Design of Enhanced CNN Model for Rice Disease Classification with Comparative Analysis on Different Variants of Dataset

Convolutional Neural Network (CNN) is extensively employed in crop disease classification and has been achieving remarkable results. But the challenge researchers face while designing the CNN model is misclassification of diseases leading to low accuracy and huge training time. Overfitting is the common problem which is faced by CNN models. Also, the dataset used is the key factor for training the CNN model. This study proposes the CNN architecture for rice disease identification with enhanced accuracy and less training time. Three convolutional layers and two dense layer are added along with application of max pooling, batch normalization and dropout layers. Hyperparameter fine tuning is applied for training the proposed architecture to obtain the optimized parameters to improve the model accuracy. The proposed CNN model is evaluated by doing comparative analysis on three variants of rice disease dataset which includes RGB images, grey scale images and augmented RGB images. It has been found in the study that proposed CNN architecture performs classification of rice diseases with high accuracy and without overfitting on all three variants of dataset dataset. It attains best training accuracy of 96.75% and testing accuracy of 95.25% with augmented RGB images with training time of 127 seconds only.