Multi-label classification for multi-temporal, multi-spatial coral reef condition monitoring using vision foundation...

Area of Science:

• Marine Biology
• Computer Science
• Artificial Intelligence

Background:

• Coral reefs provide vital ecosystem services but are threatened by climate change and human activities.
• Deep learning models struggle with complex underwater images for coral reef condition classification.
• Vision foundation models offer high accuracy but require extensive resources for fine-tuning.

Purpose of the Study:

• To develop an efficient method for classifying coral reef conditions using foundation models.
• To address the computational and carbon footprint challenges of fine-tuning large models.
• To evaluate the performance and generalizability of adapted foundation models for ecological image analysis.

Main Methods:

• Integration of the DINOv2 vision foundation model with Low-Rank Adaptation (LoRA) fine-tuning.
• Utilized multi-temporal underwater images from 15 dive sites at Koh Tao, Thailand.
• Employed universal labeling standards for citizen science-based conservation programs.

Main Results:

• The DINOv2-LoRA model achieved a superior match ratio of 64.77%, outperforming conventional models (60.34%).
• LoRA significantly reduced trainable parameters from 1136.50M to 5.91M.
• Demonstrated exceptional generalizability across different seasons and spatial settings through transfer learning.

Conclusions:

• The DINOv2-LoRA approach offers an efficient and accurate method for coral reef condition classification.
• This study is the first to explore efficient foundation model adaptation for multi-label coral reef classification in multi-temporal and multi-spatial contexts.
• The proposed method provides a valuable tool for coral reef monitoring, conservation, and management.