Loading statement…
We are building an agentic workflow that pairs every.farm (farm context and remote sensing) and crops.every.farm (an audited crop catalog) with a Claude-orchestrated reasoning layer to help researchers and growers ask two questions at the field scale: which parts of my farm are best suited to diversification, and what should I plant there? Our first test case is the Cornell Lake Erie Research and Extension Laboratory (CLEREL) in Portland, NY.
How to read this report. The narrative below explains the motivation, the system we are building, what we have produced at CLEREL so far, and how we plan to assess whether the recommendations are actually useful. Several sections are interactive — click the diagram nodes, drag the criterion sliders, or click any candidate zone on the map to drill in.
Specialty crop monocultures across the Lake Erie Grape Belt — Concord and a handful of vinifera and hybrid wine varietals — face compounding pressure from climate volatility, processor consolidation, and labor and input cost inflation. The economic case for adding complementary crops is increasingly clear, but the operational case is hard: which parts of which fields, with which crops, with what payback horizon?
March 2026 — a concrete example. Refresco, one of the major Concord grape processors serving the Lake Erie Grape Belt, abruptly canceled contracts with growers across western New York, Pennsylvania, and Ohio — after many had already invested in the 2026 season. Smaller-acreage growers who delivered only to Refresco face the steepest fall: no buyer, sunk costs, and a real possibility of mothballing vineyards to cut losses. The episode is a stark, recent illustration of the structural risk that motivates this work — that single-processor exposure on a single specialty crop can flip from steady to existential overnight.
Decision support for this question has historically meant a soils map, a yield monitor printout, and a conversation with an extension agent. That works, but it does not scale, it can’t easily integrate the full open-data toolbox of remote sensing and climate records, and it can’t cheaply re-run as new imagery arrives, prices move, or the catalog of viable crops grows.
Our hypothesis is that an agent-based workflow, built on top of well-defined APIs for farm context and crop intelligence, can give researchers and growers a faster, more auditable, and continuously-updateable starting point for these conversations — without replacing the expert review and on-farm trials that ultimately validate any recommendation.
The two research questions the system aims to answer:
1. Identify candidate geographies — the sub-field regions where diversification is most justified by underperformance, biophysical distinctness, edge/transition character, and economic opportunity.
2. Identify candidate crops — the audited catalog entries that match each region’s soil, drainage, climate, and market context, ranked for fit and economic upside.
Click any node in the diagram to expand its role in the pipeline.
Click a node above to read more.
Three layers of the system are in production. The agent layer is operational and currently being exercised against CLEREL.
Browse, retrieve, and aggregate open imagery and gridded climate data into per-block time series. Twelve open datasets are wired in today, covering optical, radar, soils, terrain, climate, and land cover.
Open archives only — easy time-series aggregation per block, including pixel statistics over arbitrary date ranges.
A curated, continuously-audited catalog of crops appropriate for U.S. growing regions. The April 2026 audit refreshed prices and added per-crop suitability matrices for soil texture and drainage class, alongside a 5-axis “snowflake” regional fit score. The catalog is still actively growing.
Full schema: pH range, GDD, frost-free days, water requirement, years to production, gross/net return, price trend, insurance availability, snowflake axes (market · climate · infrastructure · financial · risk).
The agent assembles raster stacks from every.farm, computes per-pixel sub-criteria (Underperformance, Distinctness, Edge, Opportunity), segments candidate zones, and queries the catalog with each zone’s biophysical profile to surface a ranked crop shortlist. Each step is reproducible; provenance is preserved end-to-end.
Block-aware imagery search, time-series aggregation, and field-data capture from one workspace.
Field NDVI distribution and per-block analysis on the desktop, paired with a mobile capture flow for in-situ measurements.
A grower-facing screener over the audited catalog, filterable by region, lifecycle, category, revenue, and price trend.
Crop Picker by Every.Farm — filtered to fruit / perennial / increasing-price-trend in the Lake Erie Concord Grape Belt.
Hover any point to see crop details. Position is regional fit (snowflake total) versus net return per acre.
The dashed green line is the Pareto frontier: the set of crops where no other crop in the catalog has both higher regional fit and higher net return. Crops on the frontier (haloed in green) are the candidates a rational grower could not strictly improve on without trading off one axis against the other. Crops below it are dominated — there is at least one crop with both better fit and better return.
Note that the snowflake score is a regional-fit composite (market, climate, infrastructure, financial, and risk) rather than a fit to any specific field. The pipeline composes this regional confidence with per-pixel suitability matching to produce per-zone recommendations. The chart filters to crops with snowflake ≥ 12 and a positive net return; lower-fit or unaudited crops are still in the catalog but not plotted here. The CLEREL analysis below was run against the April 2026 catalog snapshot (32 crops with both fields populated); the current live catalog is larger and a re-run would incorporate the new entries.
A composite score combines four sub-criteria: Underperformance (U), Biophysical Distinctness (D), Edge/Transition (E), and Economic Opportunity (O). Pixels above the 70th percentile within the vineyard are smoothed and segmented into candidate zones (PPZs), then matched against the catalog. Click any zone on the map below to inspect it.
Click any candidate zone on the map to drill in. The panel below updates with that zone’s biophysical context and top crop matches.
The four sub-criteria are combined into a single composite C with weights set during the experiment. There is no single correct weighting — yield-driven or market-driven framings shift which zones float to the top. Drag the sliders or pick a preset.
Weights are renormalized to sum to 1.
Highlighted rows changed position relative to the previous weight setting.
Each criterion picks out a different pattern at CLEREL. Tab through to read what each is doing.
Highest values concentrate in the northern Concord/Niagara blocks (varietal-adjusted residuals) and portions of Taft with elevated yield CV.
Subscript a denotes min-max normalization within the vineyard footprint. The widely-accessible fallback uses Sentinel-2 NDVI mean (inverted) and NDVI temporal standard deviation when yield monitor data is not available.
Picks out small hydric and high-ECa patches inside Concord — the kinds of micro-habitats that resist a one-size-fits-all viticultural prescription.
Robust covariance via Minimum Covariance Determinant. Features include DualEM ECa deep + shallow, TWI, slope, elevation, and Sentinel-2 NDVI mean. The widely-accessible fallback substitutes POLARIS clay/sand/OM/pH for ECa.
Strongest along the western vineyard boundary and in the southern headland. These are natural homes for hedgerows, pollinator strips, and N-fixers.
Edge term decays with distance to the field boundary on the order of one row width. TWI and slope contribute hydric-awkwardness and steepness.
No longer near-saturated: per-pixel texture + drainage matching against 32 audited crops yields a genuine gradient instead of a uniform upper bound.
Each crop is scored against the pixel’s drainage class, USDA texture, slope, elevation, and pH. The best-crop expected value minus a $2,500/ac Concord baseline is the pixel’s opportunity delta.
Each criterion highlights complementary patterns. The composite is a weighted average; segmentation runs against it after smoothing.
Weights default to a yield-and-economics emphasis. The slider above lets you explore alternative framings — yield-led, market-led, distinctness-led — and see how the candidate ranking responds.
A second pipeline restricts inputs to widely-accessible layers only — Sentinel-2, USGS 3DEP, POLARIS, PRISM. The two pipelines agree on roughly half the candidate pixels at CLEREL.
Interpretation: the open-data pipeline is a defensible baseline — it covers similar total acreage and converges on the largest cross-block zones (PPZ 6 and PPZ 2) — but on-farm yield and ECa pull additional precision out of the model where they are available. A service tier structure that offers an open-data baseline plus an optional ECa-survey premium tier is plausible.
Querying crops.every.farm against the dominant CLEREL diversification zone profile (silt loam, moderate-to-well drainage, ~6.2 pH, Zone 6a), the same three crops surface across most zones.
Why hardy kiwi keeps surfacing as the top expected-value pick at CLEREL, and what makes it interesting beyond the headline net return:
An agentic system that produces beautiful maps is only useful if its recommendations hold up against expert review and on-farm reality. Our research plan is a structured loop between the model, the experts, and the field.
Sensitivity analysis on composite weights and segmentation thresholds; replace single-crop matching with polyculture-combination scoring; add per-zone NPV from establishment + operating cost when those fields are backfilled in the catalog.
Walk each candidate zone with viticulturists, soil scientists, and economists. Ask each reviewer: is this a defensible recommendation? Capture disagreements as structured feedback that re-enters the scoring function.
Pilot one or two of the top crops (likely hardy kiwi and a companion) on a sub-block of a high-scoring PPZ at CLEREL. Treat the trial as a measurement of the system, not just of the crop.
Beyond CLEREL, run structured interviews with Grape Belt growers, wholesale buyers, and processors to validate market assumptions. Surface the parts of the catalog where buyer proximity should down-weight a recommendation.
Package the per-farm output into a deliverable that an extension agent can hand to a grower as a conversation starter — interactive map, per-zone shortlist, clear caveats. Iterate the format with extension staff.
Once the CLEREL loop closes, run the open-data pipeline against a sample of farms across NY and PA Grape Belt counties. Use the candidate-IoU between full and open-data pipelines as a calibration metric for service tiering.
The big research question. Can an agent-based system, paired with well-defined APIs for farm context and crop intelligence, produce diversification recommendations that experts trust and growers act on? CLEREL — the Cornell Lake Erie Research and Extension Laboratory — is our first test case. The assessment loop above is the path to answering it credibly.
CLEREL