Supervised vs unsupervised classification for a national land cover layer

Every national land cover programme eventually has this argument in a meeting room: do we let the algorithm find its own clusters, or do we feed it labelled training samples and tell it what to look for. The answer shapes your methodology note for the next decade, so procurement is the time to settle it, while the nomenclature and accuracy targets are still open for negotiation.

What the two approaches actually do

Unsupervised classification runs a clustering algorithm (k-means and ISODATA are the two you'll see most) over the spectral bands and groups pixels by statistical similarity, with no labels attached. You get clusters, then an analyst sits down and assigns each one a class: this cluster is cropland, that one is bare ground, this other one is a mix of both and needs splitting. It's fast to run and doesn't need a training archive, which is why it shows up in first-pass pilots and in regions with no existing ground-truth data to draw on. The cost shows up later, in the labelling step, where cluster boundaries rarely line up cleanly with the classes your statistics office reports against.

Supervised classification flips the order. An analyst, or an existing reference dataset, labels a sample of pixels first: this parcel is forest, that one is built-up, this one is water. The algorithm, whether it's a random forest, support vector machine, or a CNN running on the imagery, learns the spectral signature of each class from those samples and applies it across the full scene. It needs more upfront work, but the output classes are the ones you defined, which matters enormously when the layer has to match a legal nomenclature like CORINE or an equivalent national schema.

For a statutory base layer that gets cited by environmental reporting, agricultural statistics and planning departments at once, supervised classification is almost always the right default. The classes aren't up for reinterpretation each year. Cropland has to mean the same thing in year six as it did in year one, and an algorithm trained against a stable set of labelled classes gives you that consistency far more reliably than an unsupervised run that gets relabelled by whoever's on shift.

Training samples: the part that determines accuracy

The algorithm choice gets most of the attention, but the training sample set is where accuracy is won or lost. A random forest trained on 200 well-distributed, field-verified samples per class will outperform a CNN trained on 50 samples pulled from one corner of the country. A few things matter in practice:

Spatial spread. Samples clustered around one region train a classifier that overfits to that region's soil colour, crop calendar and atmospheric conditions, then underperforms everywhere else.

Class balance. Built-up and water are usually easy to over-sample because they're visually distinct and easy to label. Bare ground and transitional cropland are harder and get neglected, which is exactly where the classifier then struggles.

Temporal match. A training sample labelled from last year's imagery can mislabel this year's scene if a field was fallow, flooded, or recently cleared. Samples need refreshing against the current acquisition, not carried forward indefinitely.

Independent validation set. Samples used to train the classifier shouldn't be the same ones used to check its output. This sounds obvious and still gets skipped under deadline pressure, and it's the single fastest way to produce an accuracy figure nobody can defend in an audit.

Picking a pixel classifier for a layer that has to last

Random forests remain the workhorse for national programmes: they handle multispectral bands well, they're not fussy about sample size, and they're straightforward to retrain each year as new samples come in. Support vector machines perform comparably on smaller sample sets but scale less gracefully to continent-size scenes. Deep learning models can edge out both on complex scenes with enough labelled data behind them, but they demand a training archive most agencies don't have on day one, and a black-box model is a harder thing to explain to an audit committee than a random forest's feature importances.

The practical answer for most agencies: start with a supervised, random-forest-based classifier against a labelled schema you control, and treat the training sample library as a living asset that gets added to every acquisition cycle rather than rebuilt from scratch.

If your agency is weighing up classifier choice and training sample strategy for next year's edition, see how Land Use Mapping turns an annual acquisition into one classified layer your statistics office, environmental reporting and planning teams can all cite from the same source. Get in touch if you'd like to talk through where your current workflow sits against this.