Detection of Leaf Black Sigatoka Disease in Enset Using Convolutional Neural Network

Enset is mostly grown in central and southwest Africa and is used for a variety of things, including food, medicine, shelter, and even animals. It is also called as false banana (Enset ventricosum). Viruses, bacteria, and fungus are frequently capable of causing harm to Enset. Black spot disease or leaf sigatoka disease, which affects Enset leaves by fungi, reduces agricultural productivity and quality. Without a suitable system, sigatoka disease detection takes longer, requires more work, and is more expensive. The major goal of this project is to create an automated computer vision system that can recognize Enset leaf black sigatoka disease and suggest a course of therapy using a convolutional neural network (CNN). There are four phases in the suggested system. Dataset gathering is the initial stage, during which photographs of both healthy and diseased Enset leaves are collected from agriculture in the East African zone. Two useful characteristics are taken from the image during this phase, after which the system is trained and the model is created using the training datasets. The third phase focuses on classifying photos according to whether they are healthy or unhealthy, utilizing attributes that were extracted during training. The classification accuracy of the proposed model is tested by using testing datasets and the experiment was conducted on two CNN models. The first is the proposed model and the second is the pre-trained model (Resnet50). Both models are trained using the same image size (112 × 112) with RGB color channels. The data augmentation technique is used to increase the number of datasets during the training phase. The pre-trained model is also trained by a dataset called Resnet50. This system achieves 96.4% of overall accuracy. Finally, the model's performance is evaluated using a confusion matrix (precision, recall, and accuracy). Python is used for implementing the model. Generally, the system achieves 97.6%, 95.2%, and 96.4% precision, recall, and accuracy, respectively.