Effect of Synthetic Data Augmentation on Plant Classification Accuracy Using MobileNetV2, EfficientNet-B0, and ResNet-18

The classification of agricultural crops such as cassava, corn, rice, and sugarcane is crucial for efficient data collection and decision-making. However, the limited availability of labeled images poses a significant challenge in training deep learning models effectively. This study investigates the effect of synthetic data augmentation using Deep Convolutional Generative Adversarial Networks (DCGAN) and Wasserstein GAN with Gradient Penalty (WGAN-GP) on plant classification accuracy. The original dataset contained only 245 images, which were augmented using GANs to increase data diversity. Three convolutional neural networks—MobileNetV2, EfficientNet-B0, and ResNet-18—were trained on three dataset variations: the original dataset, a combination of original and synthetic images, and a fully synthetic dataset. Experimental results reveal that the use of GAN-generated data, particularly from DCGAN, negatively impacted classification accuracy due to artifacts and inconsistencies in the synthetic images. Despite achieving lower Frechet Inception Distance (FID) scores, DCGAN-generated images introduced distortions that reduced model performance. WGAN-GP produced more visually realistic images but still failed to improve classification accuracy. The highest accuracy of 97.37% was achieved using only the original dataset, while models trained with GAN-augmented datasets experienced performance degradation. These findings emphasize that when the amount of real data is limited, low-quality synthetic images can further degrade model performance rather than improving it. Future research should focus on enhancing the quality of GAN-generated images and exploring hybrid augmentation techniques for small datasets.