Image dataset for benchmarking automated fish detection and classification algorithms

Area of Science:

• Marine Ecology
• Oceanography
• Computer Science

Background:

• Multiparametric video-cabled marine observatories are essential for real-time remote monitoring of marine ecosystems.
• Continuous, high-frequency image data from these platforms necessitate automated methods for biological time-series extraction.

Purpose of the Study:

• To generate coastal fish time series using data from the OBSEA (Oceanic Buoy System for Environmental Acoustic monitoring) observatory.
• To create a comprehensive dataset of tagged fish and associated environmental conditions for AI development.

Main Methods:

• The OBSEA observatory collected continuous 24-h imagery from 2013-2014 at a depth of 20 meters.
• Image analysis involved tagging 69,917 fish across 30 taxa.
• Meteorological and oceanographic data were collected and quality-controlled to correlate with image quality.

Main Results:

• A substantial dataset of 69,917 tagged fish from 30 taxa was generated.
• Environmental data provided context for observed fish populations.
• The study successfully produced continuous coastal fish time series.

Conclusions:

• The tagged fish dataset is valuable for developing Artificial Intelligence (AI) algorithms for automated fish identification and classification.
• This approach enables efficient analysis of extensive time-lapse image sets from marine observatories.
• Automated analysis is key to extracting meaningful biological insights from large-scale marine monitoring data.