Food fraud detection in Octopus mimus using hyperspectral imaging and machine learning techniques

Among the many seafood products susceptible to food fraud, octopus (Octopus mimus) is often replaced by giant squid (Dosidicus gigas) due to its higher market value and desirable nutritional and sensory properties. As a result, the development of rapid and noninvasive techniques for food quality assessment is of significant interest to the food industry. In this research, we evaluated the capability of two hyperspectral imaging systems to discriminate between octopus and giant squid meat. For the experimentation, samples of both species were acquired from a local port, and the arms were selected as the most similar parts between species. After cleaning and removing the skin, 300 cuts were extracted from each species and subdivided into three groups of 100, representing fresh, frozen, and cooked samples, respectively. Hyperspectral images were acquired in two spectral ranges: the visible and near-infrared (Vis-NIR, 400-1000 nm) and the near-infrared (NIR, 900-1700 nm) ranges. Mean spectral profiles were extracted and preprocessed with smoothing and normalization. An exploratory principal component analysis (PCA) was applied to the data. Subsequently, three classification models were implemented: linear discriminant analysis (LDA), support vector machine (SVM), and k-nearest neighbors (k-NN), with 30 repetitions and cross-validation applied for robustness. These models were created using the full spectrum and were then optimized. The results indicate that the highest accuracy was achieved using LDA classifiers, with smoothing and normalization as preprocessing steps. The accuracy reached 100% for the Vis-NIR range and 98.3% for the NIR range. After optimization, which involved selecting relevant wavelengths using feature selection algorithms, LDA classifiers again provided the best results. Accuracy improved to 99.9% for the Vis-NIR range and 94.8% for the NIR range, using only ten relevant wavelengths. These findings suggest that LDA-based classifiers are the most effective for distinguishing between octopus and giant squid meat. When optimization is required, feature selection using ReliefF or correlation-based feature subset selection algorithms is recommended. These results have important implications for the seafood industry, where accurate species identification is critical for ensuring quality control and food safety.