Self-Supervised Learning for Class-Imbalanced Guava Disease Detection: A Contrastive Reconstruction Enhancement Approach

Rikkula, Vishnu Vardhan Reddy (2025) Self-Supervised Learning for Class-Imbalanced Guava Disease Detection: A Contrastive Reconstruction Enhancement Approach. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (1MB) | Preview

Abstract

Supervised learning based systems of plant disease detection are highly accurate in general, but fail utterly on off-the-record disease classes. This renders them inefficient in agriculture where isolated diseases tend to go viral. This research paper discusses the effectiveness of self-supervised learning in alleviating serious class imbalance in guava disease detection with specific attention to the pre-training of Contrastive Reconstruction Enhancement (CRE) in minority class detection in comparison to full supervision methods. The experimental design applied in the study is the three-phase one, that is, setting supervised baselines on the Bangladesh guava dataset (4,360 images, 6 classes), running the CRE pre-training procedure on the PlantVillage dataset (61,486 images, 39 disease classes), and conducting comparative analysis through fine-tuning. With the outcomes of the experiment, it is evident that CRE pretraining significantly enhances the identification of minority classes. As an example, Healthy leaves increased in the recall percentage, 67% to 81.82% and Insects eaten increased in the recall percentage, 50% to 76.92%. The macro F1 score increased to 0.9406 as compared to 0.8930 and this indicates that there was an enhancement of performance in all the classes by 5.33%. Although the total accuracy in this case decreased a little bit, setting 98.77 % at 97.71%, this trade-off is favorable to practical application when it is of great interest to find a rare disease. The paper contains some empirically proven evidence that class imbalance in the agricultural AI could be addressed through self-supervised learning, which would be a replicated methodology applicable in the situation with other special crops faced with similar data constraints. These findings indicate that the use of SSL is an effective method of developing powerful disease detection systems that may be employed in agricultural contexts where available resources are scarce.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Singh, Jaswinder UNSPECIFIED
Subjects:	Q Science > QA Mathematics > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science S Agriculture > SB Plant culture H Social Sciences > HD Industries. Land use. Labor > Specific Industries > Agriculture Industry Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	08 Sep 2026 12:01
Last Modified:	08 Sep 2026 12:01
URI:	https://norma.ncirl.ie/id/eprint/9900

Actions (login required)

View Item