11-07, 12:15– (Australia/Hobart), Anglesea Room
We present a cost-effective few-shot classification methodology applied to identifying historic sheep dip locations. Using AI foundation model features (DINOv2) and multi-modal large language model (GPT-4o) prompt engineering we can efficiently locate probable sites using only a small set of reference data – without finetuning underlying models.
This project aimed to develop and trial cost-effective low-data methodology leveraging pre-trained foundation and generative AI models to locate historic sheep dip locations within the Southland region of New Zealand. For this work our calibration dataset consisted of 32 known sheep dip sites.
Our two-stage approach (1) prefilters locations of interest using similarity matching in the semantically rich DINOv2 feature space, then for each location (2) applies a multi-stage prompt chain through the vision enabled GPT-4o large language model (LLM) to reason whether there are indications of sheep dipping activities and provide a corresponding prediction on the likelihood of a sheep dip being present. To reduce hallucinations, we employ chain of thought prompting techniques, explicitly instruct the LLM to review observations and inferences at multiple steps, and provide its final judgement in predefined subjective categories (i.e. certainly, likely, unlikely, certainly not).
Further work following on from this project is to ground truth identified historic sheep dip locations with fieldwork and soil sampling.
Environmental Scientist and Geospatial Analyst with e3Scientific.
I am the Head of Artificial Intelligence at Intranel, a small tech consultancy in Christchurch, NZ. I have 7+ years experience developing and deploying bespoke AI solutions for solving industry lead problems spanning agricultural, medical imaging, environmental, and conservation applications.
