A data-driven binary-classification framework for oil fingerprinting analysis.

Area of Science:

• Environmental Science
• Analytical Chemistry
• Data Science

Background:

• Marine oil spills pose significant environmental and societal challenges.
• Oil dispersants complicate traditional oil fingerprinting methods.
• Accurate oil fingerprinting is crucial for effective oil spill response and source identification.

Purpose of the Study:

• To introduce machine learning (ML) for oil fingerprinting analysis in the context of dispersant-treated oil.
• To develop and compare ML models for distinguishing between different crude oil types.
• To identify the most suitable biomarkers and ML algorithms for robust oil fingerprinting.

Main Methods:

• A data-driven binary classification framework using ML algorithms was developed.
• Dimensionality reduction (Principal Component Analysis - PCA) was applied.
• Five biomarker groups (terpanes, steranes, TA-steranes, MA-steranes, diamantanes) were analyzed using six ML algorithms (KNN, SVC, RFC, DTC, LRC, EVC).
• Hyperparameter optimization and cross-validation (GridSearchCV) were employed to enhance model accuracy and prevent overfitting.

Main Results:

• The Random Forest Classifier (RFC) algorithm using diamantanes achieved the highest F-score (0.871).
• Principal Component Analysis (PCA) with 95% variance was utilized.
• Diamantanes were identified as the most effective biomarker for distinguishing weathered crude oil (WCO) and chemically dispersed oil (CDO).

Conclusions:

• ML-aided oil fingerprinting offers a powerful tool for oil spill source identification.
• Diamantanes are recommended as the primary biomarker for distinguishing WCO and CDO in dispersant-affected spills.
• This study highlights the significant value of ML in advancing oil spill response research and practical applications.