Deep learning excels in plant disease diagnosis, yet its performance is often hampered by limited samples and class imbalance. To address these issues we introduce a comprehensive data generation pipeline. First, we employ traditional augmentation, CutMix, MixUp and synthetic image creation via Deep Convolutional Generative Adversarial Networks (DCGAN) to produce new samples. Then we apply the generated data in three scenarios: (S1) balancing the class distribution of the original set, (S2) increasing the number of samples per class, and (S3) expanding the dataset while keeping classes balanced. Using EfficientNetV2B0 with both pretrained and fine‑tuned weights, we evaluate how these distributions affect model accuracy. Results show a substantial boost in classification accuracy, highlighting the pivotal role of data diversity and class balance. The approach is adaptable to other crop disease detection tasks, offering promising tools for precision agriculture.
Review