Key points
- AI vision language models (VLMs) are limited as research tools in botany and agriculture because they struggle to recognise microscopic details in plants
- Software produced by UQ PhD scholar Tianqi Wei aims to address this gap by identifying exactly where AI tools are struggling to comprehend minute details in plants and plant disease
- Mr Wei said his PlantMicro benchmarking dataset ultimately provides crucial feedback AI developers can use to improve their products for use in microscopic plant research
Software created by University of Queensland scientists could help AI platforms overcome a blind spot limiting their value in plant research.
UQ PhD student Tianqi Wei said even advanced vision language models (VLMs ) had difficulty interpreting pictures of microscopic plant details.
“VLMs are widely used as research tools in other areas of microscopy but not plant science because they can’t be relied on to accurately analyse cellular details”.
“Ultimately there are blind spots preventing us from using these tools to make discoveries that advance research in areas such as food security and environmental sustainability.”
To guide VLMs towards a more robust understanding of plant microscopy, Mr Wei and his collaborators at UQ’s School of Electrical Engineering and Computer Science have created an AI benchmarking dataset called PlantMicro .
A collaboration with UQ’s School of Agriculture and Food Sustainability and partially supported by the Grains Research and Development Corporation (GRDC), the software integrates thousands of microscopic images and uses sophisticated training mechanisms to test the ability of VLMs to identify, sort, and count minute details in plants and plant disease.
“Part of the reason there is an AI performance gap in plant microscopy is because we have not been able to see exactly where VLMs are falling short, and how they can improve,” Mr Wei said.
“With PlantMicro we can clearly identify where these major AI models are struggling.”
Mr Wei said experiments carried out with PlantMicro confirmed both closed and open-source VLMs performed poorly on plant host and disease identification, with average accuracies of about 30 per cent.
“These results are marginally above random guessing, which suggests that VLMs still face considerable challenges in identifying host or pathogen types from microscopic evidence,” Mr Wei said.
He said identifying these performance gaps was crucial feedback for AI developers to improve their products for use in microscopic plant research.
“We hope our program will provide a better foundation for AI to be used in plant science for the betterment of everyone on the planet,” Mr Wei said.
Read the research presented by Mr Wei at the European Conference on Computer Vision 2026 .
Download the PlantMicro benchmarking dataset for free here .
Collaboration and acknowledgements
This work was a collaboration with UQ’s School of Agriculture and Food Sustainability and was supported by Analytics for the Australian Grains Industry (AAGI). AAGI (UOQ2301-010OPX) is a Strategic Partnership between the Grains Research and Development Corporation (GRDC), The University of Queensland, Curtin University, and Adelaide University.
