Crop yield forecasts that get sharper every week of the campaign.
Rindecast combines Sentinel-2 imagery, weather records and on-farm management data into probabilistic yield forecasts for the Argentine Pampas โ delivered per field and per 300 mยฒ hexagon, updated from planning through harvest.
Every input decision is made months before anyone knows what the field will do.
A grower commits seed, hybrid, density and fertiliser in September. The yield that actually arrives in April can swing between 6 and 14 t/ha on the same field, largely on how much rain falls in a two-week flowering window in December. Most tools answer late, with one number, for a whole field.
Decisions are staged, forecasts aren't
Choosing which crop, choosing how to sow it, and managing a standing crop are three different questions asked by three different people at three different times. A single end-of-season estimate helps none of them.
A field average hides the field
A lot averaging 10 t/ha routinely contains zones at 6 and zones at 14. Variable-rate seeding and fertilisation need that structure, not the mean.
A point estimate hides the risk
"9.8 t/ha" and "9.8 t/ha, 80% chance between 7 and 12" support completely different decisions about how much nitrogen to buy.
One model family, three decision moments.
The forecast is re-issued as information arrives. It starts as a wide distribution built from field history and seasonal ENSO signal, and tightens each week once the crop is visible from orbit.
Planning
Before the crop is chosen. Field history, soil analysis, previous crop and seasonal outlook produce a yield scenario for each candidate crop โ maize, soy, wheat/soy double-crop, sunflower, sorghum. Answers: what should this field grow?
Organisation
The crop is set; the management is not. Sowing date, hybrid, density and the N-P-S plan enter the model against a 30โ60 day sub-seasonal forecast. Answers: how should we sow it?
In-crop
The crop is growing and observable. NDVI and vegetation indices, actual rainfall and temperature, accumulated stress, achieved stand and phenological stage refresh the forecast every week. Answers: what will it actually yield?
View as table
| Stage | P10 (t/ha) | P50 (t/ha) | P90 (t/ha) | Interval width |
|---|
Forecasts at 300 mยฒ, not per farm.
Every source โ harvest monitor, satellite, weather, soil โ is normalised onto a shared Uber H3 resolution-12 hexagonal grid. That makes heterogeneous datasets from different companies directly comparable, and it puts the prediction at the scale precision equipment actually operates on.
Why the grid matters
- Hexagons have uniform neighbour distance, unlike squares โ spatial features like local slope and neighbourhood stress behave consistently.
- One join key across every source. Sentinel-2 pixels, radar rainfall, soil samples and combine yield monitors all land in the same cell.
- Prescription-ready output. Sub-field predictions map straight onto variable-rate seeding and fertilisation.
- Recoverable coverage. Where harvest maps are missing or unreliable, satellite-to-yield relationships learned on clean fields fill the gap.
The crop tells you what it's doing โ if you're watching every five days.
Sentinel-2 returns a 10 m multispectral image of every field roughly every five days. The shape of the vegetation-index curve โ when it rises, how high it peaks, how fast it senesces โ carries most of the in-crop yield signal, and a departure from the historical curve is an early warning long before harvest.
View as table
| Date | Current campaign | 5-season median | ฮ |
|---|
A spatial feature store is the asset.
Models come and go. What compounds is a clean, quality-scored, spatially standardised record of what was planted, what the weather did, what the satellite saw and what the combine measured โ across many farms and many seasons.
Cloud-native by necessity
A single Pampas season is millions of hexagons ร a 5-day image cadence ร a decade of history. Ingestion, index computation and backfill run as elastic batch jobs against object storage rather than on fixed hardware.
Open data where it exists
Sentinel-2 is published as cloud-optimised imagery in public cloud registries, so imagery is read in place instead of being copied. Public soil, climate and radar products are integrated the same way.
Physiology plus learning
Process-based crop models supply physiological structure โ degree-day accumulation, stage-specific stress sensitivity โ that machine learning alone tends to get wrong in seasons unlike the training set.
Development roadmap.
Rindecast is an early-stage company building on applied research into combining process-based crop modelling with machine learning for early maize in the Argentine Pampas.
| Milestone | Horizon | Outcome |
|---|---|---|
| Data architecture | Months 0โ6 | Common data model across partner farms, H3 r12 grid implemented, first formal QA/QC pass and per-source reliability diagnosis. |
| First predictive model | Month 12 | In-crop dynamic model for early maize with temporal and regional validation; probabilistic output per field and per hexagon. |
| Multi-model, multi-crop | Month 24 | Planning, organisation and in-crop models across maize, soy, wheat/soy, sunflower and sorghum on a stable feature store. |
| Decision simulation | Month 36 | Prescriptive layer โ simulate management and climate scenarios, evaluate technical and economic impact. |
Partner farms and agtech collaborators welcome.
We work with Argentine agricultural companies willing to contribute historical production data in exchange for forecasts on their own fields.
hello@rindecast.com