Scholarlink

<p>Global food sovereignty is increasingly threatened by the prevalence of potato diseases, causing significant damage to their yield and quality. Classical approaches to disease detection methods remain highly laborious and error-prone, relying on expert diagnosis. To mitigate these challenges, the present study introduces an ensemble modelling approach integrated with explainable artificial intelligence (XAI), developed to overcome field-level challenges of potato leaf disease identification and promote regenerative agricultural practices. A real-field dataset was collected from the Kapurthala district of Punjab, India, containing 2017 original images (augmented to 7248 samples) of two destructive potato leaf diseases, early blight and late blight, along with healthy leaf samples. The dataset was pre-processed through scaling and transformations to enhance model generalization. In the proposed system, complementary strengths of various neural network architectures, such as CNN, ResNet50, VGG16, MobileNetV2, EfficientNetB3, and InceptionV3, are integrated to solve the multiclass classification problem. The suggested model yielded promising results, obtaining an overall accuracy of 99.72% on an independent test set, reflecting high predictive reliability across all disease classes, and surpassing all standalone models, whose accuracies ranged from 94.36 to 99.48%. Additionally, the incorporation of XAI techniques offered interpretable visualization maps that gave insight into the model’s inferencing mechanism and improved transparency for plant pathologists and researchers. These results demonstrate the effectiveness of a scalable potato disease management approach where transfer learning, ensemble modelling, and interpretability methods substantially boost the classification performance. Furthermore, this ensemble approach indicates strong potential for deployment in precision farming systems, with feasible computational requirements and suitability for real-time, site-specific crop management applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Application of Explainable AI-Driven Ensemble Modelling Approach for Robust Potato Disease Identification under Field Conditions

  • 期刊论文

  • 发表日期：

    2026年5月28日

查看全文
  • Kajal Verma,
  • Mithilesh Kumar Dubey,
  • Devendra Kumar Pandey

摘要

Global food sovereignty is increasingly threatened by the prevalence of potato diseases, causing significant damage to their yield and quality. Classical approaches to disease detection methods remain highly laborious and error-prone, relying on expert diagnosis. To mitigate these challenges, the present study introduces an ensemble modelling approach integrated with explainable artificial intelligence (XAI), developed to overcome field-level challenges of potato leaf disease identification and promote regenerative agricultural practices. A real-field dataset was collected from the Kapurthala district of Punjab, India, containing 2017 original images (augmented to 7248 samples) of two destructive potato leaf diseases, early blight and late blight, along with healthy leaf samples. The dataset was pre-processed through scaling and transformations to enhance model generalization. In the proposed system, complementary strengths of various neural network architectures, such as CNN, ResNet50, VGG16, MobileNetV2, EfficientNetB3, and InceptionV3, are integrated to solve the multiclass classification problem. The suggested model yielded promising results, obtaining an overall accuracy of 99.72% on an independent test set, reflecting high predictive reliability across all disease classes, and surpassing all standalone models, whose accuracies ranged from 94.36 to 99.48%. Additionally, the incorporation of XAI techniques offered interpretable visualization maps that gave insight into the model’s inferencing mechanism and improved transparency for plant pathologists and researchers. These results demonstrate the effectiveness of a scalable potato disease management approach where transfer learning, ensemble modelling, and interpretability methods substantially boost the classification performance. Furthermore, this ensemble approach indicates strong potential for deployment in precision farming systems, with feasible computational requirements and suitability for real-time, site-specific crop management applications.