Improving Imbalanced Korean Melon Pest Data and Classification Performance through Generative AI-based Image Augmentation

Due to climate change, plant pests and diseases are becoming increasingly severe, posing significant challenges to agricultural productivity and crop quality. Additionally, although smart farming technologies attempt to address these issues, the difficulty in collecting balanced and sufficient data remains a major obstacle. This study proposes the use of generative AI models to create images as a solution to the data imbalance problem in plant pest datasets. The proposed method utilizes Generative Adversarial Networks (GAN) and Generative Pre-trained Transformer 4 (GPT-4) to generate highquality synthetic images of plant pests. The augmented dataset is then used to train and evaluate stateof-the-art classification models, including Convolutional Neural Networks (CNN) and Transformerbased architectures. Experiments were conducted using a dataset of Korean melon (Cucumis melo L.) leaves, which are particularly susceptible to various pests and diseases. The results demonstrate that the proposed augmentation techniques significantly improve the performance of the classification models, with a specific method achieving the highest F1-score and accuracy. The generated images are also shown to be visually similar to real pest images, indicating the effectiveness of the augmentation approach.