Aggressiveness Trait Identification in Portunus trituberculatus via a Self-supervised Computational Ethology Framework

High-density aquaculture of Portunus trituberculatus is severely impacted by aggressive behaviour, leading to injury and cannibalism. Despite its ecological and economic importance, aggressiveness is still commonly assessed using manual observation and arbitrary thresholds, which fail to capture the high-dimensional and continuous nature of behavioural traits. Here, we propose a deep learning-based self-supervised computational ethology framework for the automated and objective identification of aggressiveness traits from behavioural data. The framework integrates an augmentation-free self-supervised representation learning strategy with probabilistic clustering, allowing intrinsic behavioural similarity to be inferred without predefined labels or thresholds. Applying this approach to mirror-test behavioural data, we identify a stable ternary aggressiveness personality structure, challenging the binary identification paradigms commonly adopted in crustacean behavioural studies. Importantly, the biological relevance of the identified traits is independently evaluated through paired agonistic contests: among the methods tested, our framework yields statistically significant separation among aggressiveness levels ( P < 0.05). These results suggest that, under the present experimental conditions, aggressiveness in P. trituberculatus may be more appropriately represented as a multi-level personality trait rather than a strictly binary attribute. Our framework provides a robust quantitative tool for automated aggressiveness identification in precision aquaculture and illustrates how computational approaches can be meaningfully integrated with experimental ethology.