en=1439a,fn=1431 ± 6.125–39
BIEA202015-03-1513%transect22a,gn=1913a,b—25a,cn=1916a,dn=1931a,en=1935a,en=1927a,fn=1922 ± 4.816–27
BIEA222015-03-1627%transect31a,gn=2327a,b—49a,cn=2345a,dn=2344a,en=2346a,en=2350a,fn=2344 ± 8.931–50
BIEA232015-03-1626%transect29a,gn=2426a,b—36a,cn=2429a,dn=2444a,en=2449a,en=2440a,fn=2433 ± 5.429–40
BIEA242015-03-1616%transect37a,gn=1116a,b—41a,cn=1141a,dn=1149a,en=1149a,en=1139a,fn=1140 ± 1.737–41
BIEA24.52015-03-16—too few photos (n=5)a,g,h—k—too few photos (n=5)a,c,htoo few photos (n=5)a,d,htoo few photos (n=5)a,e,htoo few photos (n=5)a,e,htoo few photos (n=5)a,f,h——
BIEA252015-03-1718%transect16a,gn=1918a,b—29a,cn=1926a,dn=1931a,en=1933a,en=1930a,fn=1925 ± 6.216–30
BIEA262015-03-1724%transect15a,gn=3024a,b—21a,cn=3014a,dn=3023a,en=3028a,en=3027a,fn=3019 ± 6.214–27
BITB272015-03-1736%transect24a,gn=1636a,b—34a,cn=1624a,dn=1646a,en=1648a,en=1638a,fn=1630 ± 7.024–38
BITB282015-03-1831%transect18a,gn=4331a,b—27a,cn=4320a,dn=4334a,en=4337a,en=4329a,fn=4324 ± 5.118–29
BITB292015-03-1830%transect19a,gn=4930a,b—30a,cn=4924a,dn=4940a,en=4942a,en=4926a,fn=4924 ± 4.619–30
BITB302015-03-1843%transect14a,gn=1243a,b—30a,cn=1218a,dn=1234a,en=1238a,en=1230a,fn=1223 ± 8.214–30
BITB322015-03-1930%transect12a,gn=1930a,b—17a,cn=1912a,dn=1918a,en=1920a,en=1920a,fn=1915 ± 4.012–20
BITB332015-03-1921%transect7.9a,gn=2721a,b—20a,cn=2711a,dn=2718a,en=2618a,en=2622a,fn=2715 ± 6.97.9–22
BITB342015-03-1962%transect37a,gn=1162a,b—41a,cn=1135a,dn=1156a,en=1167a,en=1047a,fn=1140 ± 5.335–47
BITB34.52015-03-19—too few photos (n=3)a,g,h—k—too few photos (n=3)a,c,htoo few photos (n=3)a,d,htoo few photos (n=3)a,e,htoo few photos (n=3)a,e,htoo few photos (n=3)a,f,h——
BICC352015-03-2014%transect6.7a,gn=1514a,b—17a,cn=1513a,dn=1521a,en=1524a,en=1520a,fn=1514 ± 5.96.7–20
BICC362015-03-208.6%transect6.4a,gn=158.6a,b—19a,cn=1518a,dn=1521a,en=1527a,en=1525a,fn=1517 ± 7.76.4–25
BITB372015-03-2040%transect29a,gn=5640a,b—37a,cn=5630a,dn=5751a,en=5554a,en=5542a,fn=5735 ± 6.429–42
BITB382015-03-2158%transect26a,gn=4458a,b—34a,cn=4428a,dn=4454a,en=4456a,en=4431a,fn=4430 ± 3.326–34
BITB392015-03-2152%transect37a,gn=6052a,b—40a,cn=6032a,dn=6153a,en=5957a,en=5944a,fn=6138 ± 5.432–44
BITB402015-03-2145%transecttoo few photos (n=5)a,g,h45a,b—too few photos (n=5)a,c,htoo few photos (n=5)a,d,htoo few photos (n=5)a,e,htoo few photos (n=5)a,e,htoo few photos (n=5)a,f,h——
BITB412015-03-2243%transect25a,gn=1343a,b—32a,cn=1332a,dn=1340a,en=1342a,en=1340a,fn=1332 ± 6.025–40
BITB422015-03-2240%transect23a,gn=4940a,b—31a,cn=4924a,dn=5140a,en=5043a,en=5135a,fn=5128 ± 5.923–35
BITB432015-03-2230%transect22a,gn=1230a,b—36a,cn=1215a,dn=1247a,en=1250a,en=1237a,fn=1227 ± 10.715–37
BITB442015-03-2339%transect27a,gn=3639a,b—33a,cn=3629a,dn=3644a,en=3647a,en=3639a,fn=3632 ± 5.127–39
BITB452015-03-2345%transect33a,gn=1345a,b—45a,cn=1339a,dn=1351a,en=1352a,en=1336a,fn=1338 ± 4.833–45
BITB462015-03-2332%transect18a,gn=1132a,b—28a,cn=1115a,dn=1135a,en=1038a,en=1031a,fn=1123 ± 7.615–31
BITB472015-03-2453%transecttoo few photos (n=2)a,g,h53a,b—too few photos (n=2)a,c,htoo few photos (n=2)a,d,htoo few photos (n=2)a,e,htoo few photos (n=2)a,e,htoo few photos (n=2)a,f,h——
BICC502015-03-2515%transect16a,gn=1115a,b—29a,cn=1130a,dn=1137a,en=1148a,en=1129a,fn=1126 ± 6.916–30
BICC512015-03-2517%transect21a,gn=1717a,b—22a,cn=1724a,dn=1727a,en=1728a,en=1728a,fn=1724 ± 3.021–28
BIEA522015-03-2537%transect35a,gn=2237a,b—38a,cn=2222a,dn=2250a,en=2250a,en=2239a,fn=2233 ± 8.022–39
BITB532015-03-2633%transect23a,gn=2233a,b—35a,cn=2231a,dn=2241a,en=2242a,en=2227a,fn=2229 ± 5.023–35
BITB542015-03-2656%transect30a,gn=9956a,b—34a,cn=9921a,dn=9950a,en=9754a,en=9731a,fn=9929 ± 5.821–34
BITB552015-03-2651%transect33a,gn=1451a,b—43a,cn=1434a,dn=1447a,en=1451a,en=1438a,fn=1437 ± 4.733–43
BITB562015-03-2725%transect19a,gn=1325a,b—20a,cn=1320a,dn=1325a,en=1326a,en=1318a,fn=1319 ± 1.018–20
BITB572015-03-27–2843%transect43a,gn=11143a,b—44a,cn=11133a,dn=11166a,en=11168a,en=11136a,fn=11139 ± 5.333–44
BITB582015-03-2867%transecttoo few photos (n=1)a,g,h67a,b—too few photos (n=1)a,c,htoo few photos (n=1)a,d,htoo few photos (n=1)a,e,htoo few photos (n=1)a,e,htoo few photos (n=1)a,f,h——
Notes for Chagos #1 2015 (16)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [b] Gold = LOF in-water diver transect, the mean of all of the station's transects: 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 (2,500 frames).
  • [e] Coralscapes (v1 EPFL-ECEO/coralscapes-vit-b-dpt; v2 coralscapesv2-dinov3-vitb-lora-dpt) is not considered reliable for these photos: 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [g] Claude run complete (2,500 of 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), unknown points kept.
  • [h] Too few photos (under 10) at this site for a value: not shown and not averaged (photo-count rule: 10 or more photos for a value, 20 or more for a confident one).
  • [i] The author's own CoralNet points on the ground-truth frames at this site (30 − WATER − FISH, unknown points kept).
  • [j] No LOF transect at this site: the gold cell is the author's own CoralNet points (frames manually annotated in CoralNet) instead.
  • [k] No LOF transect at this site.
  • 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 in gray italic but never averaged.
  • * The site's headline coral cover rests on only 10–19 reef-habitat photos; this estimate is less certain.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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

Tonga 2013

22 sites · 422 photos

Gold standard = LOF diver transects (no photos); the value is the mean of all of each site's transects. Claude is a comparison only, at sites with at least 10 usable habitat photos.

SiteDateHeadline (site page)ClaudeGold standardCoralSCOP
TOHA032013-09-1220%transect20a,e,fn=1620a,b32a,c,dn=16
TOHA042013-09-1225%transect18a,e,fn=1425a,b26a,c,dn=14
TOHA082013-09-13–1422%transect34a,en=2622a,b28a,c,dn=26
TOHA102013-09-1440%transect29a,en=2340a,b25a,c,dn=23
TOHA142013-09-1644%transect36a,en=3844a,b33a,c,dn=38
TOHA182013-09-1753%transect48a,e,fn=1653a,b29a,c,dn=16
TOHA212013-09-1835%transect32a,e,fn=1235a,b38a,c,dn=12
TOHA222013-09-1840%transect12a,en=2140a,b21a,c,dn=21
TOHA272013-09-2031%transect25a,e,fn=1331a,b39a,c,dn=13
TOHA282013-09-2031%transect46a,e,fn=1731a,b32a,c,dn=17
TOHA292013-09-2042%transect45a,e,fn=1342a,b41a,c,dn=13
TOHA302013-09-2140%transect36a,e,fn=1040a,b36a,c,dn=10
TOHA312013-09-2134%transect37a,e,fn=1834a,b39a,c,dn=18
TOHA322013-09-2134%transect44a,e,fn=1134a,b27a,c,dn=11
TOVA332013-09-2233%transect33a,en=2133a,b36a,c,dn=21
TOVA372013-09-2328%transect12a,e,fn=1828a,b25a,c,dn=18
TOVA382013-09-2441%transect30a,en=4141a,b31a,c,dn=41
TOVA442013-09-2553%transect32a,e,fn=1553a,b40a,c,dn=15
TOVA522013-09-2847%transect43a,e,fn=1747a,b43a,c,dn=17
TOVA532013-09-2838%transect40a,en=3438a,b39a,c,dn=34
TONI542013-09-2934%transect20a,e,fn=1334a,b34a,c,dn=13
TONI552013-09-2933%transect35a,e,fn=1533a,b34a,c,dn=15
All sites (unweighted mean)—323633
Notes for Tonga 2013 (10)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [b] Gold = Living Oceans Foundation (LOF) Global Reef Expedition in-water diver transect, the mean of all of the site's transects. It is the primary value; the photo cell is a comparison only.
  • [c] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [d] CoralSCOP here = the site page's frame-matched row: run on exactly the photos behind the Claude cell (per-frame values joined by file; capped at 100% per photo; zero-filled). Comparison only.
  • [e] Claude = blind 7-call point count on the author's EDITED September 2013 survey-dive photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator; mean across photos. Comparison only: never averaged with the transect and never the headline; site means only (photos are not the transect points).
  • [f] LOW-N: 10–19 usable habitat photos at this site; indicative only, not a confident estimate.
  • 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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

Fiji 2013

10 sites · 148 photos

Gold standard = LOF diver transects (no photos); the value is the mean of all of each site's transects. Claude is a comparison only, at the 10 sites with at least 10 usable habitat photos (9 low-n). FJTV42 and FJTO09 come from the expedition's global site table.

SiteDateHeadline (site page)ClaudeGold standardCoralSCOP
FJTO022013-06-0336%transect26a,f,en=1236a,b34a,c,d,en=12
FJTO092013-06-0529%transect18a,f,en=1329a,b25a,c,d,en=13
FJML162013-06-0833%transect35a,f,en=1933a,b41a,c,d,en=19
FJML182013-06-0940%transect24a,f,en=1040a,b38a,c,d,en=10
FJML202013-06-0937%transect51a,f,en=1337a,b54a,c,d,en=13
FJNA402013-06-1644%transect28a,f,en=1544a,b31a,c,d,en=15
FJTV422013-06-1737%transect45a,fn=2337a,b38a,c,dn=23
FJCC452013-06-1850%transect41a,f,en=1850a,b41a,c,d,en=18
FJCC482013-06-1947%transect39a,f,en=1447a,b38a,c,d,en=14
FJVB522013-06-2147%transect39a,f,en=1147a,b34a,c,d,en=11
All sites (unweighted mean)—344037
Notes for Fiji 2013 (10)
  • [a] Value from the map's computer-vision method tables (denominator and photo set as on the site's page).
  • [b] Gold = Living Oceans Foundation (LOF) Global Reef Expedition in-water diver transect, the mean of all of the site's transects; FJTV42 and FJTO09 come from the expedition's global site table. It is the primary value; the photo cell is a comparison only.
  • [c] 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. It is also unstable frame to frame (the same subject has scored 94.1% in one photo and 23.3% in another), and the mask sum overstates cover most on the healthiest reefs.
  • [d] CoralSCOP here = the site page's frame-matched row: run on exactly the photos behind the Claude cell (per-frame values joined by file; capped at 100% per photo; zero-filled). Comparison only.
  • [e] LOW-N: 10–19 usable habitat photos at this site; indicative only, not a confident estimate.
  • [f] Claude = blind 7-call point count on the author's EDITED June 2013 survey-dive photos, 20 points per photo, (HC+BL)/(20 − WATER − FISH − OBST), unknown points kept in the denominator; mean across photos. Comparison only: never averaged with the transect and never the headline; site means only (photos are not the transect points).
  • 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 in gray italic but never averaged.
  • Headline (site page): the coral-cover figure the site's page, report card and map pop-up print, with its method; it is never averaged into the CV mean.
  • 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