Harnessing the power of Squidle+ to develop flexible machine learning models

Area of Science:

• Marine Ecology
• Computer Science
• Remote Sensing

Background:

• Manual annotation of seafloor imagery is a bottleneck in ecological monitoring due to time and cost constraints.
• Existing machine learning (ML) models often lack generalizability due to training on small, homogeneous datasets.

Purpose of the Study:

• To develop and validate scalable ML models for analyzing large, heterogeneous seafloor image datasets.
• To improve the efficiency and accuracy of ecological monitoring using ML classifiers.

Main Methods:

• Assembled a large, heterogeneous dataset (1.7 million annotations, 150,000 images) from 325 Australian expeditions using the Squidle+ platform.
• Developed a novel neighbour-distribution strategy and adaptive cropping for effective model training on diverse image resolutions.
• Implemented a flexible training pipeline excluding entire campaigns for independent national-scale performance assessment.

Main Results:

• Achieved high test accuracies: 85%-93% for Essential Ocean Variables (EOVs) including Seagrass, Macroalgae, and Hard Coral cover.
• Reached 96.5% accuracy for the species-specific label Ecklonia radiata on campaign test sets across Australia.
• Demonstrated robust, national-scale performance of ML models trained on diverse, large-scale datasets.

Conclusions:

• The developed ML models provide a scalable and accurate solution for seafloor image analysis, overcoming previous data limitations.
• Publicly available code and models integrated into Squidle+ reduce barriers for ecological monitoring programs.
• This study establishes a new benchmark for ML application in marine ecological assessments using large, diverse datasets.