Cassava Plant Disease Detection Using Transfer Learning with Convolutional Neural Networks

Abstract

Cassava vegetables represent one of the main food sources in some regions like Africa. It can be affected by plant diseases which can increase food insecurity in areas where the plant is relied upon. An early detection, identification and classification of cassava plant disease may be a beneficial tool. Manual field examination can be expensive in many ways, monetary, working time, physically exhausting, etc. Automated detection and identification of plant diseases can be done by analysis of field/plant images. In that case, machine learning can be used to classify the states of these plants. This paper is concerned with building a convolutional neural network (CNN) using transfer learning to create a model for an unbalanced dataset that is more accurate and efficient than existing CNN models. Tested and compared pre-trained models include ResNet101V2, ResNet50V2, EfficientNetB2, VGG16, VGG19, and MobileNet2. After initial testing and results, based on test accuracy, training accuracy, and output trends, further improvements were made by using VGG19. This model was further fine-tuned for the cassava plant disease classification. The resulting model achieved a training accuracy of 99.31 % and a test accuracy of 80.27% with the use of 50 epochs which outperforms current results from literature.