Deep Learning for Disease Identification in Cassava Plants Using RGB Imaging

Agriculture is the most critical sector in the world economy. Food security and consistent food supply are the key challenges due to rapid growth in the world population and crop diseases. Crop yield gets significantly hampered due to diseases attacking large crop fields. The precise classification of plant diseases is essential to provide appropriate pesticide supplements. This study applies deep learning (DL) models to identify diseases in cassava plants using red–green–blue (RGB) images. Three models, namely Xception-net, MobileNetV2, and EfficientNetB0, are trained and evaluated on the cassava disease dataset publicly available on Kaggle. Since the dataset is imbalanced, five-fold cross-validation is used to check the consistency of the results. Transfer learning and random weight initialization were used for model training. The Xception-net is the clear winner among the three models, with an accuracy of 87.8% using transfer learning, 81.33% using random weight initialization, and 85% average accuracy using five-fold cross-validation. Further, this work is extended to develop web and mobile applications with DL models deployed at the backend to predict plant diseases from input images.