Classification of ripeness levels and prediction of physicochemical properties of durian fruit and pulp by near-infrared hyperspectral imaging

การจำแนกระดับความสุกและการทำนายสมบัติทางเคมีกายภาพของผลและเนื้อทุเรียนโดยการถ่ายภาพไฮเปอร์สเปกตรัลอินฟราเรดย่านใกล้

Name: Sneha Sharma

LCSH: Durian

LCSH: Spectroradiometer

LCSH: Infrared spectroscopy

Abstract: Durian is one of the popular and important tropical fruit in South-East Asia. Thailand is one of the main producers and exporters of durian. Pre and post-harvest handling measures are one of the challenging aspects being encountered by the durian growers in Thailand. Farmers mostly depend on visual inspection and tapping to identify the harvest maturity. Traditional approaches are effective, however, selecting the export quality of durian is challenging the traditional methods. Non-destructive sensors and technology are ready to be implemented for the real-time monitoring of the ripening stage and physicochemical properties of the durian. With the advancement in non-destructive technologies, hyperspectral imaging (HSI) has been developed as a powerful integration of imaging and spectroscopic techniques. Near-infrared hyperspectral imaging (NIR-HSI) has been considered a promising approach for the quality inspection of agricultural products. In this thesis, the possibility of using the pushbroom NIR-HSI system to classify the ripening stage and predict physicochemical properties has been established. The pushbroom NIR-HSI system (900-1600 nm), also known as line-scanning, generates the 3-D hypercube by scanning the sample on a transition stage at 10 mm s-1. This thesis has been divided into mainly four sections. Firstly, the preliminary analysis was done for the ripening stages (unripe, ripe, and overripe) classification of durian pulp by Bayesian optimized machine learning algorithms. In this section, one of the popular hyperparameter optimization techniques known as the Bayesian method was implemented on the NIR-HSI data to develop three machine learning classification models: support vector machine (SVM), random forest (RF), and k-nearest neighbour (kNN). Using Bayesian optimization, the highest classification accuracy and kappa coefficient obtained were 88.5% and 0.83, respectively, by the SVM classifier. The result showed that the optimization method influences the model accuracy of the classifiers. In the second section, the performance of the six machine learning classifiers for the ripening stage classification of intact fruit and pulp was performed. Six machine learning classifiers selected for the classification model development were linear discriminant analysis (LDA), partial least squares discriminant analysis (PLS-DA), artificial neural network (ANN), SVM, RF, and kNN. Grid search hyperparameter optimization was used to develop the model for classification. Accuracy comparison was made according to the classification metrices between the classifiers using full wavelengths and featured wavelengths selected by genetic algorithm (GA) and principal component analysis (PCA). The results showed that LDA classification models based on the full wavelengths and eleven feature wavelengths classified the durian intact fruit and pulp successfully with 100% accuracy. The performance of SVM, PLS-DA, and RF attained fair classification results; however, in comparison, the kNN classifier showed the lowest accuracy. In the third section, the regression models were developed to predict the physicochemical properties: dry matter (DM), total soluble solids (TSS), and fat content (FC); in durian pulp. Partial least squares regression (PLSR), SVM, RF, and l-dimensional convolution neural network (1-D CNN) models were developed: custom, U-Net, and VGG-19. Feature wavelengths were selected by successive projection algorithm (SPA) and genetic algorithm (GA). DM prediction model was the most accurate, showing coefficient of determination of test set (r2), root mean square error of prediction (RMSEP), and the ratio of prediction to deviation (RPD) in the range from 0.87-0.95, 1.42¬2.28%, and 2.7-4.4%. The outstanding performance for DM prediction was obtained from the GA-PLSR model. GA-SVM model attained the highest accuracy for TSS prediction. The overall model shows a fair result with the r2, RMSEP, and RPD of 0.87, 3.05%, and 2.8, respectively. For the FC prediction, the best accuracy was attained from the GA-PLSR model. GA-PLSR model for FC prediction exhibits the r2 and RMSEP of 0.81 and 0.65%, respectively. The 1-D CNN models were only feasible for DM prediction. The CNN models for TSS and FC did not provide convincing results for further application. This part of the thesis identified the best model for the DM, TSS, and FC which can be applied for the real-time monitoring of durian pulp. In the final section, the spatial distribution of DM, TSS, and FC was represented for the visualization of changes within and between the durian pulp harvested on different days after anthesis (DAA). The spatial mapping of durian pulp shows the changes in physicochemical properties that occurred during the ripening stages. Based on the findings of this thesis, it was confirmed that the NIR-HSI system could be implemented for ripening stage classification and the internal quality prediction of durian pulp. The pushbroom NIR-HSI system offers an advantage to be installed for the online scanning of the durian in large packaging firms. The few featured wavelengths selected from the PCA and SPA can be further used to develop portable multispectral or single-shot imagers. NIR-HIS system can provide powerful information from each pixel of the image. This information can be used to develop robust models when combined with advanced machine learning techniques along with feature selection. A real-time monitoring system can be developed using the information acquired in this thesis for durian grading and quality inspection.

King Mongkut's Institute of Technology Ladkrabang. Central Library

Address: BANGKOK

Email: Lifelong@kmitl.ac.th

Name: Panmanas Sirisomboon

Role: Thesis Advisor

Email : panmanas.si@kmitl.ac.th

Created: 2021

Modified: 2024-11-08

Issued: 2024-11-08

วิทยานิพนธ์/Thesis

application/pdf

eng

DegreeName: Doctor of Engineering

Level: Doctoral Degree

Descipline: Agricultural Engineering

Grantor: King Mongkut's Institute of Technology Ladkrabang

©copyrights King Mongkut's Institute of Technology Ladkrabang

ลำดับที่.	ชื่อแฟ้มข้อมูล	ขนาดแฟ้มข้อมูล	จำนวนเข้าถึง	วัน-เวลาเข้าถึงล่าสุด
1	Sneha Sharma.pdf	4.61 MB	3	2025-04-23 14:19:30

ใช้เวลา

0.017752 วินาที