New interactive machine learning tool for marine image analysis.

Area of Science:

• Marine Biology
• Image Analysis
• Machine Learning

Background:

• Marine imaging technologies generate vast datasets, but manual analysis is time-consuming.
• Efficient analysis of marine video and image data is crucial for ecological research.
• Identifying species like Mycale lingua in large datasets presents a significant challenge.

Purpose of the Study:

• To demonstrate the effectiveness of the open-source RootPainter tool for analyzing marine image datasets.
• To assess RootPainter's capability in extracting presence and surface area data for Mycale lingua.
• To evaluate the impact of new corrective annotation metrics on model training and validation.

Main Methods:

• Utilized RootPainter, an interactive machine learning tool, on two marine image datasets (time-lapse and ROV video).
• Developed and applied corrective annotation metrics for objective model training assessment.
• Employed transfer learning to enhance analysis efficiency for Mycale lingua models.

Main Results:

• Achieved high accuracy in identifying Mycale lingua with an average Dice score of 0.94 ± 0.06.
• Analysis efficiency increased 6 to 16 times faster than manual annotation using transfer learning.
• Successfully extracted surface area measurements for Mycale lingua from both datasets.

Conclusions:

• RootPainter offers a rapid and accurate solution for analyzing large marine image datasets.
• Interactive machine learning tools can significantly improve the speed and scope of biodiversity research.
• Future applications include enhanced understanding of spatiotemporal patterns in marine biodiversity.