Evaluation of Data Augmentation Techniques for Rice Leaf Disease and Pest Classification Using EfficientNetV2B3

Abstract

Rice is a primary staple crop in Indonesia that is highly susceptible to diseases and pests such as blast, brown spot, and planthopper infestations. Early detection supported by advanced technology is essential to strengthen national food security. This study aims to enhance the classification accuracy of rice leaf diseases and pests using the EfficientNetV2B3 architecture through the application of various data augmentation methods. Three approaches were employed: non-augmentation, traditional augmentation techniques such as rotation, shift, zoom, shear, and flip, as well as StyleGAN2-ADA–based augmentation. The model was trained on Google Colab Pro+ using an NVIDIA Tesla T4 GPU for 50 epochs. The results indicate that the non-augmentation method achieved a validation accuracy of 78.80%, while traditional augmentation improved it to 88.80%. The StyleGAN2-ADA approach yielded the best performance, with a validation accuracy of 97.58%, test accuracy of 98%, and an F1-score of 98%, albeit requiring longer computational time. GAN-based augmentation proved effective in enhancing the model’s generalization capability and stability. The StyleGAN2-ADA approach was found to be optimal for improving the performance of EfficientNetV2B3 in classifying rice leaf diseases and pests.