Seals and cormorants are top predators that play important roles in the ecosystem, while also sparking emotions and debates. The impact that seals and cormorants have on fish stocks is a frequently discussed issue. However, in order to answer this, information is needed about which species and sizes of fish seals and cormorants consume. Researchers at the Swedish University of Agricultural Sciences, in collaboration with the Institute of Marine Research in Bergen (Norway), have developed a new method that, using machine learning, improves researchers' ability to perform precise diet analyses.
Traditionally, diet analyses have been based on visual identification of otoliths (ear stones) and other skeletal remains from the fish eaten by the predators. Otoliths are collected from samples taken from the predators' digestive systems, fecal samples from seals, or regurgitated pellets from cormorants. The problem is that otoliths from many fish species are very similar to each other, making it difficult to distinguish them visually. The challenge becomes even greater when the otoliths are eroded, meaning they have been partially broken down by the strong stomach acids in the predators’ stomachs.
“Analyzing diet samples is a time-consuming process that also requires a great deal of expertise. By combining image analysis of the otoliths with machine learning, an AI-based approach, we have managed to create an efficient tool that helps us refine our knowledge of what seals and cormorants actually eat,” says Monica Mion, a researcher at the Department of Aquatic Resources at SLU and the lead author of the study on the new method, which was recently published in the scientific journal ICES Journal of Marine Science.
To develop and test the new method, Monica Mion and her colleagues conducted a study analyzing otoliths from vendace (Coregonus albula) and whitefish (Coregonus lavaretus) – two common fish species in the diet of the ringed seal in the Gulf of Bothnia. The ringed seal population was heavily endangered at the end of the 20th century but has since recovered. This has led to conflicts and debates over how much fish the ringed seals actually eat and whether they compete with commercial fisheries for fish resources. Predation by the ringed seal has also been included in stock assessments for vendace in the Gulf of Bothnia.
“Vendace is important for commercial fishing in the Gulf of Bothnia, and accurate information about how much vendace the seals eat, from different age classes, is needed for reliable stock assessments and to manage the stocks,” says Monica Mion.
Vendace and whitefish have otoliths that are very difficult to distinguish, especially after passing through the seals' digestive system. The researchers conducted laboratory experiments simulating the seals’ digestion in order to train the AI model on how otoliths change after erosion. The model was able to analyze both intact and eroded otoliths with 90 percent accuracy and determine which fish species they belonged. The study showed that about three-quarters of the otoliths analyzed from the ringed seals came from vendace, while a quarter belonged to whitefish.
The challenge of identifying fish species in diet samples is not unique to vendace and whitefish. Many cod species are also difficult to distinguish visually, and analyses often result in a general classification as "cod fish."
“But there is a significant difference between saying that seals eat 100 percent cod fish, which is often interpreted as 100 percent cod, and actually finding that their diet consists of 50 percent cod, 25 percent pollock, and 25 percent haddock,” says Monica Mion.
Having accurate diet data reduces the risk of uncertain extrapolations - that is, attempting to predict conditions beyond observed data based on existing trends. For example, assuming that seals consume the same fish species year-round, even when data only covers their winter diet. Such extrapolations (or assumptions) lead to incorrect conclusions.
Accurate conclusions is crucial for fisheries management and for reducing conflicts between stakeholders, such as commercial fishermen and conservationists. Furthermore, high-quality diet data is an essential component of ecosystem models and a prerequisite for ecosystem-based management.
- Our research not only shows that AI and machine learning can support diet analyses, but also opens the door to future applications for more species and ecosystems. The method has potential to improve our understanding of marine ecosystems and the complex relationships between predators and prey, says Monica Mion.
*Machine learning means that the computer is trained to recognize and distinguish between different fish species’ otoliths by analyzing large amounts of image data.
The article Species assignment from seal diet samples using shape analyses in a machine learning framework has been published in the ICES journal of Marine Science.
Monica Mion, Environmental analyst specialist
Department of Aquatic Resources, Institute of Marine Research, SLU
monica.mion@slu.se, +46 10 478 40 87
Karl Lundström, Researcher
Department of Aquatic Resources, Institute of Marine Research, SLU
karl.lundstrom@slu.se, +46 10 478 41 38