A machine learning based image classification method to estimate fish sizes from images without a specified reference object

Abstract

Most fish populations around the world are unassessed and their status is unknown. Size based methods could provide fast and transparent assessments, but they require information on fish sizes. Citizen science programs, social media and smart phone applications generate millions of georeferenced fish images globally. Machine learning based fish species and size identification could help turn these images into valuable data for population status assessments. We present a machine learning, image classification based method to identify fish size classes from photos of anglers holding fish. To train the model we group images into ten 5-10 cm size classes, similar to classes used in underwater visual fish surveys. The model was trained using 2602 images from angler citizen science platforms MyCatch and FishSizeProject. Although the number of images was limited, the model achieved an overall accuracy of ~50%. Importantly, the misidentification of size classes was consistent across 20 separate model training rounds, each conducted with an independent, random allocation of images for training and test datasets. Our method suggests that photo based fish size class identification is feasible, and that prediction uncertainty should be incorporated into subsequent analysis, as it is done with fish ageing errors in fisheries stock assessments.