Tessera embeddings · head-model predictions
This map compares a satellite-based model against our vineyard records. The model's analysis is from the 2024 growing season.
Click a polygon, pick an answer, and add a note about what's actually there.
We're testing whether AI models trained on satellite data can reliably map vineyards across New York State. The satellite layer condenses a full year of radar and optical imagery (10 m resolution, 2024) into a signature for every spot of land; a model trained on ~11,000 known blocks in the Finger Lakes and Lake Erie regions then scores how vineyard-like each spot is — that's the blue P(vineyard) overlay, with the threshold slider controlling how confident a pixel must be to show.
On held-out areas the model is right far more often than not (ROC-AUC ≈ 0.98), but it flagged 1,687 possible unmapped vineyards (~5,000 acres) and 602 mapped blocks it can no longer see. That's where you come in: your local knowledge tells us which of these are real. Every verdict improves both the vineyard records (every.farm) and the next version of the model — the long-term goal is an accurate, self-updating vineyard census for the state.
Your name and notes are stored with your responses so we can follow up and credit contributions. Questions? Contact Nick at Orbitist.