Automated coffee leaf disease classification via deep feature extraction with PhytoV2Net and InceptionV3 architectures

Abstract

Proper categorization of diseases in coffee plants is critical for their early detection and good crop management, which in turn has a direct impact on the quantity, quality, and long-term sustainability of agriculture. Early detection of diseases makes it possible to resort to selective interventions, which not only reduce the chance of total crop loss, but also help in the control of outbreaks. Until now, detection methods have relied on manual inspections that are slow, inconsistent and prone to human errors, making them very much dependent on experts. The present research uses deep learning along with machine feature extraction techniques for coffee disease identification using PhytoV2Net and InceptionV3, respectively. Both networks were trained on the JMuBEN and JMuBEN2 datasets, which consist of a total of 58, 549 leaf images. These networks can differentiate between healthy and diseased leaves by learning to identify visual symptoms such as spots, discolorations, and lesions. The custom PhytoV2Net model produced an accuracy score of 99.87%, while InceptionV3 was based on 99.55% under K-fold cross-validation. PhytoV2Net performance was also above 95% in both precision and recall, indicating the high dependability and constancy of the model in its disease identification. The environmental changes posed difficulties, such as variations in lighting, leaf blocking, and background noise. Preprocessing techniques, particularly data augmentation applied to the JMuBEN dataset from an open-source data repository, helped improve image quality and enhance model robustness. Deploying these models in real time can significantly advance smart farming. When integrated into edge devices, handheld tools, or drone systems, they enable autonomous on-site detection of coffee leaf diseases and turn these platforms into intelligent assistants for plant health monitoring.