Abstract
Coastal ecosystems of the Mediterranean Sea are among the richest in non‐indigenous species, mostly due to the establishment of species coming from the Red Sea through the Suez Canal. Two herbivorous rabbitfishes, Siganus rivulatus and Siganus luridus , are already invasive in the south‐eastern part of the Mediterranean Sea where they cause ecological damage by overgrazing algae. The early detection and the counting of these non‐indigenous species in the rest of the Mediterranean Sea is thus a major challenge for scientists and ecosystem managers. However, analysing images from divers or remote cameras is a demanding task. Here, a dataset of 31,285 images of Siganus spp. and of six common native fishes to the Mediterranean Sea was built from 40 underwater videos recorded at three reef habitats. A deep learning algorithm was then trained to identify Siganus spp. on images containing the eight Mediterranean species. Finally, the algorithm and a post‐processing filtering were tested with an independent dataset of 2024 images. The model had a recall of 0.92 for the Siganus genus (i.e., two Siganus species combined). After a confidence‐based post‐processing, the recall increased to 0.98 with only 4 out of 272 images of Siganus spp. being misclassified. Accuracy reached a score of 0.61 meaning that experts would have to discard false positives. Images of five native species not present in the training dataset yielded similar false positive rates than species present in the training dataset. Overall, the automatic processing of images by the model and then the checking of putative Siganus images by experts required up to five times less effort than a full processing by experts. The algorithm can help to efficiently detect these two invasive fishes in underwater images to evaluate progress towards conservation objectives and accelerate citizen‐based monitoring of coastal ecosystems.