Improving plant disease classification using realistic data augmentation

Recently, several studies have used deep convolutional neural networks (DCNN) for plant disease classification based on leaf symptoms images. Most of these studies use the PlantVillage dataset, which contains simple images with simple backgrounds. This has led to classifiers achieving high accuracy in academic settings but underperforming in practical scenarios where images taken by farmers are much more complex. In this paper, we propose a data augmentation (DA) method that transforms simple images from the PlantVillage dataset into more complex images, thereby bridging the gap between academic results and practical performance. Our technique focuses on making the images more realistic by generating images containing multiple leaves and complex backgrounds, such as soil. To evaluate the impact of our proposed method, we trained DCNN models using various augmentation methods and tested them on a dataset containing highly complex images comparable to those taken by farmers, posing a substantial challenge. The experimental results showed that models trained on images generated by our method outperformed traditional data augmentation methods, such as geometric transformations. Moreover, our method is competitive with Generative Adversarial Networks (GANs) without requiring any training phase.