| dc.description.abstract | Diseases that affect coffee have serious repercussions for policymakers, seed industries, farmers, and consumers. Over the years, farmers have had to travel long distances searching for pathologists to check their coffee leaves and inform them which disease affects them. The pathologist approach is time consuming and costly. Automated technology in the agricultural sector has saved farmers’ time. The use of convolutional neural networks (CNN) to classify plant diseases has been implemented and showed reputable results. However, they still experience difficulty due to loss of information through pooling, difficulty in error propagation, computational complexities, and translation invariance, which leads to failure in pose preservation, and loss of shape and texture information. This work developed a hybrid neural network, which adopted three subnets for feature collection. To promote strong gradient flow and signal reception from all preceding layers, hybrid neural network model adopted DensNet loop connectivity where error signals were easily propagated to all the preceding layers in the subnets more directly. The Hybrid neural network model used channel-wise concatenation features from all preceding layers to collect rich feature patterns. This allows the model to combine complex and simple features, hence high generalization. This work used images from Mutira coffee plantation that were taken using a digital camera and with the help of a plant pathologist. The hybrid neural network model displayed a 99.7 % F1 score. In comparison, the conventional CapsNet model displayed 87 % F1 score accuracy on testing our framework based coffee dataset comprising 5 coffee disease categories with 58,000 images. The developed model showed relatively higher and stable accuracy when sensitivity analysis was performed by varying testing and training dataset percentages. Support Vector Machines (SVM), AlexNet, ResNet, VGGNet, Inception V3, Artificial Neural Networks (ANN), and VGG 16 deep learning approaches scored 84.5\%, 88.6\%, 99.3\%, 97.87\%, 99.14\%, and 98.2\%, respectively when the coffee dataset was used. The work used only five features with the target variable instead of the total ten features. This was done through the rigorous backward regression process. Due to the reduced number of features, model computation also reduced. This was also evident from the results of the test error. The overall test error for the developed hybrid neural network model was 0.16 while the test error for the original CapsNet model was 0.26. These findings indicate that hybrid neural network model may be a decent and, in most cases, superior and less expensive alternative for phrase categorization models founded on convolutional neural networks (CNNs). Therefore, several classifiers could be fused to enhance the accuracy of plant diseases classification. | en_US |
