Methodology & accuracy
Tarımus produces its parcel-scale analyses with methods established in the academic literature and accepted internationally. This page sets out, openly, which data we use, at what resolution, how we validate it, and our known limits.
Updated: June 2026
Data sources
Every parcel is read by aligning several independent data layers to the same boundary; where cloud affects the optical observation, radar steps in and the sources corroborate each other.
- 01
Optical satellite
Copernicus Sentinel-2 and Landsat — multi-band optical observation for vegetation, phenology and crop classification.
- 02
Radar satellite
Copernicus Sentinel-1 (SAR) — cloud-independent structure and moisture signal; fills the gaps in optical observation.
- 03
Thermal
Landsat thermal band — land-surface temperature, water stress (CWSI) and evapotranspiration indicators.
- 04
Climate series
Multi-year climate reanalysis and daily meteorology — drought, frost, temperature and rainfall indicators.
- 05
Soil & topography
Global soil databases (≈250 m) and a digital elevation model — water-holding capacity, slope and aspect.
- 06
Field data
Drone and ground verification by the DronGelsin field team; producer declarations and sample points.
Resolution and recency
Spatial: analyses are produced at parcel scale. Optical and radar observation is typically 10–30 m, while soil layers are in the ≈250 m band.
Temporal: satellite passes refresh regularly across the season (optical + radar combined, a usable observation every few days in most regions); climate indicators update daily.
Every output is stored together with the observation dates and pass information it rests on; an analysis can be reopened down to its source even months later.
Validation approach
Field validation: DronGelsin's operational field arm compares satellite-based forecasts with drone and ground observation; producer communication also runs through it.
Cross-validation: model outputs are tested for consistency against independent data layers (radar ↔ optical, neighboring parcel, past season).
Explainability: every metric arrives with a confidence level and is traceable down to its source; it is open to third-party audit.
Accuracy by crop
Today we offer field-validated crop classification for 20 crops; for these crops, classification accuracy is above 80%. Our taxonomy covers 200+ crops, and the scope is updated on this page as validation expands.
Accuracy is not a single number; it varies by crop, region, parcel size and season. We share the detailed per-crop accuracy table and current scope in a technical discussion.
Known limits
Stating the limits openly builds trust. We know up front the conditions under which the confidence interval widens:
- 01
Cloud cover
Dense, persistent cloud limits optical observation; radar largely compensates, but in some periods the confidence interval widens.
- 02
Small parcels
On very small or very thin parcels below the resolution, the mixed-pixel effect increases.
- 03
Mixed cropping
Multiple crops on the same parcel, or intercropping, can lower classification confidence.
- 04
Rare crops
Validation is more limited for rare crops with few labeled samples; they may be in the taxonomy but not yet field-validated.
The methods we build on
Our outputs are not opaque predictions but calculations built on established methods:
- Climate vulnerability: the IPCC framework (hazard · exposure · sensitivity · adaptive capacity).
- Water & evapotranspiration: a Penman-Monteith / FAO-56 based approach and green/blue/grey water accounting.
- Phenology: development-stage reading based on the BBCH scale.
- Carbon: mapping to the GHG Protocol, IPCC 2019 Refinement, PCAF and Verra VM0042 frameworks.
- Remote sensing: multi-band vegetation indices and SAR-based classification.
Invitation
Let's Write the Next Chapter of Agriculture Together.
Food company, exporter, bank, insurer, cooperative or ESG advisor — let's talk through a pilot report for your sourcing region or portfolio. We'll listen, and we'll show you.