A Visual Question Answering Dataset for Benchmarking Vision-Language Models in Plant Science

Area of Science:

• Agricultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Existing plant disease datasets primarily support classification and detection tasks.
• This limits the application of vision-language models in interactive, reasoning-based plant disease diagnosis.
• There is a need for comprehensive datasets to train AI for complex agricultural decision-making.

Purpose of the Study:

• To introduce PlantExpertVQA, a large-scale visual question answering dataset for agricultural decision-making.
• To facilitate the development of advanced vision-language models for plant disease diagnosis.
• To benchmark current AI models and demonstrate effective domain adaptation strategies.

Main Methods:

• Compiled 765,186 question-answer pairs from 45 open-source datasets, including PlantVillage.
• Utilized a two-stage pipeline for automated QA generation with expert linguistic review.
• Images covered 38 crop species and 89 disease conditions, with questions categorized by complexity and type.
• Dataset underwent iterative review by domain experts for accuracy and relevance.

Main Results:

• Current state-of-the-art vision-language models, including multimodal LLMs, exhibit poor performance on PlantExpertVQA.
• Parameter-efficient fine-tuning of a 2B-parameter model on a subset of the dataset led to significant performance gains.
• Improvements were observed across all question categories, validating the dataset's utility for domain adaptation.

Conclusions:

• PlantExpertVQA is a valuable resource for advancing AI in agriculture, particularly for diagnostic applications.
• Effective domain adaptation is achievable with compact models and targeted fine-tuning on specialized datasets.
• The findings highlight the potential for AI to support farmers in complex decision-making processes.