Machine Vision Applications for Welfare Monitoring in Aquaculture: Challenges and Opportunities

Abstract

Increasing consideration of welfare in aquaculture has prompted interest in non-invasive methods of monitoring that avoid unnecessary stress and handling. Machine vision (MV) provides a potential solution to these needs, as it can be used for non-invasive monitoring of animal health and welfare in real-time. We examined the practical applications of MV for welfare monitoring in aquaculture, the hardware and algorithms used for automated data collection, and the main challenges and solutions for data processing and analysis. The most common application of MV has been the estimation of size-related metrics (growth, biomass) in fish, but key aspects of welfare, such as monitoring of parasites and disease or detection of stress-related behaviours, are lagging behind. Numerous camera setups have been used, ranging from single to stereoscopic cameras and from emersed to submerged cameras, but these have often been used under optimal conditions that may not always reflect those prevalent in industry (high densities, low visibility), likely overestimating performance. Object detection algorithms, such as YOLO, have been the approach of choice for most MV applications in aquaculture, but our review has identified an increasing number of alternatives that can help circumvent some of the challenges posed by high densities and poor lighting typical of commercial farms. MV has the potential to transform welfare monitoring in aquaculture, but there are still important challenges that need to be overcome before it can become mainstream, namely the ability to detect ectoparasites and diseases, identify abnormal behaviours, and work across taxa, particularly in crustaceans.