Quantifying the Effectiveness of Multiple Fish Behaviors for Hypoxia Monitoring Across Different Environments

The analysis of behavioral indicators using computer vision offers a cost-effective and efficient approach for monitoring fish hypoxia. However, most existing studies primarily focus on extracting behavioral indicators and lack systematic evaluation of their effectiveness in hypoxia monitoring. Identifying key behavioral indicators that are highly correlated with dissolved oxygen (DO) levels and demonstrate strong environmental robustness remains a significant challenge. To address this issue, a multi-behavior quantification model is proposed. This model defines and measures eight commonly used indicators: respiration rate, posture, vertical distribution and position, swimming speed and angle, and horizontal distribution and position. The effectiveness of these indicators for hypoxia monitoring is then evaluated across different hardware configurations (surface top view and underwater side view) and environmental conditions (diurnal cycle, temperature variations, and ammonia nitrogen concentration) by calculating the mutual information (MI) between each behavioral indicator and DO. Results indicate that from the side view, the vertical position emerges as the most effective indicator for hypoxia monitoring, demonstrating the highest accuracy, generalizability, and computational efficiency. It achieves the highest MI (1.20) with DO and significantly outperforms other indicators (p < 0.01). From the top view, swimming speed (MI = 1.03) and horizontal distribution (MI = 1.07) are identified as the most effective indicators, both of which are significantly higher than the angle indicator (p < 0.05). Regarding computational complexity, horizontal distribution requires the lowest computational cost, while swimming speed demonstrates greater robustness under varying environmental conditions.