Scholarlink

<p>In order to ensure sustainable agricultural productivity, a reliable diagnostic framework to identify mango leaf diseases through interpretable visual symptoms is necessary. Deep learning models have high classification accuracy, but the “black box” nature of the deep learning models often makes it difficult to understand the underlying rationale of a prediction. To overcome this limitation, an explainable artificial intelligence (XAI) agent is computationally developed based on a two-stage diagnostic strategy. The proposed framework first employs hybrid vision transformer architecture for leaf-level classification, and employs local interpretability methods to determine the specific image patches that influence the decision. In the second stage, a feature detection model scans the identified regions to link the classification to visible pathological indicators such as necrotic regions, holes, and discoloration. By bridging the gap between global predictions and local geometric causes, this dual-model approach mimics the selective attention of a human specialist. The result analysis demonstrates that this strategy effectively transforms opaque diagnostic processes into a transparent and human-understandable format, thereby enhancing the reliability of automated systems for early crop management in hazardous or large-scale agricultural environments. The empirical results reveal that the suggested framework is capable of accurately classifying leaves while simultaneously localizing symptoms in an interpretable way, which can further be applied to diagnose diseases in different plants.</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

A hybrid ResNet-ViT and YOLOv3-ViT pipeline for interpretable mango leaf disease diagnosis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月8日

查看全文
  • Ravi Anand,
  • Ritesh Kumar Mishra,
  • Rijwan Khan,
  • Vipin Kumar

摘要

In order to ensure sustainable agricultural productivity, a reliable diagnostic framework to identify mango leaf diseases through interpretable visual symptoms is necessary. Deep learning models have high classification accuracy, but the “black box” nature of the deep learning models often makes it difficult to understand the underlying rationale of a prediction. To overcome this limitation, an explainable artificial intelligence (XAI) agent is computationally developed based on a two-stage diagnostic strategy. The proposed framework first employs hybrid vision transformer architecture for leaf-level classification, and employs local interpretability methods to determine the specific image patches that influence the decision. In the second stage, a feature detection model scans the identified regions to link the classification to visible pathological indicators such as necrotic regions, holes, and discoloration. By bridging the gap between global predictions and local geometric causes, this dual-model approach mimics the selective attention of a human specialist. The result analysis demonstrates that this strategy effectively transforms opaque diagnostic processes into a transparent and human-understandable format, thereby enhancing the reliability of automated systems for early crop management in hazardous or large-scale agricultural environments. The empirical results reveal that the suggested framework is capable of accurately classifying leaves while simultaneously localizing symptoms in an interpretable way, which can further be applied to diagnose diseases in different plants.