Aquaculture has grown over the past decade now producing 88 million tons annually which is 49 \% of the global fish production. By addressing production challenges and introducing automated methods, production can increase even further. Among such challenges are infectious diseases which have a profound impact. To alleviate this negative impact, prevention and early detection of diseases are of utmost importance. The latter includes both the tracking of health, for example in the form of growth monitoring, as well as the recognition of early signs of disease such as behavioral or physiological symptoms. With the development of modern technology such as data science and AI the growing aquaculture industry is switching from manual monitoring and controlling to machine-driven solutions. AI has automated information extraction from images and accurate data interpretation have facilitated better decision making and higher profitability in fish farms.
The goal of this thesis work is to utilize machine learning and computer vision for the early detection and prevention of fish diseases in aquaculture. My research work comprises of three main modules. The first module focuses on fish biomass estimation, utilizing deep learning algorithms to segment fish, classify them into five species, and estimate their biomass. The second module aims at detecting disease symptoms, employing a deep learning algorithm to classify fish into healthy and unhealthy categories, and subsequently identifying symptoms and locations of bacterial infections if a fish is classified as unhealthy. We expanded the capabilities of this module for real-time detection of flavobacterium in trout through the analysis of underwater footage.
The third module focuses on analyzing fish behavior in real time, Unlike the previous scenario where symptoms were physically visible, such as changes in color, bleeding or visible injuries, behavioral symptoms are not as prominent. Fish are sensitive to environmental changes and they exhibit a series of responses to changes in environmental factors. For example, when fish are stressed, they undergo various metabolic changes, all of which are expressed externally by variations in their behavior. Hence any kind of change in feeding behavior, swimming or skin color is a sign of unfavorable conditions, stress. Analyzing unusual behavior can provide an early warning of its health status. In this project 5 parameters (fish density, speed, direction, angle and depth) are calculated that give insights into fish health. Additionally, to overcome the typical scarcity of available data in this field, with the help of our industry partners we collected and prepared datasets from scratch. And using these algorithms industrial software solutions can be developed for an improved fish health monitoring in fish farms. These advances will facilitate the production of environmentally and economically more sustainable fish, while promoting animal welfare.
