What is our data vision? How do we combine so many different sources into one trustworthy signal?
A short walkthrough of how we ingest, validate, model and deliver crop intelligence — and the principles that hold the pipeline together when a client asks us a question we have never been asked before.
Walkthrough · HSAT data strategy
From raw signal to decision-ready forecast.
Every number we deliver passes through four stages. Each one has its own validation gates. A value only moves downstream once it reconciles with the stage before it.
Collect every relevant signal.
Satellite imagery, weather stations, in-field surveys, official statistics, news and trade data — streamed continuously into a versioned data lake.
- Multi-source, multi-scale
- Raw data is never discarded
- Provenance is preserved end-to-end
Reconcile against the field.
Ground-truth surveys and expert review calibrate remote signals. Outliers are flagged, investigated and reconciled — not silently dropped.
- Cross-source triangulation
- Agronomist sign-off in-loop
- Known-error catalogue per geography
Explainable, not just accurate.
Forecasts are built from transparent models. Every output carries the inputs that drove it, so clients can audit, challenge and retrain against their own priors.
- Per-field yield and area models
- Confidence intervals, not point estimates
- Historical back-tests published openly
Into the tools you already use.
Dashboards, API, CropGPT, and direct-to-desk briefings. Every delivery channel reads the same underlying record — no forked truth.
- API, dashboard and chat interfaces
- Versioned, reproducible outputs
- SLAs for refresh and latency
Remote and ground signals, triangulated.
No single source is sufficient on its own. The signal emerges from the overlap — which is why each HSAT product leans on a mixture of the feeds below.
Four rules that hold the pipeline together.
Ground truth beats averages.
A national figure is a starting point, not an answer. Every headline number is disaggregated down to the field wherever possible.
In practice: regional statistics are used as a prior; we publish the delta between our reconciled field-level figure and the official number — not the other way around.
Provenance is a feature.
Every delivered value can be traced back to its raw inputs — the pass, the station, the survey, the report.
In practice: forecasts ship with a metadata sidecar listing the contributing sources, their weights and their freshness.
Models are explainable by default.
No black boxes. If a number surprises you, we can tell you which signal changed and when.
In practice: we publish historical back-tests — including where we were wrong — and make confidence intervals a first-class output.
Decision-ready, not data-shaped.
Clients do not want a data lake. They want a number they can act on — with the right context next to it.
In practice: outputs are delivered through dashboards, API and CropGPT, each calibrated to the decision the user is actually making.
The Türkiye earthquake, assessed in 48 hours.
A live example of the pipeline under pressure — satellite imagery, seismic overlays and field-verified reporting combined into a rapid logistics assessment for clients.
Explore the surface area.
Want a walkthrough on your own commodity?
Tell us which crops and geographies matter to you. We will come back with a proposal — and a view of how the strategy applies to your data — within one working day.