Sensors
- Sentinel-1 C-band SAR (ESA): 10 m ground range resolution in IW mode, 6-day repeat at the equator (12-day per satellite, two satellites in constellation). VV and VH polarisations respond to canopy moisture content and volumetric scattering from the leaf canopy; backscatter rises as the soybean canopy closes and falls again at senescence, providing a phenological signal independent of cloud cover.
- Sentinel-2 MSI (ESA): 10 m (visible/NIR) and 20 m (red-edge, SWIR) spatial resolution, 5-day revisit with both satellites. NDVI and red-edge chlorophyll index (CIre) track green-up and maturity on cloud-free acquisitions; these optical observations anchor phenological timing that SAR alone cannot resolve to species level.
- MODIS Terra/Aqua: 250 m (bands 1-2) to 500 m resolution, near-daily global coverage. MODIS NDVI and EVI time series provide a coarse but temporally dense phenological backbone useful for regional-scale area estimation and for gap-filling where Sentinel revisit is interrupted.
- Landsat 8/9 OLI: 30 m multispectral resolution, 8-day combined revisit. Longer archive (Landsat 8 from 2013, Landsat 9 from 2021) supports multi-year planted-area trend analysis and cross-calibration with Sentinel-2 reflectance products.
Why cloud makes optical-only mapping unreliable here
The Brazilian Cerrado and Matopiba frontier, together with the Argentine Pampas and the soybean belts of Paraná state, experience cloud cover exceeding 70 percent of days during the main growing season from October to March. Parts of Mato Grosso, Brazil's largest producing state, can go six or eight consecutive weeks without a single usable Sentinel-2 or Landsat scene. Waiting for a clear-sky composite risks missing the narrow window between planting and canopy closure when classification is most reliable.
The same problem recurs in Heilongjiang and Jilin provinces in north-east China, where spring cloud and haze extend from April through June. A purely optical workflow either produces a map with large data gaps or forces analysts to use late-season imagery that conflates harvested soybean fields with bare soil from other crops. Neither outcome is acceptable to a commodity trader pricing a futures contract or an insurer calculating exposure before the season closes.
What the radar signal actually measures
C-band SAR at 5.405 GHz interacts with the soybean canopy in two ways that matter for mapping. First, volume scattering from the trifoliate leaves increases VH backscatter as the canopy develops, typically rising 3 to 6 dB between emergence and peak biomass. Second, double-bounce between stems and the soil surface contributes to VV backscatter early in the season. The temporal trajectory of VH/VV ratio across a 12-week window carries a phenological fingerprint that differs from maize, which has a coarser leaf structure and different moisture dynamics, and from pasture, which has a smoother, lower-backscatter profile.
The honest limit is that soybean and other broadleaf crops, particularly cotton and some pulses, produce similar C-band signatures during vegetative growth. SAR alone cannot reliably separate them. That is precisely why the optical component is not decorative: even two or three cloud-free Sentinel-2 acquisitions per season, timed around flowering or pod-fill, are sufficient to resolve the ambiguity when combined with the SAR time series in a supervised classifier.
Fusion architecture: how the two data streams are combined
The standard fusion approach stacks Sentinel-1 VV and VH backscatter time series (typically 10 to 15 acquisitions per season) with Sentinel-2 NDVI, red-edge index and SWIR-based indices on the dates when cloud cover is below a usable threshold, perhaps 20 percent scene cloud fraction. The combined feature stack is fed to a random forest or gradient-boosted classifier trained on field-verified reference parcels. Published studies using this architecture in the Brazilian Cerrado have reported overall accuracies in the range of 88 to 94 percent for soybean versus non-soybean classification, though figures vary with the density of training samples and the proportion of cloud-free Sentinel-2 scenes available.
MODIS time series can substitute for Sentinel-2 in the temporal backbone when fine-resolution optical data is scarce, at the cost of spatial precision. For area estimation at regional scale, the coarser resolution is often acceptable; for individual farm-level insurance exposure, 10 to 30 m data is necessary.
Double-cropping, common in Mato Grosso where soybean is followed by maize in the same season, introduces genuine attribution ambiguity. A single-date map will capture the first crop correctly but may misclassify the second crop's establishment phase as bare soil or early pasture. Handling this requires a second classification pass on imagery from the February to April window, with separate training data for the safrinha season.
From area map to commodity and insurance numbers
A planted-area polygon layer is the starting point, not the end product. For commodity supply forecasting, the area estimate is multiplied by a regional yield expectation derived from vegetation index trajectories, producing a pre-harvest production estimate that can be compared against USDA WASDE or CONAB official figures weeks before they are published. The value to a commodity desk lies in the lead time and in the independence from self-reported national statistics.
For crop insurers, the planted-area map serves two distinct functions. Before the season, it establishes the insured area, replacing or auditing farmer-declared hectarage that may be inflated. After a loss event, it provides the denominator for damage calculations: knowing that a policy covers 2,400 hectares of confirmed soybean, rather than taking the farmer's word for it, materially changes the exposure calculation. Parametric products triggered by area shortfalls relative to a declared planting intention require this kind of independent verification by design.
Satellize applies this fusion methodology operationally, building on experience from crop-estimation work including the Kingdom of Tonga programme, and can configure seasonal soybean monitoring for specific administrative regions, river basins or portfolio footprints.
Resolution floors, latency and what the map cannot tell you
At 10 m resolution, Sentinel-1 IW mode can detect fields down to roughly 1 to 2 hectares with reasonable reliability; smaller parcels are subject to mixed-pixel contamination from field boundaries and access tracks. In fragmented smallholder landscapes, this is a genuine constraint. The Brazilian soybean belt, dominated by large commercial farms typically exceeding 500 hectares, is well-suited to 10 m data. Parts of the Chinese soybean belt, with smaller average parcel sizes, push against this floor.
Processing latency from satellite acquisition to a classified area layer is typically 3 to 7 days using near-real-time Sentinel-1 and Sentinel-2 data streams, depending on pipeline configuration and the volume of the study area. This is adequate for seasonal supply forecasting but not for daily trading signals.
The map shows planted area. It does not show yield, crop health, or whether the crop will be harvested rather than abandoned. Those questions require separate analysis of vegetation index trajectories through the season. Treat the area estimate as a necessary but not sufficient input to a production forecast.
Typical figures
| Primary SAR spatial resolution | 10 m (Sentinel-1 IW mode, ground range) |
| Primary optical spatial resolution | 10 m visible/NIR, 20 m red-edge and SWIR (Sentinel-2 MSI) |
| SAR revisit (two-satellite constellation) | 6 days at equator; shorter at higher latitudes due to orbit geometry |
| Optical revisit (two-satellite constellation) | 5 days (Sentinel-2A + 2B combined) |
| SAR frequency and polarisations | C-band 5.405 GHz; VV and VH dual-polarisation in IW mode |
| Minimum reliably mapped field size | Approximately 1 to 2 hectares at 10 m resolution; smaller parcels subject to mixed-pixel error |
| Typical processing latency | 3 to 7 days from acquisition to classified area layer |
| Archive depth | Sentinel-1 from 2014; Sentinel-2 from 2015; Landsat from 1972 (OLI from 2013) |
| Reported classification accuracy (published literature) | 88 to 94 percent overall accuracy for soybean vs. non-soybean in Brazilian Cerrado studies |
| Delivery formats | GeoTIFF area mask, GeoPackage parcel polygons, CSV area summary by administrative unit or portfolio zone |
Analytics Satellize can run
| Seasonal planted-area polygon layer | Supervised random forest classification on stacked Sentinel-1 VV/VH time series and Sentinel-2 NDVI/red-edge features | GeoTIFF binary mask and GeoPackage polygon layer with area-per-parcel attributes, delivered at season mid-point and close |
| Regional area estimate with uncertainty bounds | Pixel-count aggregation with bootstrap confidence interval based on classification accuracy from held-out validation parcels | CSV table of planted hectares by department, state or custom zone, with 90 percent confidence range |
| Double-crop detection flag | Two-pass seasonal classification comparing first-season (Oct-Jan) and second-season (Feb-Apr) SAR and optical stacks to identify parcels with sequential crop cycles | Polygon layer with single-crop / double-crop attribute; summary table of double-cropped area by region |
| Insured-area verification report | Overlay of classified soybean extent against declared policy polygons; discrepancy flagging where satellite-derived area differs from declared area by more than a configurable threshold | Per-policy PDF report and GIS layer showing confirmed, unconfirmed and flagged parcels |
| Pre-harvest production index | Planted-area estimate combined with in-season NDVI trajectory relative to historical baseline (MODIS or Sentinel-2), following published crop-monitoring protocols | Tabular production index updated fortnightly through the season, formatted for commodity-desk consumption |
| Multi-year planted-area trend series | Consistent annual classification applied to Landsat 8/9 and Sentinel-2 archive, normalised for inter-sensor differences using published cross-calibration coefficients | Annual area time series from 2013 to present, GeoTIFF stack and summary CSV |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.