Rice Disease Image Classification using MobileNetV2 Pretrained Model with Attention Visualization using Gradient-weighted Class Activation Mapping (Grad-CAM)

Rice is one of the staple foods in Asia, particularly in Indonesia. With an ever-increasing demand of food, a stable rice production is mandatory and pest is one of the major challenges faced in sustainable rice production. In this research, we proposed an image classification model based on the MobileNetV2 pretrained model combined with a visual explanation based on the gradient-weighted class activation (Grad-CAM) algorithm to build a robust and accurate classification of rice diseases. The model is based on convolutional neural network (CNN) architecture. First, transfer learning is done from the MobileNetV2 pretrained model to create the classification model, followed by Grad-CAM to produce the visual explanation of the CNN. Finally, the model is trained on 7,077 rice images containing four different diseases (bacterial blight, blast, brown spot, and tungro) with a data augmentation process to increase the dataset’s overall variance. This process yields a model with a classification accuracy of up to 99,9%, combined with visual feature explanation making this model a robust and efficient classification model.