DATA AUGMENTATION EFFECTS ON PROTONET FEW-SHOT YELLOW DISEASE SEVERITY IN CHILI LEAVES

Yellow curling disease in chili plants is one of the leading causes of declining horticultural productivity because it reduces the quality and quantity of crops. Variations in symptoms at each level of severity make the identification process difficult, especially when labeled data is minimal. This study proposes a Prototypical Network-based Few-Shot Learning (FSL) approach with VGG16 architecture as a feature extractor. Five augmentation techniques, namely horizontal flip, rotation, zoom, brightness, and contrast adjustment, were used to increase data diversity in data-scarce conditions. Experiments were conducted with N-way K-shot configurations (2–5 classes; 1, 5, and 10 examples per class) to evaluate the impact of augmentation on prototype representation stability. Results show that increasing the number of examples per class consistently improves accuracy from 34.6% in 5-way 1-shot to 49.4% in 5-way 10-shot without augmentation. However, the use of augmentation decreases performance in higher N-way scenarios because it increases intra-class variability. The t-SNE visualization reinforces this study, where the healthy and severely diseased classes are clearly separated, while the intermediate class shows overlap. The novelty of this study is that it is the first to evaluate the impact of augmentation strategies on prototype representation stability in the agricultural domain with limited data. The results of this Few-Shot Learning approach are effective for plant disease classification despite the limited dataset.