Commercial optical time-series crop type classification
Crop types diverge in their seasonal spectral trajectories far more than in any single snapshot. Dense time-series from Planet Dove and Sentinel-2, classified with machine-learning models trained on field labels, can map what is growing and where.
Sensors
- Planet Dove / SuperDove: 3-metre ground sampling distance, daily revisit globally. SuperDove adds eight spectral bands including red-edge at 705 nm, improving separation of chlorophyll-rich canopies. High revisit is the primary advantage in cloudy tropical environments where Sentinel-2's 5-day repeat collapses to far fewer clear observations per season.
- Sentinel-2 MSI: 10-metre resolution in four visible and near-infrared bands, 20-metre in six red-edge and SWIR bands. Five-day repeat at the equator with both satellites. Free and open archive from 2015. The SWIR bands at 1610 nm and 2190 nm help distinguish bare soil from senescent vegetation and separate some crop types that look identical in the visible.
- Landsat 8/9 OLI: 30-metre resolution, 16-day repeat per satellite (8-day combined). Useful for extending time series back to 1972 via earlier Landsat missions, and for cross-calibrating with Sentinel-2. Resolution is often too coarse to resolve smallholder plot boundaries below roughly one hectare.
- Airbus Pléiades Neo: 30-centimetre panchromatic, 1.2-metre multispectral. Tasked on demand. Not suited to time-series classification due to cost and narrow swath, but valuable for generating training labels and validating classifier outputs at field boundaries that Sentinel-2 cannot resolve.
Why a single date misleads
Maize and sorghum at peak greenness are nearly indistinguishable in the visible and near-infrared. So are young rice paddies and flooded grassland. A classifier trained on a single midsummer image will confuse them reliably. The spectral-temporal trajectory is what separates them: maize in West Africa reaches canopy closure roughly six weeks after sowing and senesces sharply; sorghum peaks later and holds its greenness longer before harvest. Rice shows a flooded-field signature in SWIR bands early in the season that upland cereals do not. These timing differences, measured across dozens of cloud-free observations, are far more discriminating than any single spectral snapshot.
This is the core argument for time-series classification. The phenological trajectory of a crop, its path through spectral space across a growing season, is a fingerprint. Dense imagery makes that fingerprint legible.
Revisit frequency is not a luxury in the tropics
In temperate Europe, Sentinel-2's 5-day repeat typically yields 20 to 30 usable cloud-free observations per growing season. In humid tropical environments, the same 5-day repeat might yield four or five. That is often insufficient to reconstruct a reliable phenological curve. Planet's Dove constellation, which images every land point on Earth daily at 3 metres, changes the arithmetic substantially. Even at 50 percent cloud cover, daily imaging produces roughly 180 candidate observations per season, of which a statistically adequate sample will be clear.
The practical minimum for resolving a tropical crop calendar through cloud is generally cited in the published literature as one clear observation per phenological stage, with at least four stages (emergence, vegetative peak, grain fill, senescence) needing representation. Daily revisit makes that achievable. Sentinel-2 alone frequently does not.
How classifiers are trained and where they fail
Random forest classifiers remain the standard workhorse for crop type mapping. They accept a feature stack, typically a time series of vegetation indices and raw band reflectances sampled at fixed intervals across the season, and output a per-pixel class probability. Training requires georeferenced field survey labels: plots of known crop type, ideally collected in the same season as the imagery. The classifier learns which spectral-temporal patterns correspond to which crop. Published studies using Sentinel-2 and random forests in sub-Saharan Africa report overall accuracies of 75 to 90 percent for dominant crop types, with performance degrading sharply for minority classes and for crops grown on plots smaller than two or three pixels.
Recurrent neural networks, particularly long short-term memory (LSTM) architectures, treat the time series as a sequence rather than a flat feature vector. They can in principle capture temporal dependencies that random forests miss. In practice, they require larger training datasets and are more sensitive to irregular observation gaps caused by cloud. Neither approach resolves the fundamental confusion between spectrally similar crops at coarse resolution. Maize and sorghum remain a persistent problem at Sentinel-2's 10 metres; separating them reliably generally requires either higher-resolution imagery, additional SWIR observations, or plot-level ground truth dense enough to anchor the classifier at field boundaries.
Transfer learning, training a model in one region and applying it in another, degrades predictably when crop calendars or varieties differ. A classifier trained on Kenyan maize phenology will not transfer cleanly to Tongan root crops without retraining on local labels.
The Tonga programme: small islands, fragmented plots
Satellize's crop-estimation programme in the Kingdom of Tonga illustrates the practical constraints of applying these methods in a smallholder island context. Tonga's agricultural plots are small, often below half a hectare, which means Sentinel-2's 10-metre pixels frequently straddle plot boundaries and mix crop signatures. Planet Dove's 3-metre imagery resolves individual plots more cleanly, reducing mixed-pixel contamination in the training data. The island geography also limits the volume of available field survey labels, which constrains classifier complexity and pushes the workflow toward simpler models with regularisation rather than deep networks that would overfit on sparse ground truth.
Honest limits of the method
Cloud is the dominant operational constraint. No amount of revisit frequency recovers information from a pixel that is cloud-covered at every critical phenological moment. In persistently overcast regions, gap-filling using SAR backscatter as a proxy for vegetation state is an active research area but not yet operationally mature for crop type classification.
Spatial resolution imposes a hard floor on the minimum mappable plot size. At 10 metres, plots below roughly 0.05 hectares are effectively unresolvable. At 3 metres, that floor drops to around 0.003 hectares, but Planet Dove's radiometric calibration and band configuration differ from Sentinel-2, so models trained on one cannot be applied to the other without cross-calibration. Archive depth matters for multi-year analysis: Planet's commercial archive extends to roughly 2016 at useful global density, whereas Sentinel-2 runs from 2015 and Landsat from 1972, though at coarser resolution. Accuracy figures from published studies are also geographically specific; they should not be assumed to transfer to a new region without local validation.
From classified map to decision-ready product
A crop type map is an intermediate product, not an answer. The questions that drive procurement are typically about area under cultivation, expected yield, or import dependency. Crop type classification feeds area estimation, which combined with yield models, whether statistical, agronomic, or satellite-derived, produces the numbers that matter to a ministry of agriculture or a commodity trader. The classification layer also enables change detection year on year: which fields shifted from cassava to maize, and where is fallow land being brought into production.
Delivering this as a GIS layer in a standard format such as GeoTIFF or a vector polygon file with per-field class probabilities allows downstream integration with national agricultural statistics systems. Confidence intervals on area estimates, derived from the classifier's per-pixel probability outputs and the sampling design of the validation set, are as important as the point estimate itself.
Typical figures
| Spatial resolution (primary) | 3 m (Planet Dove / SuperDove); 10 m visible/NIR, 20 m red-edge/SWIR (Sentinel-2) |
| Revisit frequency | Daily (Planet Dove, global); 5-day equatorial (Sentinel-2, both satellites combined); 8-day (Landsat 8+9 combined) |
| Spectral bands used | Blue, green, red, red-edge (~705 nm, ~740 nm), NIR (~842 nm), SWIR (~1610 nm, ~2190 nm); SuperDove adds coastal blue and yellow bands |
| Minimum mappable plot size | ~0.003 ha at 3 m; ~0.05 ha at 10 m; ~0.3 ha at 30 m |
| Minimum clear observations per season (tropical) | 4 per phenological stage recommended; daily revisit required to achieve this under persistent cloud |
| Typical classification accuracy (dominant crops) | 75–90% overall accuracy reported in published studies for major crop types; lower for minority classes and sub-pixel plots |
| Archive depth | Planet Dove: ~2016 at useful global density; Sentinel-2: 2015–present; Landsat: 1972–present (30 m) |
| Delivery formats | GeoTIFF (per-class probability raster), vector polygons with class and confidence attributes, tabular area estimates with confidence intervals |
| Latency (operational season mapping) | End-of-season map: days to weeks after harvest depending on cloud clearance and processing pipeline; in-season updates possible at lower accuracy |
Analytics Satellize can run
| Crop type map | Random forest classifier trained on field survey labels, feature stack of multi-date band reflectances and vegetation indices | GeoTIFF raster with per-pixel class probability and dominant class label; vector polygon layer with per-field classification |
| Crop area estimates with uncertainty | Design-based area estimation using classifier output and stratified random sample validation; confidence intervals from sampling theory | Tabular report of area by crop type, region and confidence interval, formatted for integration with national agricultural statistics |
| In-season crop type monitoring | Partial-season LSTM classifier using observations accumulated to date; updated iteratively as new clear imagery arrives | Periodic GIS layer updates through the growing season; accuracy statement qualifying the partial-season estimate |
| Year-on-year crop rotation and land-use change | Multi-year time-series comparison of classified maps; change matrix quantifying transitions between crop types and fallow | Change detection GeoTIFF and tabular transition matrix; flagged fields showing unexpected rotation or abandonment |
| Training label quality assessment | Spectral-temporal separability analysis (Jeffries-Matusita distance or similar) applied to field survey label sets before classifier training | Report identifying which crop pairs are separable in the available imagery and which require additional field collection or finer resolution data |
| Smallholder plot boundary delineation | Object-based image analysis or deep-learning segmentation on Planet 3 m or Pléiades Neo imagery to define field units prior to classification | Vector polygon layer of delineated field boundaries for use as classification units rather than raw pixels |
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.