Plant Disease Detection Through Deep Learning Using The Plant Village Dataset

Abstract

Challenges The diseases in the plants still remain a major threat that is faced by global agricultural systems, which imposes serious effect to plant productivity quality and food security, while early detection of these pathogens could be essential tools for sustainable agriculture. Conventional visual and laboratory detection methods are labor-intensive, time-taking, and less suitable for large-scale field applications; thus, it is easy to understand why artificial intelligence (AI) and deep learning research into agriculture have been in the ascendant. In this work, we investigate the performance of deep learning models for plant disease detection in hundred or thousand-level classes using the PlantVillage dataset that contains 26000+ labelled images of healthy and diseased leaves. We had experimented with three architectures I/i.e. Custom CNN, ResNet-50 and MobileNetv2, where the custom CNN is architecture which is constructed and trained on the dataset, ResNet-50 extracts features by fine-tuned transfer learned knowledge (pre-trained model) while MobileNetv2 as an architecture that was specifically designed to be easy to deploy on low power device. The performance was evaluated based on accuracy, precision recall and F1-Score for an overall consideration. From these results, we can see that our Custom CNN achieved the best result among benchmark models (99.3% accuracy and near perfect through all other rates) which indicates that the performance of model can learn subtle disease patterns while producing very confident predictions. It also performed well in the ResNet-50 test, achieving 96% accuracy as well as a favorable balance between precision and recall that would be suitable for generalizable deployments. Despite being 6.4% less accurate than MobileNetv2, it is computationally more efficient therefore can be appropriately chosen [selected] for use in mobility or edge-device based application, if one plans to do a real-time in field diagnosis. The findings of the study illustrate the importance of deep learning to precision agriculture for scalable accessible and robust disease monitoring, yet more challenges need further investigation including the long-tail distribution nature in terms of diseased versus non-diseased plants, poor generalization performance under real situations, and high demand on computational resources.