Citation
Ram Kumar Thirumugam, . and Yasmin Mohd Yacob, . and Wan Mahfuzah Wan Ibrahim, . and Salina Mohd Asi, . and Suhizaz Sudin, . (2025) Machine learning precision for mangosteen maturity: a comparative analysis of conventional classifiers. Pertanika Journal of Science & Technology (Malaysia), 33 (1). pp. 283-312. ISSN 2231-8526
Abstract
Identifying mangosteen maturity stages pre-harvest is crucial for postharvest quality, as fruit disease and pest infestation often occur at specific stages. Deep learning, while popular for classification, struggles with false negatives. Conversely, conventional machine learning methods now effectively handle false negative issues. The main goal of this research is to determine the significant comparison between different conventional classifiers, namely Random Forest (RF), Decision Tree (DT), Support Vector Machine (SVM) and K-Nearest Neighbour (K-NN), in terms of their accuracy, validity, and False Negative Rate (FNR) in predicting six distinct classes. Image samples of 253 mangosteens across six maturity stages were used, with 20 regions of interest (ROIs) each. 112 Gray-level Co-Occurrence Matrix (GLCM) and colour features were extracted to train models using texture, colour, and combined features. The evaluation metrics used for assessing the validity of predictions included precision, recall, F1-score, accuracy, and Cohen’s Kappa. The RF classifier achieved high validation scores, with an accuracy of 0.76 and Cohen’s Kappa of 0.70 for combined features, 0.75 and 0.69 for coloured features, and 0.46 and 0.33 for texture features. The Friedman test on false positive rates (FNR) across the four models shows significant differences (p < 0.05) for colour, texture, and their combination, with p-values of 0.00134, 0.00572, and 0.00071, respectively. RF is the best method, with the lowest mean FNR scores: 1.16 for texture, 1.16 for colour, and 1.00 for combined features. In conclusion, the RF classifier outperforms other classifiers in accuracy, validity, and mean FNR across six classes with three category features, achieving statistical significance in the Friedman Test.
Download File
Official URL: http://www.pertanika.upm.edu.my/resources/files/Pe...
|
![[img]](http://webagris.upm.edu.my/style/images/fileicons/text.png)
Abstract
Identifying mangosteen maturity stages pre-harvest is crucial for postharvest quality, as fruit disease and pest infestation often occur at specific stages. Deep learning, while popular for classification, struggles with false negatives. Conversely, conventional machine learning methods now effectively handle false negative issues. The main goal of this research is to determine the significant comparison between different conventional classifiers, namely Random Forest (RF), Decision Tree (DT), Support Vector Machine (SVM) and K-Nearest Neighbour (K-NN), in terms of their accuracy, validity, and False Negative Rate (FNR) in predicting six distinct classes. Image samples of 253 mangosteens across six maturity stages were used, with 20 regions of interest (ROIs) each. 112 Gray-level Co-Occurrence Matrix (GLCM) and colour features were extracted to train models using texture, colour, and combined features. The evaluation metrics used for assessing the validity of predictions included precision, recall, F1-score, accuracy, and Cohen’s Kappa. The RF classifier achieved high validation scores, with an accuracy of 0.76 and Cohen’s Kappa of 0.70 for combined features, 0.75 and 0.69 for coloured features, and 0.46 and 0.33 for texture features. The Friedman test on false positive rates (FNR) across the four models shows significant differences (p < 0.05) for colour, texture, and their combination, with p-values of 0.00134, 0.00572, and 0.00071, respectively. RF is the best method, with the lowest mean FNR scores: 1.16 for texture, 1.16 for colour, and 1.00 for combined features. In conclusion, the RF classifier outperforms other classifiers in accuracy, validity, and mean FNR across six classes with three category features, achieving statistical significance in the Friedman Test.
Additional Metadata
| Item Type: | Article |
|---|---|
| AGROVOC Term: | mangosteens |
| AGROVOC Term: | classification |
| AGROVOC Term: | machine learning |
| AGROVOC Term: | image analysis |
| AGROVOC Term: | optical properties |
| AGROVOC Term: | preharvest operation |
| AGROVOC Term: | accuracy |
| AGROVOC Term: | forecasting |
| Geographical Term: | Malaysia |
| Uncontrolled Keywords: | Conventional machine learners, decision tree, fruit maturity, K-Nearest Neighbour, random forest, support vector machine |
| Depositing User: | Ms. Azariah Hashim |
| Date Deposited: | 17 Aug 2026 03:56 |
| Last Modified: | 17 Aug 2026 03:56 |
| URI: | http://webagris.upm.edu.my/id/eprint/4354 |
Actions (login required)
 |
View Item |
