Coral-YOLO: An Intelligent Optical Vision Sensing Framework for High-Fidelity Marine Habitat Monitoring and...

Area of Science:

• Marine biology
• Artificial intelligence
• Computer vision

Background:

• Coral reefs face catastrophic decline from climate-induced bleaching, threatening marine biodiversity.
• Automated monitoring is crucial, but current object detectors struggle with complex underwater scenes due to spatial and feature robustness deficits.

Purpose of the Study:

• To develop a novel framework, Coral-YOLO, for improved detection and forecasting of coral reef health.
• To address limitations in spatial reasoning and feature robustness in underwater object detection.

Main Methods:

• Introduced the Holistic Attention Block Head (HAB-Head) for deep cross-scale reasoning.
• Implemented MCAttention, a randomized training mechanism for scale-invariant and robust features.
• Utilized a newly curated, multi-year CR-Mix dataset for evaluation.

Main Results:

• Coral-YOLO achieved state-of-the-art 50.3% AP, outperforming the YOLOv12-m baseline by +1.8%.
• Demonstrated significant gains in detecting small objects (+2.6% in APS).
• The temporal forecasting module achieved 82.7% accuracy in predicting future coral health.

Conclusions:

• Coral-YOLO sets a new benchmark for automated coral reef monitoring and forecasting.
• The framework enables proactive reef conservation by identifying at-risk corals earlier.
• Coral-YOLO offers a powerful tool to combat the decline of critical marine ecosystems.