BWENet: Detection and Grading of Bacterial Wilt Using Deep Convolutional Neural Network

Abstract

Objectives: To design a Bacterial Wilt Enset (BWENet) model that can detect and grade the level of bacterial wilt disease in Enset using a deep convolutional neural network. Methods: Convolutional neural network is used for detection of bacterial wilt and 4-way Softmax is used for grading bacterial wilt into a specific level (normal, early-stage, infected stage, and completely wilted Enset). About 1600 images were used to evaluate the performance of the model out of which 70% is for training, 15% for validation and 15% for testing. We collected dataset from kefa, Sheka agricultural research center and some selected areas in SNNPR. The proposed model was trained using these images and data augmentation techniques were applied to generate more images. Findings: BWENet model achieved an identification test grading accuracy of 90.53% bacterial wilt disease. BWENet model is found to be faster to train and has a smaller model size as compared to State-of-the-art models like Alex Net, Google net, and VGG19. Novelty: we are proposing a new CNN architecture to grade bacterial wilt disease severity measurement. By our experiments, we have shown the kernel size selection, poling layer selection, and segmentation achieved better accuracy results.