Introduction
Innovation has played a major role in shaping European aquaculture , and it remains essential for the sector’s continued survival. Advances in digital technology offer new opportunities for innovation that might contribute to the industry’s continued success , stimulating initiatives for the development of artificial intelligence, automation, and Internet of Things based solutions. The iFishIENCi project aims to optimise feeding in fish production systems through combin ed application of machine learning algorithms , video observation of fish behaviour, and water quality digital sensors and software. The approach uses FishMET , a model simulation of fish feeding behaviour and related physiological and metabolic processes. The ability of FishMET to accurately simulate the se processe s is limited by a lack of data describing their functional relationships for different species within varying conditions. Obtaining the required data is challenging as these relationships have a dynamic and apparently complex nature . To address these issues, an ongoing series of experiments are being performed to a) validate a combination of methods for obtaining data that describe rainbow trout (Oncorhynchus mykiss) feeding behaviour, feed intake, and stomach fullness, to b) identify and describe relationships between these processes and c) generate data that is useful for the development and training of machine learning algorithms. If distinct levels of stomach fullness and hunger are associated with different behaviours , it may be possible to use a combination of machine learning and real-time video monitoring of fish behaviour to assist with the optimisation of feeding regimes beyond what is currently possible.
Materials and methods
A recirculation aquaculture system facility was stocked with rainbow trout of varying size and held at densities of approximately 35kg/m3. The fish were subject to varied meal regimes delivered by automated feeders, and feed consumption, stomach fullness, behaviour and water quality data were assessed. Behaviour was recorded by video camera, and feed consumption was quantified using a labelled diet and X-ra y imaging, as described by McCarthy et al. (1992). Digital sensors and cloud integrated monitoring were used to record t emperature and dissolved oxygen . Machine learning algorithms based on the Sciit- learn python library were trained and used for target acquisition and fish tracking, and k-means clustering techniques were used to evaluate inter-fish distances and grouping.
Results
Fish in video images were successfully targeted and tracked using algorithms. Aggregations of fish below feeders were observed before feeding events and were effectively characterised by k-means clustering algorithms. Strong s chooling is observed post feeding but increased clustering pre-feeding was not consistently observed. In feed markers are visible in X-ray images and calculated feed consumption quantities compare well to known quantities of feed consumption.
Discussion
The ability of algorithms to identify, track, and characterise clustering is promis ing, although more exploration is required. k-means clustering scores were not consistently higher before mealtimes despite the absence of aggregations post-feeding. The incidence of aggressive behaviours, which may disrupt clustering, appear to increase before meals, and singles behaviour types may not be reliable as standalone cues. Aggregation may be an anticipatory response in fish conditioned to receiving feed at set times and from the same location , and so its usefulness as an indicator of hunger is yet to be determined. Behaviour such as aggression or higher swimming velocity may be more appropriate indicators of hunger state, although behavioural expression may differ across cultivation conditions and size classes. The McCarthy et al. (1992) method was successfully used to quantity feed consumption . Its suitability for quantifying relative fullness across separate regions of the gastrointestinal track is now being investigated . This information can be useful for understanding hunger, satiety, anticipation, and associated behaviours. Ongoing experimentation continues to assist the development of machine learning algorithms that may reveal relationships between stomach fullness, hunger state and behaviour.
References
McCarthy, I. D., Houlihan, D. F., Carter, C. G., and Moutou, K., 1993. Variation in individual food consumption rates of fish and its implications for the study of fish nutrition and physiology. Proceedings of the Nutrition Society, 52, 427–436.