Identifying Potential Historic Sheep Dip Locations in Southland usi
summary
Project uses 32 labeled sites to calibrate a two-stage pipeline that first tiles historic aerial imagery and runs DINOv2 prefilter similarity matching, then passes candidate tiles to a vision LLM — GPT‑4o multimodal with chain-of-thought — to describe and classify likelihood. Trial on 186 tiles surfaces ~760 candidate sites (194 high, 298 medium, 268 low); high-confidence false negatives are ~5%, overall accuracy rises from ≈65% to ≈75%, and the two-stage design avoids an estimated $2M single-pass cost.
AI-written summary · about
chapters (12)
- 0:00Introduction and project overview
- 2:28Challenge of finding sheep dips
- 3:31Machine learning approach limitations
- 4:58Transformer models for visual data
- 6:24Two-stage AI processing method
- 8:53Image preprocessing and feature extraction
- 10:39Reference data and similarity matching
- 12:24Multimodal model classification
- 13:50Human review and categorization
- 15:38Results and accuracy assessment
- 17:22Cost and efficiency benefits
- 18:46Future applications and Q&A
related talks
31mFrom Radar Echoes To Real-Time Recognition: Deep Learning on Another Planet -An Open Source Workflow
36mClassification of Satellite Imagery
45mFrom Radar Echoes To Real-Time Recognition: Deep Learning on Another Planet an Open Source Workflow
41mLarge Scale Field Delineation
33mWhat and where? - Machine learning for geospatial image analysis
21mNeedle in the Haystack: Applied ML for Inspections of Power Infrastructure
28mGeoAI: Evolution of AI-Based Tree/Vegetation Detection Models Using Imagery
46mML Use Cases at Think Big Analytics