How-to 02 · Scoring

Point intercept in CoralNet

CoralNet, run by the University of California San Diego, is the platform most benthic monitoring uses: images go up, points go on, a classifier proposes a label at each point and a person confirms or corrects it. The course's class list is the one in general use there, which is why your Entry can sit beside other people's data.

Two rules before you start. The machine proposes and you decide: the human-confirmed label is the score of record, the raw model output is logged and never reported, and the model version is written down. And your buddy scores the same frame at the same points without conferring — which in CoralNet means two people, two passes, and no peeking at each other's labels before both are done.

Once: the account and the source

Create an account at coralnet.ucsd.edu. It is free. Each scorer needs their own account — you and your buddy are two users.
Create a source. A source is CoralNet's container for a project: one set of images, one label set, one set of point-generation rules. Name it for the site and the station, not for yourself — the Dossier belongs to the station, and the source should outlive your involvement. Set its visibility as you and your instructor decide; private is fine.
Set the labelset. Add the course's classes from CoralNet's shared label list: hard coral, soft coral, dead coral, macroalgae, crustose coralline algae, turf algae, other invertebrates, sand, rubble, rock — and an indeterminate (unknown) label. Use the existing CoralNet labels where they exist rather than inventing your own; that is the whole reason for using somebody else's list. Write the labelset you ended up with, in full, in Section 4 of the Entry.
Set point generation in the source settings: simple random, 50 points per image (25 is the course minimum; 50 the recommendation). This is done once per source and applies to every image uploaded after it, so set it before the first upload.
Add your buddy to the source with permission to annotate.

Each capture set: upload and score

Upload the frame you are scoring — the one the frame check chose, and only that one. Keep the camera's filename so the archive manifest can match it. Upload from the camera's JPEG; if the record is raw, export a JPEG for scoring and note in Section 12 that the scoring copy is a derivative of the raw file that remains the record.
Open the annotation tool for the image. CoralNet places the 50 random points. Any point that lands on a reference chart is re-drawn — the charts occupy area that cannot be scored — so note how many you moved in Section 4 (points re-drawn).
Score every point. If the source has a trained classifier it will propose a label at each point; confirm or change it. If there is no classifier yet — a new source starts without one — you label each point yourself. Either way the question at each point is the same: what is directly underneath this one? A point on bleached tissue is hard coral. A point you cannot call — white tissue or white skeleton, coral or coralline, a saturated highlight — is indeterminate, never the likelier guess.
Mark the image confirmed when all points are labeled. That is your score of record.
Your buddy scores the same image at the same points, logged in as themselves, without looking at your labels first. CoralNet keeps one label per point per image, so the practical way to get two independent scores is one of: your buddy records their labels on the printed point list before either of you enters anything; or you duplicate the image in the source under a second name and each of you scores one copy; or you export your annotations before your buddy starts and they re-label the same points. Any of these works — what matters is that the two passes are independent and that both figures go in Section 4 side by side, never averaged.

Getting the numbers

Count the points in each class. CoralNet shows the per-image label counts and the source's statistics pages show them as percentages; either way it is points out of the point count. 16 of 50 on hard coral is 32 %. Read the sampling Standard Error for that cover and point count from the table in the Forms document, and write both in Section 4. A class seen at fewer than three points is reported as a count — "2 of 50" — not as a percentage with a ±.

Export: the archive needs the points

Export annotations from the source (the export tools are on the source's pages; choose the annotations CSV). The file lists, for every point, its row and column in pixels and its label — that is the point-position record the Entry requires, so that anyone can score the same spots again.
If a classifier proposed labels, also export the machine suggestions where CoralNet offers them, and record the classifier's version and the scorer-model agreement in Section 4. The raw model output is logged, never reported as the result.
Name and file the exports with the frame in the archive, and list them in Section 12 as the point list for that frame. Record "CoralNet" and the date in Section 12 as the scoring tool; the site itself changes continuously and has no version number to cite, so the date stands in for one.

Why the export matters more than the score. A cover figure without its point positions is a claim nobody can check. Ten years from now the images may be rescored against a class nobody thought to ask about; the point positions are what let that rescoring be compared with yours.

Last checked against coralnet.ucsd.edu in September 2026. CoralNet's interface changes on its own schedule; the workflow here — labelset, point generation, upload, confirm, independent second scorer, export the points — is the course's, and does not. If a step no longer matches the site, say so, and use the manual overlay or the Fiji fallback in the meantime — both are entirely legitimate.