Trawling for photographic bycatch: using computer vision to detect habitat-forming kelp in citizen science fish photographs

Photographs of organisms taken by citizen scientists often contain a wealth of secondary data such as habitat composition information captured in the background. Computer vision has the potential to unlock these secondary data and increase the value of this resource for biodiversity monitoring. Here, we trained a computer vision model using approximately 4000 underwater fish photographs to detect the presence of the common kelp species Ecklonia radiata (hereafter Ecklonia ) in the background. The best performing model returned an average area under the precision-recall curve (AUPRC) value of 0.798 on the validation set. The model had an average precision of 0.79 across unseen test sets, meaning that approximately 4 out of every 5 Ecklonia detections were correct. However, the model had a relatively low recall of 0.52 across test sets, meaning it was only detecting half of the fish images that contained Ecklonia. When the best performing model was deployed on an unseen data set of 55000 iNaturalist photographs of fish, approximately 3900 secondary Ecklonia observations were obtained (after accounting for the false positive rate). This represents a 15-fold increase in the number of underwater photographs of Ecklonia in our study region when compared to iNaturalist observations where Ecklonia was the primary subject of the photograph. This study demonstrates how extracting secondary data using computer vision can considerably increase the information obtained from citizen science photographs.