Trust & Transparency

How Qroplet works
under the hood

Every recommendation, alert, and forecast traces back to a specific data source and a documented method. This page is the audit trail.

01 · Data Sources

What we measure, and where it comes from

Every parcel insight is grounded in named, third-party data — not opaque proprietary datasets. Updates flow on the cadence each provider publishes.

  • Satellite

    Sentinel-2 (Copernicus / ESA)

    Optical imagery used for NDVI, NDWI, and EVI. 10 m resolution; revisit every 5 days under cloud-free conditions.

    We ingest Sentinel-2 L2A surface reflectance (Bands B03, B04, B08, B8A, B11) and mask clouds via the scene classification layer (SCL).

  • Satellite

    Sentinel-1 (Copernicus / ESA)

    Synthetic Aperture Radar (SAR) gives us all-weather visibility — critical when clouds hide optical sensors during early- and late-season weeks.

    VV/VH polarizations from Sentinel-1 GRD; used as a fallback when Sentinel-2 coverage is insufficient and as a moisture / structure cross-check.

  • Climate

    ERA5 (Copernicus Climate Change Service)

    Long-run climate baselines: temperature, precipitation, aridity per parcel. Used by the crop suitability model and by anomaly thresholds.

    Hourly ERA5 reanalysis aggregated to monthly and seasonal climatologies for each parcel centroid.

  • Weather

    Open-Meteo

    Short-term and decadal forecasts (frost, heat, rain) used by alert logic and the decadal weather card.

    ECMWF IFS and ICON models routed through Open-Meteo, queried per parcel centroid with 24-hour cache.

  • Soil

    ISRIC SoilGrids

    Soil pH, clay/sand/silt fractions, and organic carbon for the suitability model — your parcel's actual soil profile, not a regional average.

    SoilGrids v2.0 raster sampled at parcel centroid for the 0–30 cm horizon (six depth slices weighted by depth).

  • Topography

    SRTM / Copernicus DEM

    Elevation, slope, and aspect inputs. Drives part of the crop suitability score and is used by water-stress context.

    30 m SRTM where available, falling back to Copernicus GLO-30. Slope and aspect derived per parcel polygon.

02 · Models & Algorithms

From raw data to a decision

Every model has a name, a training set, and a known set of failure modes. We document them so you know what you are trusting.

  • Crop suitability

    Matches 135 crop profiles to your parcel's soil, climate, and elevation. Returns Top-N crops with an explainable score breakdown.

    Technical notes
    • LightGBM gradient-boosted trees, trained on regional yield archives and EcoCrop / FAOSTAT suitability envelopes.
    • Features: 12 soil variables (ISRIC), 9 climate variables (ERA5), 3 topographic variables (SRTM).
    • Per-recommendation factor attribution shown in the UI ("why this score").

  • NDVI / NDWI / EVI time series

    Per-parcel vegetation and moisture indices, smoothed to remove cloud and sensor noise before downstream models see them.

    Technical notes
    • Indices computed from Sentinel-2 surface reflectance (e.g. NDVI = (NIR − Red) / (NIR + Red)).
    • Cloud masking via SCL; per-week per-parcel aggregation weighted by clear-pixel coverage.
    • Savitzky-Golay smoothing applied to the weekly series with a fallback to Whittaker smoothing for sparse coverage.

  • Phenology detection

    Identifies Start of Season (SOS), Peak (POS), End (EOS), and Length of Season (LOS) from the smoothed NDVI curve.

    Technical notes
    • Asymmetric Gaussian fit to the smoothed NDVI series.
    • Rule-based threshold detection for SOS/POS/EOS, validated against ground-truth observations and regional crop calendars.
    • Confidence bounds derived from fit residuals and clear-observation density.

  • Water stress detection

    Flags water-stress events while the crop still looks healthy to the naked eye, so irrigation decisions arrive in time.

    Technical notes
    • Adaptive NDWI threshold relative to the parcel's seasonal baseline (not a global cutoff).
    • Alert triggers only when the anomaly persists for ≥3 consecutive observations to suppress single-frame noise.
    • Phenology phase is cross-checked so EOS-phase decline does not generate false alerts.

  • Yield forecast

    A parcel-level yield estimate, with a confidence interval. Built to inform plans — not to replace ground-truth measurement.

    Technical notes
    • LightGBM quantile regression with conformal prediction bands at 80 and 95 percent.
    • Inputs: phenology timing (LOS, SOS), NDWI anomaly duration, soil water proxies.
    • Calibrated against NUTS-3 regional yield archives; mean absolute percentage error tracked per crop × country.

  • Parcel zoning

    Splits each parcel into Low / Mid / High productivity zones — the foundation for variable-rate fertilizer and irrigation.

    Technical notes
    • Q33 / Q66 quantile classification on the within-season NDVI distribution per parcel.
    • Zone stability measured across seasons (Jaccard overlap) and reported alongside the current zoning.
    • Boundaries exported as GeoJSON, ready for variable-rate equipment.
03 · Confidence & Explainability

Every output tells you how much to trust it

A number on its own is just a guess. Qroplet ships every recommendation with the signals that produced it and a confidence band where applicable.

  • Driver attribution

    Every crop recommendation shows which factors moved the score most — soil pH, water availability, temperature band, etc. No black-box "trust us" outputs.

  • Confidence intervals

    Yield forecasts ship with 80 and 95 percent prediction intervals from conformal calibration. Wider intervals mean lower-quality inputs — and we surface that, not hide it.

  • Observation coverage

    Phenology and monitoring outputs report how many clear satellite observations they were built from. Low coverage downgrades the confidence band automatically.

  • Plain-language explanations

    Raw indices stay one click away. The default view says "moderate water stress detected" — not "NDWI dropped 0.18 over 5 days" — so you can act first and inspect later.

04 · Validation

How we know it is working

Model quality is monitored continuously — drift, accuracy, and ground-truth fit are all tracked per release.

  • Phenology ground truth

    SOS / POS / EOS predictions are validated against field-observed phenology dates and regional crop calendars. Per-crop accuracy snapshots are tracked over time.

  • Yield calibration

    Predictions are compared against NUTS-3 regional yield archives. Mean absolute percentage error is recorded per crop × country, with drift detected via population stability index.

  • Drift monitoring

    Live PSI on input features. When a model's feature distribution drifts past a threshold, the model is flagged for retraining before its quality silently degrades.

  • Model lineage

    Every prediction is tagged with the model version that produced it. Audit trails answer "which version made this decision and what data did it see".

05 · Limitations

What Qroplet is not

Trust is built as much by what we refuse to claim as by what we promise. Here is where the platform is honest about its limits.

  • Not a ground-truth yield predictor

    Yield forecast is a decision support baseline. Use it to compare parcels, plan resources, and time interventions — not to replace harvest measurement or an agronomist's judgment.

  • Not a replacement for field inspection

    Satellite indices catch most stress events early, but pest damage, micro-scale issues, and ground-level details still need eyes on the field.

  • Cloud-cover dependent (optical)

    Sustained cloud cover delays NDVI / NDWI updates. Sentinel-1 SAR fills part of the gap, but extended overcast periods will widen confidence bands.

  • Regional model coverage

    Yield calibration and crop suitability quality depend on the depth of regional training data. New regions ship with broader confidence intervals until enough local data accumulates.

See it on your own parcel

The fastest way to evaluate the methodology is to point it at land you know.

Add Your Parcel See pricing