BenthicNet: A global compilation of seafloor images for deep learning applications
Scott C Lowe1, Benjamin Misiuk2,3, Isaac Xu4
1Vector Institute, Toronto, Ontario, Canada.
Scientific Data
|February 7, 2025
Summary
BenthicNet is a new global dataset of seafloor images to train machine learning models for analyzing underwater ecosystems. This resource aids in understanding marine environments more efficiently.
Area of Science:
- Marine Biology
- Ecology
- Computer Science
Background:
- Underwater imaging generates vast seafloor datasets crucial for monitoring benthic ecosystems.
- Analyzing these large image datasets is a bottleneck, limiting environmental information mobilization.
- Developing efficient machine learning (ML) models requires large, consistent datasets, which are currently scarce.
Purpose of the Study:
- To present BenthicNet, a global compilation of seafloor imagery.
- To support the training and evaluation of large-scale image recognition models for benthic analysis.
- To provide an open-access resource for advancing ML applications in marine science.
Main Methods:
- Collected and curated over 11.4 million seafloor images, with a representative subset of 1.3 million images.
- Generated 3.1 million annotations using the CATAMI scheme across 190,000 images.
- Trained a large deep learning model on the BenthicNet compilation.
Main Results:
- BenthicNet offers a diverse representation of seafloor environments.
- Preliminary results indicate the trained deep learning model can automate image analysis tasks.
- The dataset and model demonstrate utility for both large- and small-scale benthic image analysis.
Conclusions:
- BenthicNet addresses the scarcity of large, consistent datasets for ML in marine ecology.
- The open availability of BenthicNet and the trained model facilitates reuse and further research.
- This resource has the potential to significantly improve the efficiency of seafloor ecosystem monitoring.
