A Comparative Study of the Performance of CNN-Based Models for Rice Leaf Disease Classification

Abstract

      The objectives of this study were: 1) to develop an artificial neural network model for classifying rice leaf diseases, and 2) to compare the performance of the models trained with augmented data against those trained without argumentation in classifying rice leaf diseases. The RiceLeafBD Dataset, consisted of 1,555 images divided into 4 disease categories: bacterial leaf blight, fungal leaf blight, brown leaf spot, and tunggro virus, was used for this study. This dataset is publicly available on the Kaggle Dataset platform. Three neural networks, namely InceptionNet-V2, MobileNet-V2, and EfficientNet-V2, were experimented. These models were modified by adding Squeeze-and-Excitation (SE) blocks to better capture the importance of each channel. The experiments were conducted on two datasets: the baseline dataset and the augmented dataset. Model training and testing were performed for 20 and 50 epochs. The statistics used for data analysis were accuracy, precision, recall, and F1-score.
The results of model development and comparison of the efficiency of artificial neural networks in classifying rice leaf diseases with and without augmented data showed that the modified models achieved higher performance in classifying rice leaf diseases than the base models. Among the three modified models, EfficientNet-V2 achieved the highest accuracy. Without augmentation, EfficientNet-V2 revealed accuracy, precision, recall, and F1-score values of 92.60%, 92.75%, 92.50%, and 92.25%, respectively. With augmentation (EfficientNet-V2 + DA), the model achieved 95.50%, 95.75%, 95.75%, and 95.75% for accuracy, precision, recall, and F1-score, respectively. These results indicate that the developed models performed better than the base models in all cases.