AgML-FM: Benchmarking and Adaptation Infrastructure for Robust Agricultural Vision Foundation Models
Year 5 Research Project

AgML-FM: Benchmarking and Adaptation Infrastructure for Robust Agricultural Vision Foundation Models

Description

Vision-language and vision foundation models are increasingly applied to agricultural imagery for scouting, diagnosis and phenotyping. However, general-purpose models often fail on fine-grained, crop-specific distinctions under real field conditions. Our recent work shows that zero-shot vision-language models reach only about 30% accuracy on 27 AgML disease, pest and weed classification datasets. This project expands the AIFS-supported AgML ecosystem, which already standardizes more than 69 public datasets with about 5 million labeled images, into AgML-FM: an open, reproducible infrastructure to evaluate and adapt foundation models for agriculture. We will (1) define an evaluation rubric that covers crop and context diversity, biological understanding, visual task diversity and diagnostic reasoning; (2) build a two-tier benchmark for in-distribution performance and out-of-distribution robustness, including held-out crops and regions and visually confusable cases; (3) compare zero-shot, few-shot and parameter-efficient fine-tuning strategies; and (4) release a containerized evaluator and scoring scripts through AgML on GitHub and Hugging Face. AgML-FM helps growers, extension specialists and agtech companies avoid deploying brittle models.

Team

Portrait of Mason Earles

Mason Earles

Principal Investigator