Two open-source, validated data layers for evaluating agricultural climate risk in the U.S. Corn Belt (Iowa and South Dakota). Layer 1 detects crop failure from satellite cropland maps ("did the crop fail?"); Layer 2 derives phenology-anchored extreme-weather stress covariates from daily weather grids ("did the weather warrant it?"). Both are validated against independent ground truth.
Layer 1 — Crop failure detection. Annual 30 m binary rasters of fallow / unplanted cropland from the USDA Cropland Data Layer (code 61 only). Validated against USDA FSA prevented-plant records.
Layer 2 — Phenology-anchored stress covariates. Per-pixel-year weather-stress covariates anchored on calibrated phenology (corn silking, soybean R5): 10 individual indices, 6 spring water-balance terms, and 3 compound scores, built from GridMET. Validated against NASS county yields (leave-one-year-out cross-validation).
Both layers use EPSG:5070 and ship as Cloud-Optimized GeoTIFFs.
A partnership between Fractal Agriculture and Earth Genome, funded by the Cisco Foundation.
Open-source, validation-grade data layers describing agricultural climate risk for the U.S. Corn Belt (Iowa and South Dakota). Two independent layers:
Both layers are distributed as Cloud-Optimized GeoTIFFs (COG) in EPSG:5070 (NAD83 / Conus Albers), with supporting tabular validation data as CSV.
A partnership between Fractal Agriculture and Earth Genome, funded by the Cisco Foundation.
All rasters follow the naming convention <variable>_<YEAR>.tif.
Note the two layers are at different native resolutions (30 m vs ~4 km) and carry different nodata conventions — they are separate products that share a projection and study area, not a single stacked cube.
crop-failure/)Annual binary rasters flagging fallow / idle cropland, the remote-sensing signature of an unplanted or failed field.
Method. Each annual CDL is clipped to the IA + SD study area and reclassified to a binary fallow mask using CDL code 61 (Fallow/Idle Cropland) only, per Lobell et al. (2024, Nature Sustainability). Earlier work counted codes 122/190/195 (developed open space, woody/herbaceous wetlands) as fallow; that broad definition produced a spurious ~4× "fallow spike" in the 2025 CDL (the first Sentinel-2-based release) and was retired. See caveats below.
fallow_by_county.csv — one row per county-year: year, fips, county_name, fallow_pixels, valid_pixels, fallow_acres, valid_acres, fallow_pct.
Validation. Fallow area is validated against USDA FSA prevented-plant acreage at the county level. 2019 (a regional prevent-plant year) is the clearest positive signal; most other years show low background fallow, as expected.
extreme-weather/)Per-pixel, per-year weather-stress covariates anchored to calibrated crop phenology rather than fixed calendar windows, so the stress-accumulation window tracks each pixel's actual reproductive period.
Layer 2 is phenology-anchored only. An earlier static-window representation (Apr–Oct EDD, growing-season precipitation z-scores, PDSI/SPEI 6-category classification) was retired after head-to-head validation against NASS county yields showed it had no usable predictive signal at county scale. It is not included in this product.
phenology/ — calibrated phenological dates (day-of-year rasters)Calibration offsets shift model first-crossing dates to the NASS-50% mid-crop reference (+8 days corn, +11 days soy). Post-calibration bias vs NASS is 0.0 ± 2.7 days (corn) and −0.3 ± 5.1 days (soy). QA tables (*.csv) hold the NASS comparison and the S_soy sensitivity sweep.
covariates/ — individual stress covariates16 covariates, each a per-pixel-year scalar over a phenology-anchored window (days relative to S_corn / S_soy):
corn_edd (extreme degree-days > 29 °C), corn_vpd (VPD exceedance), corn_eddi_onset / corn_eddi_peak (evaporative demand drought index), corn_night_heat (nighttime-heat exceedance), corn_spi1 (1-month SPI).soy_edd (> 30 °C), soy_vpd, soy_spi6, soy_spei6 (6-month SPI / SPEI, trailing to S_soy).spring_wb_{apr,may,jun}_mm (mm) and spring_wb_{apr,may,jun}_sd (standard-deviation anomaly vs 1981–2010).Percentile-based covariates use a 1981–2010 (WMO-standard) historical baseline.
compound_stress/ — compound scores + excess moistureThe corn compound is an empirically-derived weighted sum (the multiplicative 5-component form underperformed corn_spi1 alone and was dropped); the soy compound stays multiplicative because co-occurrence adds large tail-risk signal. See the methodology writeup for the derivation.
validation/ — NASS county-yield validationLeave-one-year-out cross-validation of the covariates against detrended NASS county yields (1995–2024, ~4,100–4,500 corn/soy county-years). Headline discrimination of yield-failure events (reduced feature set, auc_headline.csv):
By-state AUC reaches 0.85 (corn, SD) and 0.81 (soy). The phenological features beat the retired static-window features by ΔAUC +0.09 to +0.21 across every crop × threshold combination. Supporting tables: auc_headline.csv, auc_reduced.csv, auc_ablation.csv, lofo_auc.csv, model_coefficients.csv, predictions.csv, county_panel.csv, and the task*.csv series.
soy_compound correlates somewhat more strongly with corn yields than soy yields at county scale; this is documented and expected to be revisited at pixel-level validation.Licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0). You may use, adapt, and redistribute this data for any purpose, including commercial use, provided you give appropriate credit. See the LICENSE file for details.
Fractal Agriculture & Earth Genome (2026). Open Ground: validated data layers for agricultural climate risk (Iowa & South Dakota). Funded by the Cisco Foundation.
Emma Fuller · emma@fractal.ag · Fractal Agriculture