Optimized Stacking Ensemble Machine Learning for Nondestructive Toxicological Risk Assessment of Fish Meat Using Hyperspectral Imaging

Abstract

The increasing occurrence of chemical contaminants and preservative misuse in fish products poses significant challenges to food safety and public health. Conventional analytical techniques, including gas chromatography–mass spectrometry (GC–MS) and distillation‐based chromatography, provide reliable contaminant detection but are often limited by high operational costs, extensive sample preparation, destructive testing, and dependence on specialized expertise. To address these limitations, this study presents an optimized stacking ensemble learning (OSEL) framework integrated with hyperspectral imaging (HSI) for rapid and nondestructive toxicological risk assessment of fish meat. Hyperspectral data were acquired using a Specim FX10 camera covering the 400–1000 nm spectral range and preprocessed with the Savitzky–Golay filter to improve spectral quality and reduce noise. Classification performance was evaluated using both a reduced subset of 25 informative spectral bands and the complete set of 224 bands. Multiple machine learning algorithms were investigated, including logistic regression, decision trees, support vector machines, linear discriminant analysis variants, and the proposed OSEL model. Among all evaluated models, OSEL demonstrated superior predictive capability, achieving 98.9% validation accuracy and 99.1% fivefold cross‐validation accuracy using only 25 spectral bands. The model further attained a Jaccard index of 97%, a high Cohen's kappa, F1‐score, and MCC values, with zero Hamming loss. Using the complete 224‐band dataset, OSEL achieved 100% classification accuracy across all evaluation metrics. These findings demonstrate the potential of combining HSI with ensemble learning as an efficient, accurate, and scalable tool for nondestructive toxicological screening of fish meat, supporting enhanced food safety monitoring and quality assurance.