Hyperspectral Imaging-Based Adulteration Detection of Edible Oil using Optimized Machine Learning Models

Technical Summary: Hyperspectral Imaging-Based Adulteration Detection of Edible Oil using Optimized Machine Learning Models

Problem Statement
Edible oil adulteration, involving the deliberate mixing of superior oils with cheaper or inedible components, poses significant risks to public health and compromises food quality. Traditional detection methods, such as High-Performance Liquid Chromatography (HPLC) and Gas Chromatography (GC), are often time-consuming, expensive, require toxic reagents, and involve complex sample preparation, making them unsuitable for high-throughput or real-time industrial screening. There is a critical need for rapid, non-destructive, and scalable technologies capable of identifying contamination in complex food matrices.

Methodology
The study proposes a non-destructive framework integrating Hyperspectral Imaging (HSI) with optimized Machine Learning (ML) models to detect adulteration in edible oils.

• Data Acquisition: A dataset comprising 480 oil samples was generated using a Specim FX10 HSI camera system (spectral range: 400–1000 nm). The samples included pure oils (tea tree, rose, turmeric, rosehip, lemon, garlic, vitamin E, and orange) and their adulterated variants mixed at four concentration levels (2 ml, 4 ml, 8 ml, and 16 ml).
• Preprocessing: To ensure data quality, the Savitzky–Golay (SG) filter (3rd-order polynomial, window size 11) was applied to remove noise and smooth spectral curves while preserving peak shapes. Reflectance data were converted to absorption spectra using the empirical line method.
• Feature Optimization: Instead of using the full spectral dataset (224 bands), the study employed a feature selection strategy to identify the top 20 most informative spectral bands. This dimensionality reduction was performed using multi-criteria methods to enhance computational efficiency.
• Machine Learning Models: Four classification models were trained and evaluated on the optimized dataset: Linear Discriminant Analysis (LDA), CatBoost, Gradient Boosting Machine (GBM), and Extreme Gradient Boosting Machine (XGBM). The dataset was split into training and testing sets using an 80:20 ratio.

Key Contributions

1. Non-Destructive HSI-ML Framework: Development of an end-to-end method for edible oil quality identification that utilizes radiometric preprocessing and SG filtering to generate reliable spectral signatures.
2. Comprehensive Dataset: Creation of a robust hyperspectral dataset containing 120,000 spectral data points derived from 480 samples (160 pure and 320 adulterated), covering multiple oil types and adulteration levels.
3. Optimized Feature Selection: Demonstration that selecting the top 20 spectral bands significantly reduces data dimensionality and system complexity while maintaining high discriminative power, facilitating real-time industrial implementation.
4. Model Performance Benchmarking: A comparative analysis of linear and ensemble learning techniques, identifying XGBM as the superior model for this specific application.

Results
The experimental results indicated that the Smoothing + XGBM model achieved the highest performance across all metrics:

• Accuracy: 100% training accuracy, with 98% accuracy on both validation and testing sets.
• Metrics: The model demonstrated high precision, recall, F1-score, and balanced accuracy (98%). It also achieved low log loss (5%) and Hamming loss (2%), indicating high confidence and low misclassification rates.
• Comparison:
  • GBM: Showed strong performance with ~97% accuracy but exhibited signs of overfitting compared to XGBM.
  • LDA: Achieved moderate performance (93% accuracy), struggling with complex nonlinear spectral relationships, particularly at low adulteration levels.
  • CatBoost: Yielded slightly lower accuracy (~91%) and higher error rates compared to the other ensemble models.
• Feature Importance: Spectral significance analysis revealed that XGBM effectively prioritized specific, highly discriminative wavelengths (e.g., indices around 48, 146, and 155), whereas LDA relied on a broader, less selective distribution of bands.

Significance and Claims
The paper asserts that the proposed HSI-integrated approach offers a fast, reliable, and scalable solution for detecting edible oil adulteration. By combining spectral feature optimization with advanced ensemble learning (specifically XGBM), the method achieves high classification accuracy while minimizing computational costs. The authors claim this framework is suitable for real-time applications in the industrial sector for food safety and quality control.

The study acknowledges limitations, noting that the current framework is restricted to the Visible-Near Infrared (VNIR) region (400–1000 nm), which primarily captures surface information. The authors suggest that future work should extend the analysis to the NIR-SWIR region (1000–2500 nm) to enhance molecular-level characterization and further improve detection accuracy.