Crop type classification from multispectral imagery
Multispectral time-series imagery lets analysts distinguish wheat from maize from soya by tracking how each crop's reflectance evolves through the growing season. The method is proven but demands dense, cloud-free observations and reliable ground truth.
Sensors
- Sentinel-2 MSI (ESA): 13 spectral bands from 443 nm to 2190 nm; 10 m resolution in visible and NIR, 20 m in red-edge and SWIR. Five-day revisit at the equator with both satellites. The three red-edge bands (705, 740, 783 nm) are particularly diagnostic for separating crops at similar canopy development stages. Free and open archive from 2015.
- Planet SuperDove: Eight bands including two red-edge channels; 3–4 m ground resolution; near-daily revisit globally. The high spatial resolution resolves small field parcels (under 0.5 ha) that Sentinel-2 pixels straddle. Commercial licence required; latency typically under 24 hours.
- Landsat 8/9 OLI: Seven reflective bands at 30 m resolution; 16-day revisit per satellite, 8-day combined. Longer archive (Landsat 7 back to 1999, Landsat 8 from 2013) supports multi-year phenological baselines. SWIR bands (1565–1651 nm and 2107–2294 nm) help distinguish crop residue from bare soil between seasons.
- PRISMA (ASI): Hyperspectral imager covering 400–2500 nm in 239 contiguous bands at roughly 30 m resolution. Single-pass revisit (around 29 days at mid-latitudes without tasking). Resolves subtle biochemical differences between species that broadband sensors conflate, at the cost of much larger data volumes and a shallower archive.
Why the calendar matters as much as the spectrum
A single multispectral image taken in mid-July cannot reliably distinguish winter wheat from spring barley: both appear as dry, yellowing canopies at that moment. The discriminating power comes from watching the full phenological arc. Winter wheat greens up in March, peaks in May, and is harvested by July in temperate Europe. Maize stays bare soil until May, builds canopy through June and July, and senesces in September. Those trajectories, expressed as time-series of indices such as NDVI, the red-edge chlorophyll index, and SWIR-based moisture indices, form crop-specific signatures that supervised classifiers can learn.
Dense time-series are therefore the foundation of the method. Sentinel-2's five-day revisit is the practical minimum for capturing rapid phenological transitions; cloud cover in humid regions can easily remove half of all acquisitions, so analysts typically need 18 to 30 usable observations across a growing season to achieve stable classification. In persistently cloudy regions, optical-only classification remains genuinely difficult, and SAR fusion is required (covered in the companion page on yield estimation).
What the spectral bands actually measure
Visible red reflectance (around 665 nm) is suppressed by chlorophyll absorption; NIR reflectance (around 842 nm) is driven by leaf cell structure. The ratio of the two produces NDVI, which tracks green biomass. But NDVI saturates at moderate canopy densities, making it a poor discriminator between dense-canopy crops like maize and soya at peak growth.
The red-edge region (700–740 nm) is more informative at that stage. Chlorophyll absorption falls off steeply across this range, and the slope of reflectance through it correlates with chlorophyll content per unit area rather than just canopy closure. Sentinel-2's three red-edge bands, and Planet SuperDove's two, exploit this. SWIR bands add a separate axis: water content in the canopy and, after harvest, the lignin and cellulose content of residue. A classifier using red-edge and SWIR features alongside NIR can separate crop types that look identical in a standard false-colour composite.
Classifiers, ground truth, and the limits of transferability
Random forest classifiers trained on field-survey labels remain the most widely used approach, partly because they handle the high dimensionality of multi-temporal feature stacks without requiring enormous training sets. Deep learning methods, particularly recurrent neural networks applied to pixel-level time series, have shown higher accuracy in published studies where training data is abundant, but they transfer poorly across agro-ecological zones without retraining.
Ground truth is the binding constraint. A classifier trained on labelled parcels in the Po Valley will perform poorly in the Nile Delta without local labels, because planting calendars, variety choices, and irrigation practices differ. Published overall accuracies for well-resourced national mapping programmes (France's Registre Parcellaire Graphique, for instance) reach 85–95% for major crops. In data-sparse regions, 70–80% is a more honest expectation for the first season, improving as ground-truth accumulates. Minimum mappable parcel size is roughly 2–3 times the pixel area: at Sentinel-2's 10 m, that means parcels below about 0.03 ha are unreliable.
Hyperspectral sensing: more bands, narrower use case
PRISMA's 239 contiguous bands allow detection of specific pigments, water stress indicators, and even some fungal metabolites that broadband sensors cannot resolve. For crop type classification specifically, the marginal gain over a well-designed multispectral time series is modest for common species. Hyperspectral data earns its keep when the task is separating closely related varieties (durum versus bread wheat, for instance), detecting minor crops with unusual biochemistry, or building spectral libraries for later use with cheaper sensors.
The practical constraints are real. PRISMA's revisit without tasking is around 29 days, which is too coarse to capture rapid phenological transitions reliably. Data volumes are large and processing pipelines more complex. The sensor is best used in targeted campaigns over areas of specific interest rather than as the primary source for regional mapping.
From classified pixels to a usable crop map
Raw per-pixel classification outputs are noisy. Standard post-processing applies spatial smoothing within cadastral parcel boundaries (where parcel data exists) or object-based segmentation (where it does not), assigning a single dominant class to each field polygon. This removes the salt-and-pepper artefacts that confuse downstream users and aligns the output with how farmers and ministries actually think about area statistics.
Accuracy assessment requires a held-out validation set drawn from the same season and region, not from a prior year. Crop rotations mean that a field labelled as maize in 2022 may be soya in 2023; using stale labels inflates apparent accuracy. Satellize's crop-estimation work in Tonga applies this discipline to a small-island context where field sizes are sub-hectare and cloud frequency is high, which makes the classification problem harder than it appears in temperate case studies. The output of a classification run is a GIS polygon layer with per-parcel crop type, confidence score, and area estimate, suitable for ingestion into national agricultural statistics systems.
One honest caveat worth stating plainly: crop type classification tells you what is planted, not how well it is growing or what yield to expect. Those questions require separate analytical steps, covered in adjacent pages.
Typical figures
| Typical spatial resolution | 10 m (Sentinel-2 VIS/NIR), 20 m (Sentinel-2 red-edge/SWIR), 30 m (Landsat OLI), 3–4 m (Planet SuperDove) |
| Revisit frequency | 5 days (Sentinel-2 combined), 8 days (Landsat 8+9 combined), near-daily (Planet SuperDove) |
| Spectral bands used | Blue, Green, Red, Red-edge (×2–3), NIR, Narrow NIR, SWIR-1, SWIR-2; 8–13 bands depending on sensor |
| Minimum mappable parcel size | ~0.03 ha at 10 m resolution; ~0.5 ha at 30 m resolution (rule of thumb: 2–3 pixels across shortest dimension) |
| Typical classification accuracy (major crops, good ground truth) | 85–95% overall accuracy in well-resourced programmes; 70–80% in data-sparse first seasons |
| Observations needed per season | 18–30 cloud-free acquisitions for stable time-series classification in temperate zones |
| Archive depth | Sentinel-2 from 2015; Landsat OLI from 2013; Landsat 7 ETM+ from 1999 (30 m, 6 reflective bands) |
| Delivery format | GeoTIFF raster (classified), GeoPackage or Shapefile polygon layer with crop-type attribute and confidence score, CSV area summary |
| Latency (open data path) | Sentinel-2 L2A available within 3–5 days of acquisition via Copernicus Data Space; Landsat Collection 2 within 24 hours |
| Cloud cover impact | Pixels under cloud masked; classification quality degrades if fewer than ~12 usable observations are available across the season |
Analytics Satellize can run
| Seasonal crop type map | Random forest or LSTM classifier trained on multi-temporal spectral feature stack (NDVI, red-edge CI, SWIR indices) with field-survey ground truth | GIS polygon layer with dominant crop class, confidence score, and area per parcel; updated at end of growing season |
| Crop area statistics by administrative unit | Pixel-counting within classified raster, aggregated to district or national boundaries; uncertainty bounds from validation confusion matrix | CSV or Excel table of crop areas by region and class, with 90% confidence intervals |
| Phenological calendar extraction | Time-series smoothing (Savitzky-Golay or TIMESAT) applied to NDVI or EVI stack to extract green-up date, peak, and senescence per parcel | GIS layer of per-parcel phenological metrics; useful for planting-date verification and variety discrimination |
| Year-on-year crop rotation map | Multi-year stack of seasonal classifications; transition matrix computed per parcel to identify rotation patterns | GIS layer showing rotation class (e.g. continuous maize, wheat-soya rotation); annual update |
| Classification confidence and uncertainty layer | Out-of-bag error from random forest or Monte Carlo dropout from neural classifier; expressed as per-pixel posterior probability | Raster of class probability for top-1 and top-2 predicted classes; flags low-confidence parcels for field verification |
| Training-data gap assessment | Spatial coverage analysis of existing ground-truth labels against agro-ecological zones; identifies under-sampled crop types and regions | Field-survey prioritisation report with map of recommended sampling locations for next season |
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.