Even coral experts disagree on about one label in twelve
A reef dataset built from 76 scattered sources shows what it takes to make labelled images fit for training.


Picture made for ABUREAU.
Ask two coral scientists what a patch of reef is, and they will not always agree. In ReefNet, a dataset paper whose latest version is dated 6 November 2026, experts audited 8,962 patches and agreed on 92% of them. The authors say this is not a full measure of accuracy.
The set starts from about 925,000 hard-coral point labels pooled from 76 public CoralNet sources, plus a separate collection from Al-Wajh. After the audit, the authors filtered it to a benchmark of 479,027 labels across 39 classes, and mapped them to a shared marine taxonomy.
In the paper's tests on the held-out Al-Wajh data, models did best when adapted with supervised training, and the best reached 65.79% macro recall. The authors describe that test as a regional stress test, not proof the models work worldwide, and list label uncertainty and uneven geography as open problems.
For people who own image or survey data, the lesson is plain: pooled labels from many hands need an expert check and a common naming scheme before they are worth much to a model builder. Where the data came from, and how far it was checked, are part of its value.