Deepdive: Leveraging Pre-trained Deep Learning for Deep-Sea ROV Biota Identification in the Great Barrier Reef

Area of Science:

• Marine Biology
• Computer Science
• Ecosystem Monitoring

Background:

• Deep-sea ecosystems are vital and require effective conservation strategies.
• Automated classification of deep-sea biota aids in creating habitat maps for biodiversity and ecosystem health assessments.
• Large, labeled datasets are crucial for training robust deep learning models.

Purpose of the Study:

• To establish a significant dataset for deep-sea remotely operated vehicle (ROV) image classification.
• To benchmark deep learning models for classifying deep-sea biota within this dataset.
• To support biodiversity assessments and ecosystem health evaluations in deep-sea environments.

Main Methods:

• Manual labeling of 3994 ROV images featuring deep-sea biota across 33 classes using human-in-the-loop quality control.
• Implementation and benchmarking of deep learning models including ResNet, DenseNet, Inception, and Inception-ResNet.
• Evaluation of model performance on a dataset characterized by class imbalance.

Main Results:

• The Inception-ResNet model achieved a mean classification accuracy of 65% on the deep-sea biota dataset.
• Area Under the Curve (AUC) scores exceeded 0.8 for each class, indicating strong classification performance.
• The developed dataset and model benchmarks provide a foundation for further deep-sea image analysis.

Conclusions:

• The created dataset is a valuable resource for advancing deep-sea image classification research.
• Deep learning models, especially Inception-ResNet, demonstrate significant potential for automated analysis of deep-sea imagery.
• This work contributes to improved methods for monitoring and conserving deep-sea ecosystems.