A Novel Machine-Learning Framework Based on a Hierarchy of Dispute Models for the Identification of Fish Species...

Area of Science:

• Analytical Chemistry
• Machine Learning
• Food Science

Background:

• Global seafood mislabeling rates reach approximately 20%, posing significant health, economic, and environmental risks.
• Traditional fish identification methods like DNA analysis and Polymerase Chain Reaction (PCR) are costly, time-consuming, and require specialized expertise and equipment.

Purpose of the Study:

• To develop a rapid, accurate, and cost-effective method for identifying fish species, addressing the limitations of current techniques.
• To enhance the accuracy of fish species identification using a combination of spectroscopy and a novel machine learning framework.

Main Methods:

• Employed three spectroscopic modes: fluorescence (Fluor), visible near-infrared (VNIR), and short-wave near-infrared (SWIR) across 43 fish species.
• Developed a hierarchical machine learning framework with specialized classifiers for groups of similar fish types.
• Integrated global and dispute classification models to create a decision process that improves overall classification accuracy.

Main Results:

• Spectroscopy accuracies improved: Fluor from 80% to 83%, VNIR from 75% to 81%, and SWIR from 49% to 58%.
• Certain species showed identification accuracy increases of up to 40% with single-mode identification.
• Fusion of all three spectroscopic modes boosted the best single-mode performance by an average of 9% across all species.

Conclusions:

• The proposed spectroscopy and hierarchical machine learning method provides a real-time, accurate alternative to traditional fish identification techniques.
• This novel hierarchical system of dispute models is a versatile machine learning tool applicable to various classification problems with numerous classes.