Scholarlink

<p>Every year, the global agrarian economy faces severe losses inflicted by plant diseases. Overcoming this is a significant challenge due to the expertise required in plant disease identification and classification. While machine learning applications are not new to solving this problem, they rely on heavy, non-local datasets which is a hurdle when they have to be deployed in specific regions. To address this, we scrape a novel dataset, tailored to the Indian state of Tamil Nadu. To provide additional context to the images, we propose to develop a second mode of data, text descriptions, which are more abundantly available and can be augmented using large language models. Given the lack of large, structured data for the task, we intend to use a few-shot learning (FSL) approach. Specifically, we propose to use a two-step approach for FSL that incorporates contrastive pre-training to extract better latent features and a prototype-based classification network to perform disease detection. The detected disease combined with the farmers’ question is used to query an advisory document using a retrieval augmented generation pipeline. Our approach achieved an accuracy of 93% in disease classification and a context precision score of 0.87 in query resolution. This shows great potential to contribute to the food security and sustainable agriculture landscape of Tamil Nadu.</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

Multimodal few-shot learning for plant disease detection with contrastive pre-training and query addressal

  • 期刊论文

  • 发表日期：

    2025年7月1日

查看全文
  • P Pranith,
  • Varsha Yeshwanth,
  • Durairaj Thenmozhi

摘要

Every year, the global agrarian economy faces severe losses inflicted by plant diseases. Overcoming this is a significant challenge due to the expertise required in plant disease identification and classification. While machine learning applications are not new to solving this problem, they rely on heavy, non-local datasets which is a hurdle when they have to be deployed in specific regions. To address this, we scrape a novel dataset, tailored to the Indian state of Tamil Nadu. To provide additional context to the images, we propose to develop a second mode of data, text descriptions, which are more abundantly available and can be augmented using large language models. Given the lack of large, structured data for the task, we intend to use a few-shot learning (FSL) approach. Specifically, we propose to use a two-step approach for FSL that incorporates contrastive pre-training to extract better latent features and a prototype-based classification network to perform disease detection. The detected disease combined with the farmers’ question is used to query an advisory document using a retrieval augmented generation pipeline. Our approach achieved an accuracy of 93% in disease classification and a context precision score of 0.87 in query resolution. This shows great potential to contribute to the food security and sustainable agriculture landscape of Tamil Nadu.