Fish-bait-efficiency and benthic stock assessments using deep learning

The rise of new technology is continuously generating gigantic datasets, known as “big data”. Imagery to tackle biological and ecological questions, is no exception. Observing and learning from these data is crucial but remains a tedious and labour-intensive process. This project aims to address to what extent computer vision based on deep learning can solve ecological questions while minimizing – if not removing human validation. For this purpose, a convolutional neural network (CNN) was trained on two types of data representing different sampling conditions and species communities. The first aimed at detecting the attraction levels of different types of biodegradable baits using baited remote underwater videos (BRUV). The BRUV footage analysis showed promising results with an average precision (AP), the standard metrics to assess the performance of deep learning models, of 0.827 for fish for the best performing model. An Interest index was introduced to assess each of the different bait types and a cockle bait functioned as the control. The resulting analysis – manual and automated - showed that the biodegradable plastic bait C17 has the greatest potential of replacing an old-fashioned cockle’s bait. The UWTV footage had more diverse classes (17 species, genus, or other taxa) and showed more mitigated results. The fish Callionymus spp., the crustacean Munida spp. and the Pennatulaceidae classes were accurately detected with AP values of 0.86, 0.82 and 0.80 respectively. In comparison, the main focus class Nephrops norvegicus slightly underperformed, with an AP value of 0.69. Other classes were more difficult to identify as such as “hydrozoa” and “crustacean” (AP of 0.23 and 0.24), due to their high diversity of shapes, colours and sizes. Nevertheless, in regard to other studies and given the challenging nature of marine-derived data, these values are satisfying. This project highlights the promising potential of replacing the labour-intensive human-validated analysis, while identifying the gaps that still need to be overcome. The generated models will help moving toward non-invasive methods with direct applications in marine conservation and fisheries management.