University of Peloponesse
Application of machine learning methods for the prediction of plant diseases and quality of agricultural products
Abstract
dc:descriptionAgriculture plays a critical role in global economic development by meeting humanity’s need for food and sustaining employment worldwide. At the same time, it faces growing challenges from climate change, resource scarcity, and increasing demand. Recent advances in machine learning enable the analysis of big agricultural data and support earlier, data-driven interventions compared to traditional monitoring approaches. Machine learning applications in agriculture fall into four broad fields: (i) crop yield prediction, (ii) disease detection and prediction, (iii) weed/crop identification, and (iv) food quality and safety assessment. This thesis explores how machine learning methods can address real-world problems in agriculture, where limited data size, heterogeneous variables, and high dimensionality impose practical modeling constraints and require a balance between predictive performance and interpretability. The methodological framework examines supervised and unsupervised learning across regression, classification, clustering, and dimensionality-reduction tasks, demonstrating that model selection should be guided by data characteristics, problem complexity, and the need for interpretable results that support decision-making, rather than by the assumption of a universally optimal method. The original research contributions of this thesis focus primarily on food quality and safety and plant disease prediction. In the food safety domain, a modeling framework was developed to predict aflatoxin contamination risk in dried figs using environmental, soil, and microbiological variables, demonstrating that non-imaging data can provide reliable early indicators of contamination. Predictive performance was enhanced through dimensionality reduction and a transformation of the feature space that increased discriminative separation among samples. In the food quality domain, original studies on olive oil revealed interpretable relationships between chemical composition and sensory attributes, enabling prediction of key sensory characteristics and supporting sensory panel training. A complementary study further showed that sensory attributes alone can capture meaningful varietal patterns, revealing groups of olive oil varieties with similar sensory profiles. Based on these findings, a machine learning model was developed to represent each group through prototypes, enabling accurate classification of olive oil samples into varietal groups. This approach provides valuable support for sensory panel training. Within the same domain, the thesis introduces an interactive web application, which integrates multivariate analysis and clustering to classify Koroneiki extra virgin olive oils into quality subgroups within the extra virgin category. The main contribution lies in (i) providing a data-driven subcategorization of the broad extra virgin olive oil class, which encompasses a wide spectrum of quality levels, (ii) establishing this subcategorization using only the sensory attributes, and (iii) delivering a practical and user-friendly tool for classifying new samples into these quality groups. The thesis also introduces a producer-specific software tool for monitoring the quality of olive oil from a single producer using the three sensory attributes of olive oil samples produced over time by a single producer. Clustering identifies natural groupings based on sensory similarity, enabling detection of improvements, consistency patterns, or weaknesses in production practices, and supporting data-driven quality improvement. In the domain of disease prediction, the thesis examines whether soil, foliar, and orchard variables can serve as early indicators of olive anthracnose. The study represents one of the first attempts to predict the disease using real field-based soil and foliar data. The analysis addressed the challenge of a small, high-dimensional dataset and included a comparative evaluation of multiple models to identify the most suitable approach for the task. This process enabled the identification of the most influential parameters associated with disease occurrence. The resulting model provided interpretable insights for agronomists, supporting understanding of key disease drivers and enabling early intervention for effective disease management. Overall, this thesis demonstrates the potential of machine learning to support disease prediction and food quality and safety assessment. Beyond methodological contributions, the work highlights the importance of interpretability, domain knowledge, and practical applicability in real-world agricultural problems. The developed models and software tools provide not only predictive capability but also meaningful insights that support stakeholders in understanding and improving their practices.
Degree
thesis:*- Grantor dc:publisher
- University of Peloponesse
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Kottaridi, Klimentia
- Κοτταρίδη, Κλημεντία
Subjects
dc:subject × 26- Τεχνητή νοημοσύνη
- Μηχανική μάθηση
- Φυτοπαθολογία
- Ασφάλεια τροφίμων
- Αλγόριθμοι μηχανικής μάθησης
- Προγνωστικά μοντέλα
- Γεωργία ακρίβειας
- Ποιότητα τροφίμων
- Artificial intelligence
- Machine learning
- Plant diseases
- Food safety
- Food quality
- Machine learning algorithms
- Predictive models
- Precision agriculture
- Άλλες Γεωπονικές Επιστήμες
- Γεωπονικές Επιστήμες και Κτηνιατρική
- Άλλες Επιστήμες Μηχανικού και Τεχνολογίες
- Επιστήμες Μηχανικού και Τεχνολογία
- Μηχανική, διεπιστημονική προσέγγιση
- Other Agricultural Sciences
- Agricultural and Veterinary Sciences
- Other Engineering and Technologies
- Engineering and Technology
- Engineering, interdisciplinary
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62137