Scholarlink

Rice diseases cause severe yield losses, accounting for up to 30% of global rice production. Traditional machine learning approaches for image-based classification face notable limitations, particularly in misclassifying diseases with visually similar characteristics. To address this challenge, we propose XAI-ML-Rice. This model leverages descriptive text for rice disease classification, a modality more intuitive for farmers, while incorporating explainable AI techniques (the SHAP framework) to evaluate the strengths and weaknesses of different machine learning models. The dataset comprises 435 textual descriptions of five rice diseases, with preprocessed Vietnamese texts averaging 92–101 tokens. Experimental results demonstrate that the SVM model achieved an accuracy of 97.7%. Furthermore, Beeswarm plots and Global Feature Importance analyses revealed that the model successfully identified disease-relevant features, confirming its effectiveness. This study highlights the potential of diverse text-based approaches. It suggests future directions that integrate multimodal data (images and text) and adaptive explanation mechanisms, thereby enabling more interactive and context-aware decision support for end users.

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

SHAP-Based Interpretability of Machine Learning Models for Rice Disease Classification from Textual Descriptions

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Luyl-Da Quach,
  • Thanh-Khang Tran,
  • Chi-Ngon Nguyen,
  • Nguyen Thai-Nghe

摘要

Rice diseases cause severe yield losses, accounting for up to 30% of global rice production. Traditional machine learning approaches for image-based classification face notable limitations, particularly in misclassifying diseases with visually similar characteristics. To address this challenge, we propose XAI-ML-Rice. This model leverages descriptive text for rice disease classification, a modality more intuitive for farmers, while incorporating explainable AI techniques (the SHAP framework) to evaluate the strengths and weaknesses of different machine learning models. The dataset comprises 435 textual descriptions of five rice diseases, with preprocessed Vietnamese texts averaging 92–101 tokens. Experimental results demonstrate that the SVM model achieved an accuracy of 97.7%. Furthermore, Beeswarm plots and Global Feature Importance analyses revealed that the model successfully identified disease-relevant features, confirming its effectiveness. This study highlights the potential of diverse text-based approaches. It suggests future directions that integrate multimodal data (images and text) and adaptive explanation mechanisms, thereby enabling more interactive and context-aware decision support for end users.