Automatic Hierarchical Classification of Kelps Using Deep Residual Features.

Area of Science:

• Marine biology
• Computer science
• Remote sensing

Background:

• Remote image data is crucial for assessing marine benthic communities globally.
• Manual identification of organisms in this data is time-consuming, limiting its exploitation.
• Automated and machine learning methods are needed to accelerate marine biota classification.

Purpose of the Study:

• To present an automatic hierarchical classification method for identifying kelp in underwater images.
• To compare the proposed method against traditional flat and parallel multi-class classification techniques.
• To apply the method for studying changes in kelp cover over time.

Main Methods:

• Developed an automatic hierarchical classification approach using local binary classification.
• Employed deep residual networks for feature extraction, outperforming traditional CNN and hand-crafted features.
• Compared hierarchical training approaches, finding the sibling hierarchical training superior.

Main Results:

• The hierarchical classification method significantly outperformed traditional parallel multi-class classifications (e.g., 90.0% vs. 57.6% on Benthoz15).
• Learned features from deep residual networks showed superior performance compared to off-the-shelf CNN and hand-crafted features.
• The sibling hierarchical training approach demonstrated a significant performance advantage over the inclusive hierarchical approach.

Conclusions:

• The proposed automatic hierarchical classification method is effective for classifying kelp in underwater imagery.
• This method offers a significant improvement over existing classification techniques for marine benthic surveys.
• The approach has practical applications in monitoring kelp cover changes over time using autonomous underwater vehicle data.