National-scale land-cover classification

Wall-to-wall land-cover maps built from multi-seasonal satellite composites are the baseline every forest, agriculture and planning policy rests on. Getting the class definitions, sensor mix and accuracy reporting right is harder than the imagery makes it look.

Sensors

Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands, 20 m in red-edge and shortwave-infrared. 5-day revisit at the equator with both satellites. The 13-band stack is the primary optical input for supervised classifiers; multi-seasonal compositing suppresses cloud and phenological noise.
Sentinel-1 SAR (C-band): 6-day revisit, 10 m pixel spacing in Interferometric Wide Swath mode. Cloud-transparent and sun-independent. VV and VH backscatter distinguish flooded surfaces, dense canopy and bare soil in ways optical data cannot, particularly during monsoon seasons when optical revisit collapses to near zero.
Landsat 8/9 OLI: 30 m multispectral, 16-day revisit per satellite (8-day combined). The Landsat archive back to 1972 enables change baselines; the thermal band (100 m) adds surface-temperature context useful for separating irrigated cropland from dryland. Free and open since 2008.
MODIS Terra/Aqua: 250 m to 500 m resolution, daily global coverage. Too coarse for parcel-level work but useful for generating seasonality metrics and phenological indices across an entire country quickly, which then inform the training strategy for finer-resolution classifiers.

Why a map is only as good as its legend

Before a single image is processed, the most consequential decision in any national land-cover programme is the class legend. 'Forest' means different things to a carbon accountant, a hydrologist and a timber concession manager. The FAO Land Cover Classification System (LCCS) provides a hierarchical, parametric framework that separates land-cover facts (what is physically on the ground) from land-use interpretation (what people do with it). Anchoring a national map to LCCS makes it comparable across years and across borders, which matters the moment a government needs to report to the UNFCCC or cross-validate with a neighbour's dataset.

The practical consequence is that more thematic classes demand more training samples, more spectral separability and, inevitably, lower overall accuracy for a given sensor. A six-class map distinguishing closed forest, open woodland, shrubland, cropland, built-up and bare ground is achievable at high accuracy with Sentinel-2. Push to fifteen classes and the confusion between, say, degraded open woodland and shrubby fallow becomes a genuine problem that no classifier resolves without very dense, well-distributed field data.

What multi-seasonal compositing actually does

A single cloud-free Sentinel-2 scene is not a land-cover map. Cropland in the dry season looks spectrally similar to bare soil; a water body at low stage can resemble scrub. The standard approach is to build seasonal composites, typically four per year aligned with phenological quarters, by selecting the least-cloudy or median-reflectance pixel from all acquisitions in each window. This produces a stack of 52 or more spectral-temporal features per pixel, which a Random Forest or gradient-boosted classifier can use to distinguish classes that are ambiguous in any single date.

Cloud cover is the honest constraint. In persistently cloudy regions such as the Congo Basin or maritime Southeast Asia, optical compositing over a 90-day window may still leave data gaps. That is where Sentinel-1 SAR backscatter earns its place: C-band VH backscatter correlates with canopy volume and changes detectably between forest, shrub and open land regardless of cloud. Studies published in Remote Sensing and similar journals have demonstrated that fusing Sentinel-1 and Sentinel-2 features consistently improves overall accuracy by several percentage points compared with optical data alone, particularly for wetland and flooded-forest classes.

Supervised classification: the training-data problem

Random Forest remains the most widely used classifier for national land-cover work, partly because it handles mixed feature types (reflectance, backscatter, indices, elevation derivatives) without normalisation, and partly because its out-of-bag error gives an honest internal accuracy estimate. Support Vector Machines and, increasingly, convolutional neural networks applied to image chips are competitive alternatives, but they demand more tuning and larger labelled datasets.

The training-data bottleneck is real. A national programme covering, say, 500,000 km² needs thousands of geographically distributed, seasonally appropriate reference points that are correctly labelled. Collecting those points from field campaigns is expensive. The pragmatic alternative is a stratified random sample drawn from a pre-existing coarser map (MODIS Land Cover, ESA WorldCover, or a prior national inventory), supplemented by expert visual interpretation of high-resolution imagery. The risk is that errors in the reference dataset propagate into the classifier. Semi-supervised methods, which use a small labelled set to guide clustering of the unlabelled majority, reduce this dependency but introduce their own ambiguity in class assignment.

Accuracy reporting must follow the Olofsson et al. (2014) framework, or an equivalent, to be credible: area-weighted producer's and user's accuracy per class, overall accuracy, and confidence intervals from a probability sample. A headline figure of '92% overall accuracy' is nearly meaningless without knowing the class breakdown and sampling design.

The resolution floor and what it hides

At 10 m, Sentinel-2 resolves features down to roughly one-quarter of a hectare. That sounds fine until you consider that smallholder farm plots across much of sub-Saharan Africa and South Asia average between 0.5 and 2 hectares, and that mixed pixels at field boundaries systematically confuse classifiers. At 30 m (Landsat), the problem is worse: a significant fraction of pixels in fragmented agricultural landscapes contain two or more cover types, and the classifier assigns a single label to a statistical mixture.

This is not a reason to avoid Landsat; its archive depth and thermal band are genuinely useful. It is a reason to be explicit about the minimum mapping unit in the product specification. A national map at 30 m resolution with a minimum mapping unit of 1 hectare is an honest product. Claiming parcel-level accuracy from the same data is not.

Validation, uncertainty and what a government can actually act on

A land-cover map delivered without an independent validation dataset and a per-class confusion matrix is not a finished product. It is a draft. Independent validation means reference samples that were not used in training, collected by a different team or from a different source, and distributed according to a probability sampling design rather than convenience.

National governments using land-cover maps for forest policy, agricultural planning or infrastructure licensing need to understand where the map is reliable and where it is not. Delivering a per-pixel confidence layer alongside the classification, derived from classifier posterior probabilities or ensemble variance, gives decision-makers a spatial guide to where field verification is worth the cost. Satellize incorporates this kind of uncertainty quantification into its analytics outputs; the Tonga crop-estimation programme, for instance, required explicit confidence bounds on area estimates to satisfy the client's reporting obligations.

The practical ceiling for a well-executed national programme using open data is roughly 85 to 92 percent overall accuracy for a six-to-eight-class legend, depending on landscape fragmentation and cloud climatology. Pushing beyond that requires either commercial very-high-resolution imagery for training data, intensive field campaigns, or both.

Archive depth and the change-detection dividend

A single-epoch land-cover map answers 'what is here now'. A time series answers 'what changed, when, and at what rate'. The Landsat archive from 1972 and the Sentinel-2 archive from 2015 together give national programmes a multi-decadal baseline at no data cost. Consistent classification methodology applied across multiple epochs produces change matrices that are far more policy-relevant than any single snapshot: net forest loss, cropland expansion into former woodland, urban growth rates.

The discipline required is methodological consistency. A classifier trained on 2024 composites and applied to 2015 imagery without accounting for sensor differences, atmospheric correction changes or class-definition drift will produce spurious change signals. Cross-calibration between Landsat and Sentinel-2 is well-documented and tractable, but it requires deliberate effort. Skipping it is the most common source of artefacts in multi-temporal national land-cover products.

Typical figures

Primary optical resolution	10 m (Sentinel-2 visible/NIR), 20 m (Sentinel-2 SWIR/red-edge), 30 m (Landsat 8/9)
SAR resolution	10 m pixel spacing, Sentinel-1 IW mode; 20 m nominal ground resolution
Revisit (optical, equatorial)	5 days (Sentinel-2 combined); 8 days (Landsat 8+9 combined)
Revisit (SAR)	6 days (Sentinel-1 per track)
Spectral bands used	Sentinel-2: 13 bands (443–2190 nm); Landsat OLI: 7 bands (430–2290 nm) plus thermal; Sentinel-1: C-band (5.4 GHz), VV and VH polarisations
Minimum mapping unit (typical)	0.25 ha at 10 m; 1 ha at 30 m
Archive depth	Sentinel-2: 2015–present; Landsat: 1972–present; MODIS: 2000–present
Typical overall accuracy (6–8 class legend)	85–92%, dependent on landscape fragmentation and cloud climatology
Delivery formats	Cloud-optimised GeoTIFF, GeoPackage, OGC WMS/WFS, per-class confidence raster
Coverage	Global; no tasking required for open constellations

Analytics Satellize can run

Wall-to-wall national land-cover map	Random Forest classifier trained on multi-seasonal Sentinel-2 spectral-temporal feature stack, FAO LCCS-aligned legend	GeoTIFF classification layer with per-pixel confidence band and per-class accuracy report
SAR-optical fused classification for cloud-prone regions	Feature-level fusion of Sentinel-1 VV/VH backscatter with Sentinel-2 composites; gradient-boosted classifier	Improved wetland and flooded-forest class layers with reduced cloud-gap artefacts
Multi-epoch change matrix	Consistent classifier applied to cross-calibrated Landsat and Sentinel-2 annual composites; Olofsson-framework area estimation	Tabular change matrix with confidence intervals and GIS change-polygon layer
Training and validation sample design	Stratified random sampling from coarse prior map; expert visual interpretation of high-resolution imagery	Labelled point dataset with sampling weights for unbiased area estimation
Phenological feature stack	MODIS NDVI and EVI time-series decomposition to extract seasonality metrics as classifier inputs	Raster stack of amplitude, phase and greenness metrics at 250–500 m, resampled for training guidance
Accuracy and uncertainty report	Probability-sample validation against independent reference data; confusion matrix per class; confidence interval estimation	PDF technical report and machine-readable confusion matrix meeting UNFCCC reporting standards

Sources

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.