WIO-ReefFish: A High-Resolution Dataset for Taxon-Aware Coral Reef Fish Detection in the Western Indian Ocean

To understand the health of a coral reef, scientists often look at the fish swimming above it. The variety and number of different species present act as a living report card for the ecosystem, revealing whether the reef is thriving or struggling. For decades, researchers have relied on divers to swim along the reef, count the fish they see, and write down what they found. This manual work is slow, expensive, and difficult to scale up to cover the vast stretches of ocean that need monitoring. In recent years, computers have begun to learn how to spot these fish automatically in underwater videos, offering a way to speed up the process. However, for these computer programs to learn effectively, they need to be trained on large collections of real-world images where every fish has been carefully identified by a human. Such collections have been rare, especially for the Western Indian Ocean, leaving a gap in our ability to monitor these critical habitats with modern technology.

A new study addresses this gap by introducing a massive, high-quality collection of underwater images specifically designed to teach computers how to find and identify reef fish. The researchers compiled one thousand ultra-clear photographs, each capturing a wide view of the reef as a diver would see it during a standard survey. These images are not just snapshots; they are accompanied by thousands of precise outlines drawn around individual fish, marking exactly where each one is and what kind it is. The dataset covers twenty-four different groups of fish, preserving the complex way these animals live together in their natural environment. By making this resource public, the team has provided a realistic training ground for computer vision tools, moving beyond simple laboratory tests to the messy, crowded reality of a coral reef.

The team used these images to test how well nine different computer models could perform two specific tasks. The first task required the computer to find a fish and name its specific type, a process known as class-aware detection. The second task asked the computer only to find where a fish was located, without needing to know its exact species, which they called class-agnostic localization. The results showed a clear difference in performance. When the computers were asked to simply find the fish, they performed very well, successfully locating them even when the images were new and unseen. However, when the task required the computer to also identify the specific type of fish, the performance dropped significantly. The best model improved its accuracy from a moderate level when identifying species to a much higher level when it only needed to find the fish. This indicates that while computers are becoming quite good at spotting that a fish is there, telling one species from another remains a much harder challenge in the complex visual noise of a reef.

The study also tested how well these models would work in places they had never seen before. When the researchers evaluated the systems on images from different survey locations and different countries, the ability to identify specific fish types faltered noticeably. The models struggled to generalize their knowledge to new environments when asked to name the species. In contrast, the ability to simply locate the fish remained strong and reliable across these different settings. This suggests that while automated tools are ready to help count how many fish are in an area, relying on them to sort through the specific types of fish in a new region requires further development. The work establishes a new standard for measuring progress in this field, offering a foundation for building more robust tools that can eventually support large-scale monitoring of coral reef biodiversity in the Western Indian Ocean.