Envisage phenotyping : integrating artificial intelligence in image analysis for selective breeding in aquaculture

Automated phenotyping is a rapidly evolving interdisciplinary field focused on high-throughput, non-invasive phenotype collection to support breeding and management decisions in aquaculture. Current development primarily targets production traits, leaving a gap in traits linked with metabolism, health and physiological well-being. These traits are complex and difficult to measure, but advancements in AI-based prediction models combined with imaging technologies provide opportunities. This thesis used fish image data collected in breeding facilities, to broaden the scope of predicted phenotypes that contribute to the breeding goals for these species.For production traits, a structural framework was introduced to determine optimal image-based prediction strategies, demonstrated in gilthead seabream (Sparus aurata). The framework captures both simple and high-dimensional relationships between images and target traits. Interpretation of the model was provided by visualizing predictive imaginal features.To dissect the genetic basis of swimming performance in rainbow trout (Oncorhynchus mykiss), a novel workflow identified accessible and heritable morphological indicators. Among fish of the same weight, relatively larger and broader epaxial muscles, larger heads, and smaller caudal fins were associated with poorer swimming performance. This study illustrates how images and AI can be interpreted physiologically using Explainable AI.Individual re-identification is essential for longitudinal traits. Although image-based re-identification is promising, it remains infeasible for Atlantic salmon (Salmo salar) under realistic aquaculture conditions due to variation in the phenotyping environment, especially lighting, and limited phenotypic stability. In search for solutions, a review was conducted on the current state of vision-based re-identification in aquaculture. Limitations in data acquisition, data analysis, and practical implementation were identified. Without high-tech data acquisition, multi-dimensional data quality frameworks, and benchmarking methods, vision-based re-identification is found too immature to replace PIT tags. Future progress depends on dataset standardization, methodological refinement, and strong alignment with aquaculture systems.The objectivity of qualitative phenotypic measurements using images was improved and applied to body shape in gilthead seabream (Sparus aurata). Two moderately heritable traits, “Distance to the Best” (DtB) and “Distance to the Worst” (DtW), were derived for the deviation of each fish from an ideal or undesirable shape. This approach is objective, automated, visually intuitive, and flexible for updating target shapes.This thesis emphasizes that successful integration of AI and image analysis requires an understanding of the entire breeding program, not just phenotypic data collection and analysis. This thesis contributes advanced phenotyping and decision-support AI tools to the ongoing integration of emerging AI trends to improve animal production, health, and welfare.