Detection of adulteration and its classification based on the geographical area of coconut milk using a combination of NIR spectroscopy, machine learning and deep learning approach

Name: Agustami Sitorus

LCSH: Machine learning

LCSH: Deep learning (Machine learning)

LCSH: Food adulteration and inspection

Abstract: In this thesis, the research and development of chemometric tools based on machine learning (ML) and deep learning (DL) for near-infrared spectroscopy (NIRs) techniques to detect coconut milk adulteration and classify it based on its geographical area of origin were demonstrated. It is divided into 2 supporting section as a General Introduction (Chapter 1), and Conclusions, Recommendations, and Future Works (Chapter 8); 5 main sections (Chapters 2 to 6); and 1 section to challenge our model with a cross-over unknown dataset (Chapter 7). Chapter 1 briefly outlines the background of the samples, chemical content analysis of coconut milk, NIR instruments, and analysis tools used in this work. It also provides research problems, limitations of the research, and objectives of this thesis. Chapter 2 overviews the previous research paper on applying NIRs and IRs to detect and discriminate against the adulteration of food and agro-products based on recent research. Key findings from this study find that NIRs and IRs are non-destructive, rapid, simple-preparation, analytical rapidity, and straightforward methods for predicting adulteration in food and agro-products, so it is suitable for large-scale screening and on-site detection. Chapter 3 uses NIR and ML approaches to classify coconut milk types (fresh coconut milk, FCM; instant coconut milk, ICM; adulterated fresh coconut milk, A-FCM) and predicts distilled water (DW) adulteration in fresh coconut milk. A data sciences approach that combines appropriate preprocessing discovery and hyperparameter optimization concurrently is presented in this study. Partial least squares (PLS), linear discriminant analysis (LDA), support vector machine (SVM), and multilayer perceptron (MLP) with its hyperparameter were employed together with combining 18 preprocessing types and evaluated by 5-fold cross-validation (5f-CV) All regressors obtained the same satisfactory results to distinguish FCM, ICM, and A-FCM. Reressor from SVM obtained acceptable results, with Rc2 and Rp2 over 0.93, RMSEc and RMSEp below 8.30%, and RPD over 3.80. In Chapter 4, a novel approach to automatically select preprocessing (single up to multiple) and tuning hyperparameters simultaneously of ML algorithms based on their best performance in 5f-CV for FT-NIR and Micro-NIR spectroscopy data of coconut milk adulteration by DW and coconut water (CW) in the range 0 to 50%. This uses as many as 9 single preprocessing types and 3 types of ML classifier (LDA, KNN, MLP) and regressor (PLS, KNN, MLP). The performance strategy demonstrates that our proposed approach effectively addressed and produced satisfactory outcomes in classification and regression challenges and problems from coconut milk adulteration using NIRs. In Chapter 5, we explore DL algorithms that are only standardized using SNV preprocessing to identify the level of adulterated coconut milk using FT-NIR and Micro-NIR. Coconut milk adulteration samples came from intentional adulteration with corn flour and tapioca starch in the 1 to 50% range. Four types of DL algorithm architecture that were self-modified to a 1D framework were developed and tested, including CNN, S-AlexNET, ResNET, and GoogleNET. The results confirmed the feasibility of DL algorithms for predicting the degree of coconut milk adulteration by corn flour and tapioca starch with reliable performance (R2 of 0.886–0.999, RMSE of 0.370–6.108%, and Bias of −0.176–1.481). In Chapter 6, we explore the discrimination model using FT-NIR and Micro-NIR for geographical source areas of coconut milk in tandem with the classical (PCA, PLS-DA, LDA) to modern chemometrics classifier, including classifiers from ML (SVM, KNN, ANN) and DL (S-CNN, S-AlexNET, ResNET). Three sources as geographical areas of coconut milk originally from Thailand were used, including Chumphon, Samut Songkhram, and Chonburi Province. Our findings showed that an SVM and ResNET classifier could yield the optimal performance for discriminating the geographical source area of coconut milk, with an accuracy of 99.1% for the training and 100% for the testing using FT-NIR. Furthermore, when using Micro-NIR, the LDA, SVM, and KNN, the ResNET classifier delivered the highest accuracy of 99.5% for the training and 100% for the testing

King Mongkut's Institute of Technology Ladkrabang. KMITL Lifelong Learning Center

Address: BANGKOK

Email: lifelong@kmitl.ac.th

Name: Ravipat Lapcharoensuk

Role: Thesis Advisor

Created: 2024

Modified: 2025-10-02

Issued: 2025-10-02

วิทยานิพนธ์/Thesis

application/pdf

eng

DegreeName: Doctor of Engineering

Level: Doctoral Degree

Descipline: Food and Agricultural Intelligence Engineering

Grantor: King Mongkut's Institute of Technology Ladkrabang

©copyrights King Mongkut's Institute of Technology Ladkrabang

ลำดับที่.	ชื่อแฟ้มข้อมูล	ขนาดแฟ้มข้อมูล	จำนวนเข้าถึง	วัน-เวลาเข้าถึงล่าสุด
1	Agustami Sitorus.pdf	11.49 MB

ใช้เวลา

0.032135 วินาที