ALMI-A Generic Active Learning System for Computational Object Classification in Marine Observation Images.

Area of Science:

• Marine biology
• Computer science
• Data science

Background:

• Cabled Fixed Underwater Observatories (FUOs) generate vast amounts of image data.
• Manual analysis of this data for species identification is time-consuming and inefficient.
• Current methods often require extensive annotated datasets, which are difficult to obtain.

Purpose of the Study:

• To develop a more efficient machine learning strategy for interpreting underwater observatory images.
• To reduce the dependency on large annotated datasets for training.
• To improve the speed and accuracy of quantitative information extraction from image time series.

Main Methods:

• Proposed a novel strategy combining active learning and deep learning feature representations.
• Developed a machine learning-based approach named ALMI.
• Tested the method on two distinct underwater image datasets.

Main Results:

• ALMI achieved high classification accuracy (>90%) with fewer than 258 training samples on one dataset.
• ALMI achieved >80% accuracy after only 150 training iterations on another dataset.
• The method outperformed a reference approach in both accuracy and training data efficiency.

Conclusions:

• The proposed active learning-based machine learning interpretation (ALMI) method significantly enhances training efficiency for underwater image analysis.
• ALMI offers a practical solution to the bottleneck of manual data analysis in marine observatories.
• This approach enables more effective utilization of image time series data for ecological studies.