Leveraging deep learning algorithms for the timely detection of diseases in bean leaves

Abstract

In sub-Saharan Africa, where agriculture is a major commercial activity, the security of staple crops like beans is threatened by persistent diseases, notably bean rust and angular leaf spot. Uromyces appendiculatus, the fungus that causes bean rust, produces rust-colored pustules, whereas Pseudomonas syringae pv. phaseolicola, the cause of angular leaf spot, produces recognizable angular lesions. It is estimated that these diseases cost the agricultural sector millions of shillings each year in Uganda due to reduced bean yields, increased costs for disease control measures as well as the need to remove infected bean crops. A 2017 study found that Angular Leaf Spot caused a major yearly production loss of 384.2 tons especially in the Eastern region of the country where over 63% of the people participate in the activity, this raised major questions. In response to the demand for contemporary, data-driven approaches, this study presents a Deep Learning-based approach for the rapid and precise detection of angular leaf spot and bean rust by utilizing CNN algorithms with the free and open-source TensorFlow package and a public dataset of bean leaf images. This study has trained five models to detect bean rust and angular leaf spot in bean leaves. The prediction accuracies of the models were evaluated and the accuracies were 96%, 95%, 94%, 33% and 88% for Xception, ResNet50, DenseNet201, VGG19 and InceptionV3 respectively. Additionally, the performance of the models is evaluated using different metrics like F1-score, Precision and Recall. The Xception model with the highest prediction accuracy and Recall of 0.96 stood out as the top-performing model which was selected for further usage where the model was tested on images of two unhealthy classes and a healthy class. These algorithms demonstrate increased diagnostic accuracy and present a viable way to reduce the financial burden that agricultural diseases impose on Uganda and sub-Saharan Africa. Furthermore, the precise bean leaf disease identification system uses explainable AI frameworks such as LIME (Local Interpretable Model-Agnostic Explanations) to improve interpretability by visualizing the layer-wise feature extraction. These frameworks present an understanding of the attributes driving the categorization of diseases and provide details about the Deep Learning models’ choices hence promoting trust in the diagnostic results.