Hierarchical Dual-Model Detection Framework for Spotted Seals Using Deep Learning on UAVs.

Area of Science:

• Marine Biology
• Ecology
• Computer Science

Background:

• Accurate monitoring of marine endangered species like spotted seals is crucial for conservation.
• Traditional monitoring methods face challenges including weak target features and background interference.
• Limited edge computing capacity on Unmanned Aerial Vehicles (UAVs) hinders real-time data processing.

Purpose of the Study:

• To develop a hierarchical dual-model deep learning framework for accurate spotted seal monitoring.
• To address challenges in UAV-based ecological monitoring, including computational limitations and detection accuracy.
• To provide an efficient technical solution for long-term monitoring of marine endangered species.

Main Methods:

• Deployment of an optimized FF-YOLOv10 lightweight model on UAVs for rapid target localization.
• Utilization of an enhanced PP-YOLOv7 model on ground stations for precise detection.
• Implementation of a hierarchical framework combining edge and ground station processing.

Main Results:

• The FF-YOLOv10 model achieved a 24.2% reduction in computational complexity and a 33.3% increase in inference speed.
• The PP-YOLOv7 model demonstrated 94.2% precision and a 1.9% increase in recall rate.
• The dual-model framework significantly improved detection efficiency and accuracy for spotted seals.

Conclusions:

• The proposed framework offers an efficient and precise technical solution for marine endangered species monitoring.
• This approach supports habitat conservation policy formulation and ecosystem health assessments.
• Deep learning on UAVs presents a viable strategy for ecological monitoring in challenging environments.