Effects of Image Augmentation Techniques for Rice Leaf Disease Detection

Abstract

Rice leaf diseases substantially reduce crop yield, resulting in food shortages and financial losses. Early identification and control of these diseases can be aided and enhanced by automated computer vision-based detection systems. However, existing techniques suffer from low accuracy and inconsistency due to several issues. To improve model resilience against corrupted inputs and adversarial cases, this study examines the effects of image augmentation techniques on three transfer learning models for diagnosing rice leaf diseases. A consolidated dataset, which includes rice leaf images of five different classes, was used to train these models. The VGG-16 model trained on images augmented using the Random Flip technique achieves a maximum accuracy of 99.47%. However, we present the lightweight EfficientNet-B0 model, trained on MixUp augmented images, with an accuracy of 98.01%, as an alternative model that is more robust and suitable for deployment in mobile/web applications. Our results demonstrate that image augmentation techniques can enhance the model’s robustness against synthetically altered images without affecting its ability to detect and predict rice leaf diseases.