Remote Sensing Labs
Browse all docs

Classification Basics

A remote sensing image is a grid of measurements, but a map is a grid of meanings, and classification is the step that crosses between the two. It takes the per-pixel values an instrument records and assigns each location to one of a fixed set of surface categories — water, forest, cropland, bare ground, built surface. Where the Spectral Bands concept explains how different surfaces leave different patterns across the wavelengths, classification formalizes the move from those patterns to a named category, producing a thematic map in which every pixel carries a label rather than a raw number.

Classes, labels, and the mapping question

A class is a category of surface the map is meant to distinguish; a label is the class assigned to a particular pixel or object; and a classifier is simply the rule that turns measured values into labels. Everything downstream depends first on the class scheme — the list of categories and their definitions — being chosen deliberately. A good scheme is exhaustive, so every pixel has somewhere to go, and its classes are mutually exclusive, so none is ambiguous; it is also pitched at a level of detail the data can actually support. Asking a coarse sensor to separate two crop species that look nearly identical is asking a question the measurements cannot answer, however sophisticated the classifier.

That is the mapping question: not “what categories would I like” but “what categories can these observations actually tell apart”. It is bounded by the spectral, spatial, and temporal detail available, and by the mixed pixels the Resolution concept describes, where a single cell straddling two surfaces returns a blend that belongs cleanly to neither class. Classification can proceed pixel by pixel, labeling each cell from its own values, or object by object, first grouping neighboring pixels into homogeneous segments and labeling those, which adds features describing shape and context, not just color, at the cost of depending on how well the segments match real objects.

Features: what the classifier reads

A classifier never sees the landscape; it sees a vector of numbers per pixel, and those numbers — the features — are what it actually reasons over. The richest features are spectral: the raw bands and indices built from them, since a normalized index of the kind described under Spectral Bands can compress a signature into one discriminating value that simpler classifiers find easier to use. But surfaces that share a spectral signature can differ in other ways that good feature design exploits. Texture — the local variability of values around a pixel — separates a smooth water body from an equally dark but rougher surface. Terrain features derived from an elevation model, such as slope and aspect, help where land cover follows the landform. And temporal features, the seasonal profile of a pixel across many dates that the Time Series concept develops, separate classes that look alike on any single day but behave differently through the year, as evergreen and deciduous cover do.

Each pixel is thus a point in a feature space, and classification is the task of carving that space into regions belonging to each class. More features can sharpen the separation, but not without limit: features that carry no class information only add noise, and a very high-dimensional space demands more training data to populate. The features must also mean the same thing everywhere in the scene, so calibrated, consistently processed inputs matter as much here as anywhere.

Supervised and unsupervised classification

There are two broad ways to attach labels to those regions of feature space. In supervised classification you supply the answer for a set of known locations — this pixel is forest, that one is water — and the algorithm learns the boundaries that best reproduce your labeling, then applies them to the rest of the scene. You control the class scheme completely, but you must provide representative examples of every class. In unsupervised classification, or clustering, you provide no labels; the algorithm groups pixels into clusters of similar feature values on its own, and only afterwards do you inspect each cluster and assign it a meaning. Here the structure emerges from the data rather than from you, which can reveal distinctions you had not anticipated but may also produce clusters that split or merge the categories you actually cared about.

Neither is simply better. Supervised methods align the map to your intended classes but are only as good as the training data behind them; unsupervised methods need no labeled examples but shift the labor to interpreting the clusters, and their groupings may not match any scheme cleanly. In practice the two are often combined: clustering to explore the natural groupings, then supervised labeling to pin them to a defined scheme.

Training data, sampling, and class balance

For supervised classification the training data is the map’s foundation, and no algorithm recovers from a poor one. Training samples must be representative: a class is rarely a single value but a range — cropland spans many crops, growth stages, and soils — and the samples must span that internal variation, or the classifier learns too narrow a definition and mislabels the rest. They must also be accurate, since a mislabeled training point actively teaches the wrong boundary. Sampling design therefore covers each class across the conditions it takes in the scene, not just the convenient sites where it is obvious.

Class balance is a related trap. When one class dominates the training set and a class of real interest is rare, many classifiers favor the common category and quietly erode the rare one, so the sampling has to give scarce but important classes enough presence to be learned. And the data used to train must be kept separate from the data used to judge the result: evaluating a map on the same points that trained it flatters the accuracy and hides exactly the errors assessment is meant to expose. Separation has to be spatial as well as nominal: points taken from the same field or polygon as training points, or simply very close to them, share their conditions and flatter the result almost as much as reusing them.

Accuracy assessment and uncertainty

A classified map is a hypothesis about the surface, not a record of it, and its quality is unknown until it is checked against independent reference information at a sample of locations chosen by a sound sampling design. The standard summary of that check is the confusion matrix, which cross-tabulates the reference class against the predicted class: correct labels fall on the diagonal, and every off-diagonal cell records a specific confusion between two classes. From it come the headline measures. Overall accuracy is the share of checked locations labeled correctly. Producer’s accuracy takes each reference class and asks how much of it the map caught, exposing omission error — real instances the map missed. User’s accuracy takes each mapped class and asks how much of it is trustworthy, exposing commission error — labels the map claimed but got wrong. A single overall figure can hide a class that is mapped badly, so the per-class view is where the honest picture lives.

Two refinements matter in practice. When the reference sample is stratified — more points in rare classes than their area would give — each cell of the matrix has to be weighted by its class’s share of the map before accuracies are computed, or they are biased. And the area of each class is better estimated from the reference sample, with a confidence interval, than by counting map pixels, since pixel counts carry every omission and commission error straight into the total. The kappa coefficient, still widely reported, adds nothing reliable to these measures and is best left out.

Uncertainty runs deeper than one accuracy number. Many classifiers can report a per-pixel confidence or class probability — often poorly calibrated, and never an accuracy in itself — and the least certain pixels cluster predictably — at class boundaries, over mixed cells, and wherever two classes overlap in feature space. Because a classified map so often becomes the input to something else — the post-classification comparison the Change Detection concept describes inherits the error of every map it compares — carrying that uncertainty forward rather than treating the labels as certain is what keeps later conclusions honest. Read this way, classification is less a single act of labeling than a disciplined chain of judgments — which classes, which features, which examples, and how well it was checked — and the map is only ever as trustworthy as the weakest link in that chain.

Sources