Improving Potato Disease Image Classification Through Advanced Deep Learning Models

Abstract

Agriculture is an important part of human society, serving as a crucial source of food and a significant contributor to the global economy. However, it faces numerous challenges, one of which is the devastating impact of plant diseases. Among the various crops affected by plant diseases, potatoes, which are a staple food for millions of people around the world, are particularly at risk. Potatoes are susceptible to a wide range of diseases, which can lead to significant crop loss and threaten sustainable food production.

Traditional methods of diagnosing and managing plant diseases rely heavily on human intervention. While these methods are effective to some extent, they can be error-prone and inefficient, especially for large-scale operations. The need for more efficient and automated solutions for disease diagnosis and management is evident, particularly in regions such as Uganda, where agriculture is not just an economic activity but a way of life. The motivation for this research stemmed from the imperative to tackle these challenges and enhance the classification and management of potato diseases. Recognizing that these challenges intersect with the field of computer vision, an area that has seen significant advancements and has the potential to change the way diseases are diagnosed and managed, this study employed advanced deep-learning models to address these issues. Vision transformers (ViT) and convolutional neural networks (CNNs) such as EfficientNetV2B3, MobileNetV3, VGG16, ResNet50, and DenseNet121, renowned for their efficacy in image classification tasks have been employed. To further improve performance, a hybrid model framework that combines the strengths of both CNNs and the ViT has been introduced. These models were trained and evaluated on the “Potato Leaf Disease Dataset,” a diverse dataset reflecting real-world agricultural conditions, marking a significant improvement over previously used datasets.

The results reveal that the hybrid model, specifically EfficientNetV2B3+ViT, significantly outperforms the methods in the previous study in all evaluation metrics. It achieved an accuracy of 85.06%, representing an 11.41% improvement over the results of the previous study. This suggests that hybrid models can further improve potato disease classification by improving their ability to generalize to unseen data.

Furthermore, this study identified several areas for future research. These include exploring other hybrid architectures and implementing Xplainable AI to enhance the models’ interpretability and provide more insights into their decision-making processes.

This study has the potential to make a significant contribution to the field of disease classification in agriculture. The findings provide a solid foundation for future research, with the ultimate goal of developing more accurate, reliable, and interpretable disease classification systems. This research’s potential impact extends beyond the immediate benefits of improved disease classification. By improving the classification of potato disease, the research could support more effective and timely interventions by reducing crop losses and bolstering agricultural productivity. This in turn could have broader economic benefits by supporting the livelihoods of farmers and contributing to sustainable food production.