Agricultural University of Athens
Determination of minced meat quality using machine learning
Abstract
dc:descriptionCurrently, consumers expect constant reassurance of the origin, quality and compliance to label of the food products they purchase. Therefore, food industries, retailers and authorities have to develop advanced, effective and relatively low-cost solutions for quality assurance and detection of fraudulent practices. In this context, this study focuses on (a) Multispectral Imaging (MSI), (b) Fourier Transform Infrared (FTIR) spectrometry and (c) the application of advanced data analysis and machine learning methodologies. In terms of quality, the efficacy of the abovementioned methods concerning (a) non-compliance to label/ fraud detection and (b) microbiological spoilage is examined and the following analyses took place:In the first case, minced beef adulteration with pork was investigated. MSI data from 220 meat samples from four independent experiments (different meat batches) were extracted. Adulteration was performed with a 10% w/w step, creating 11 classes of samples (including pure beef and pork). After an image preprocessing step, Hierarchical Cluster Analysis (HCA) and Principal Component Analysis (PCA) were applied. Meat batches displayed significant differences and different classes were less distinguishable when the first three batches were included in the analysis. After partitioning in training and validation sets, the fourth batch was retained for independent/ external validation and Linear Discriminant Analysis (LDA) and Partial Least Squares Discriminant Analysis (PLSDA) were performed for classification among 11 classes and pure vs. adulterated samples. For the case of 11 classes, 98.48% and 96.97% of the samples were classified within a ±10% category of adulteration for LDA and PLSDA respectively, whereas the 3-class case yielded 98.48% overall correct classification (OCC). Results for the external validation set, proved LDA significantly less accurate compared to PLSDA where all samples were classified correctly, proving that 10% is within the method’s detection limit.In the second case, with 110 samples from three different batches of minced beef and horsemeat and additional images captured after 6, 24 and 48 h, model performance is investigated in terms of detection of adulteration. PCA was used for visualization purposes, while PLSDA and Random Forest (RF) for classification among different percentages of beef (4 classes), pure beef vs. pure horsemeat vs. adulterated samples, pure beef and horse vs. adulterated samples and freshly-ground vs. stored minced meat. Models significantly underperformed in independent validation. In the end, a two stage Support Vector Machine (SVM) methodology was utilized, where freshly-ground samples are separated from stored and then pure separated from adulterated. The OOC of the SVM model was equal to 95.31% for independent model validation.In the third case, multispectral images and FTIR spectra from seven different batches of freshly-ground beef, along with MSI and FTIR spectral data after being frozen (-20°C) for 7 and 32 days and then thawed were acquired (in total, 105 measurements per sensor). In terms of data analysis methods, PCA was used for data exploration, while PLSDA and SVM yielded 100% correct classification between fresh and frozen-then-thawed samples for MSI test and external validation sets. FTIR proved less accurate, as PLSDA yielded 93.3 and 96.7% classification accuracy for the test and external validation set, respectively.In cases 4 and 5, meat and specifically minced beef spoilage is explored. In case 4, using “MeatReg”, a web-based application, seven methods were tested for the prediction of bacterial counts. The dataset included 105 samples, stored in air or under modified air packaging (MAP) conditions (20% CO2/ 80% O2) at 4 and 10°C, microbiologically analyzed (Pseudomonads, Lactobacilli, B. thermosphacta and Enterobacteriaceae, as well as TVC). FTIR and MSI data were compared with electronic nose, High Performance Liquid Chromatography and Gas Chromatography coupled to Mass Spectrometry data. Results were mixed depending on the sensor and species counts, while RF regression yielded an overall good performance.Similarly to case 4, in case 5, 168 minced beef samples were analyzed for TVC, while FTIR measurements took place. Samples were stored in air or under modified air packaging (MAP) conditions at 4 and 10°C. An ensemble-based approach was employed where spoilage estimation was a fusion of several artificial neural networks yielding a mean squared error equal to 0.16 (log CFU/g)^2.
Degree
thesis:*- Grantor dc:publisher
- Agricultural University of Athens
- Year dc:date
- 2017
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Ropodi, Athina
- Ροπόδη, Αθηνά
Subjects
dc:subject × 18- Ποιότητα κρέατος
- Ανάλυση δεδομένων
- Μηχανική μάθηση
- Πολυφασματική απεικόνιση
- Φασματοσκοπία υπερύθρου με μετασχηματισμό Fourier
- Meat quality
- Data analysis
- Machine learning
- Multispectral imaging
- Fourier transform infrared spectroscopy (FTIR)
- Φυσικές Επιστήμες
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Γεωπονικές Επιστήμες και Κτηνιατρική
- Άλλες Γεωπονικές Επιστήμες
- Natural Sciences
- Computer and Information Sciences
- Agricultural and Veterinary Sciences
- Other Agricultural Sciences
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/41807