Enhanced rice disease classification through pre-trained densenet model using transfer learning

Rice disease poses serious risks to rice, a staple food that is essential for more than half of the world's population and puts food security at risk. Rice disease detection has always required skill to visually inspect the farm and locate the lesion area, which is inefficient, prone to errors, and unsustainable for commercial farming. To overcome this challenge, this research work provides an enhanced model for identifying rice diseases using transfer learning methods. A total of 5,678 augmented rice leaf images were used to train and validate the model. The model focuses on classifying four classes, namely: Brown Spot, Leaf Blast, Hispa, and healthy leaves. The dataset was split into training (80%) and test (20%) sets to achieve optimum model evaluation. DenseNet201 model was utilised as the pre-trained model for determining the parameters of the proposed model (weights and biases), based on its improved feature propagation, decreased redundancy of parameters, and better accuracy on different plant disease datasets. The parameters of the pre-trained model are transferred to the new model for classification of rice diseases. The lower layers of the pre-trained model were maintained and frozen, while the upper layer was modified. This method makes it possible for the proposed model to adapt to the rice leaf dataset and effectively classify it. The proposed model incorporated depthwise separable convolutions at the modified upper layer of the pre-trained model to reduce training time while maintaining high accuracy. The proposed model achieved a higher accuracy of 96.0% compared to the existing models. This research work will provide farmers with solutions to quickly identify rice disease and take necessary actions on their farms.