Batch size impact on enset leaf disease detection

Abstract

Enset, also known as the “false banana,” is a staple food in southern and southwestern Ethiopia that could potentially alleviate poverty among smallholders. Recently, a bacterial wilt disease that damages enset leaves has resulted in massive economic losses for farmers. The use of deep learning for automated plant leaf disease diagnosis in crops has grown in popularity in recent years; however, the impact of hyperparameter selection, particularly batch size, on model performance in the context of enset leaf disease detection remains unidentified. In this research, we looked at how batch size affects the effectiveness of a deep learning model to detect enset leaf disease. The study investigated how different batch size settings affected model performance during the detection of enset leaf disease. To confirm this, five commonly used batch sizes [16, 32, 64, 128, and 256] were combined in the proposed experiments. For the study, we have collected a total of 2132 infected and healthy leaves of enset from the south-west area of Ethiopia. Before training the convolutional neural network (CNN) model, the images in the dataset are preprocessed to enhance feature extraction and consistency. Based on the results of the experiments, we determined that the model’s efficiency was even better, but only when the batch size employed in the model was less than the size of the test dataset. The study uses deep learning to detect bacterial wilt in enset leaves and provides academics and practitioners with heuristic information to help boost enset production when CNN is used in agriculture