Identification of Potato Plant Leaf Diseases Using Image Segmentation and Deep Learning Techniques

Abstract

This study presents a novel approach to the automated identification of potato plant leaf diseases using advanced image segmentation techniques coupled with deep learning algorithms. The research addresses the critical need for efficient and accurate disease detection in potato crops, which are vital for global food security. We developed a comprehensive dataset of 5,000 high-quality images representing healthy leaves and four common potato leaf diseases: Early Blight, Late Blight, Septoria Leaf Spot, and Bacterial Leaf Spot.

Our methodology employed three distinct image segmentation techniques: Otsu's thresholding, K-means clustering, and the Watershed algorithm. These were comparatively analysed to determine their effectiveness in isolating diseased areas on leaf surfaces. The segmented images were then used to train and evaluate several deep learning models, including a custom-designed Convolutional Neural Network (CNN) and a transfer learning approach utilizing a pre- trained ResNet50 architecture.

The results demonstrated the superiority of K-means clustering for image segmentation, achieving a 92.1% accuracy in disease detection when used as a preprocessing step. The transfer learning approach with ResNet50 emerged as the most effective classification model, attaining an impressive 96.3% accuracy across all disease categories. This model also exhibited high precision (0.964) and recall (0.963), indicating its robust performance in both disease detection and healthy leaf identification.

Our approach outperformed existing state-of-the-art methods in potato leaf disease detection, albeit by a narrow margin. The study's findings highlight the potential of integrating advanced image processing techniques with deep learning for creating highly accurate, automated disease detection systems in agriculture. Such systems could significantly enhance crop management practices, potentially leading to improved yields and food security.

This research contributes to the growing field of smart agriculture by demonstrating the feasibility of AI-driven disease detection in potato crops. Future work should focus on expanding the model's applicability to more diverse field conditions and exploring ensemble methods to further improve accuracy.

Index Terms: