Improved research on coral bleaching detection model based on FCOS model.

Area of Science:

• Marine Biology
• Computer Vision
• Environmental Monitoring

Background:

• Coral reef health assessment is critical for marine ecosystem preservation.
• Existing coral bleaching detection methods often lack accuracy and real-time capabilities.
• Efficient and accurate detection models are needed for timely intervention and conservation efforts.

Purpose of the Study:

• To introduce an improved deep learning model, FCOS_EfficientNET, for enhanced coral bleaching detection.
• To optimize model performance in terms of accuracy, recall, and real-time processing speed.
• To develop adaptable model variants for diverse marine monitoring applications.

Main Methods:

• Utilized EfficientNet as the backbone for optimized parameter throughput.
• Implemented ReLU activation, cosine similarity, and softmax for dataset weighting.
• Modified attention structure and integrated BiFPN for improved feature extraction and reduced memory consumption.
• Developed four model variants (FCOS_EfficientNETb0-b3) and employed an improved training strategy.

Main Results:

• FCOS_EfficientNETb3 achieved 48.5% mAP on MS COCO and 81.5% accuracy with 59.3% recall on a custom coral bleaching dataset.
• FCOS_EfficientNETb0 demonstrated a high frame rate of 167.17 fps.
• Variants FCOS_EfficientNETb1 and FCOS_EfficientNETb2 provide a balance of performance metrics for mobile and edge computing.

Conclusions:

• The FCOS_EfficientNET series significantly enhances coral bleaching detection accuracy and real-time performance.
• Model variants offer tailored solutions for different marine monitoring scenarios, including tracking marine traffic.
• These advancements contribute to more effective coral reef health assessment and conservation strategies.