AI Development & Model Performance

Deep Learning Model & Training Pipeline

At the core of AquaTRAC AI lies the YOLOv8 (You Only Look Once, Version 8) object detection framework—selected for its optimal balance between high detection accuracy and real-time processing capability, making it ideal for continuous environmental monitoring and edge deployment.

AI Training Pipeline

Our model development follows a structured and adaptive training workflow:

  1. Model Configuration and Optimization
    The YOLOv8 architecture is carefully customized to address the unique challenges of aquatic environments, including:
    • Variable water clarity and turbidity
    • Complex and dynamic backgrounds
    • Low-visibility and low-contrast conditions
  2. Initial Training (River-Specific Learning)
    The model is first trained using curated datasets from the Yodo River Basin, enabling:
    • Accurate identification of local fish species
    • Adaptation to site-specific environmental characteristics
    • High precision in controlled ecological conditions
  3. Future Enhancement (River–Sea Transfer Learning)
    To further improve model robustness and generalization:
    • Training datasets from marine environments (e.g., collaborative datasets from Kobe University partners) will be incorporated
    • Cross-domain learning will be applied to enhance feature extraction capabilities
    • The model will then be fine-tuned back to river datasets, improving performance under diverse and unseen conditions

This hybrid strategy ensures the development of a scalable and transferable AI model, capable of deployment across different aquatic ecosystems.

Training Results & Model Evaluation

The initial training phase using localized Yodo River datasets has yielded highly promising and scientifically robust results, demonstrating the effectiveness of the AquaTRAC AI framework.

Key Performance Indicators

  • Model Loss (Training & Validation)
    Both training and validation loss curves exhibit a consistent and stable decrease, indicating:
    • Efficient model convergence
    • Strong generalization to unseen data
    • Minimal overfitting
  • Classification Accuracy
    • Top-1 Accuracy: Exceeding 99%, reflecting extremely high precision in identifying the correct species
    • Top-5 Accuracy: Reaching 100%, ensuring reliable performance even in complex or ambiguous cases
  • Confusion Matrix Analysis
    Evaluation results show:
    • High confidence across all target species classes
    • Minimal misclassification, including among visually similar species
    • Strong class separation and model reliability

Performance Significance

These results confirm that AquaTRAC AI can deliver high-precision, real-time detection and classification under realistic environmental conditions. The model’s strong performance provides a solid foundation for:

  • Scaling to larger datasets and extended deployment areas
  • Integrating behavioral and ecological analytics
  • Supporting decision-making in biodiversity conservation and river management