Cassava plant disease detection with imbalanced dataset using transfer learning

Plant disease has jeopardized the agriculture industry and is the biggest threat that influences global food security. Therefore, the fundamental guiding the control of disease proliferation is the effective diagnosis of diseases induced in plants at an early stage. This work put forward a convolutional neural network using transfer learning to perform cassava disease detection. The proposed approach has undergone substantial training and testing on cassava leaf images containing five distinct classes. After a pre-processing step using a contrast enhancement technique, oversampling techniques combined with data augmentation methods are used on the dataset to counter the high-class imbalance. Because not all data in the actual world is balanced, efficient categorization of unbalanced data is an important area of study. Experimental results demonstrate that the balanced dataset has increased the model accuracy by 4.3%. The proposed methodology achieved an accuracy score of 94.02% when data augmentation techniques were coupled with oversampling techniques.