Question 3 of 5: Supervised versus Unsupervised Classification

Question 3: Supervised versus Unsupervised Classification (20 marks)

Question text not reproduced: the examination questions are © Engineers and Geoscientists BC. Open the official past paper (linked at the top of this page) to read the question, then follow the worked solution below.

Definitions

Supervised classification is a classification strategy in which the analyst supplies prior knowledge of the scene by delineating representative training samples for each desired information class (e.g. water, forest, urban). The algorithm estimates the spectral signature (mean vector, and for maximum-likelihood the covariance matrix) of each class from these training pixels and then labels every image pixel by the decision rule — minimum-distance-to-means, maximum-likelihood, parallelepiped, and so on. The classes are defined before the machine runs; the analyst directs (supervises) the process.

Unsupervised classification reverses that order. The algorithm (for example ISODATA or $k$-means) is given only the imagery and a target number of clusters, and it groups the pixels into natural spectral clusters purely from the statistics of the data, with no training samples. The analyst intervenes only afterward, assigning an information-class label to each spectral cluster by comparison with reference data. The machine discovers the groupings; the human interprets them.

Supervised classification — advantages and disadvantages

Advantages. The output categories are the specific information classes the analyst wants, so the map is directly meaningful and its classes are known and labelled from the start. The analyst controls class definitions and can detect gross errors (e.g. a training site producing an implausible signature). Where good ground reference exists, accuracy is generally high and the classes correspond to recognizable cover types.

Disadvantages. It demands reliable prior knowledge and representative training data, which are costly and time-consuming to collect and require analyst familiarity with the area. Training sites that are unrepresentative, mislabelled, or that do not capture a class's full spectral variability degrade the result, and spectral classes the analyst did not think to train on may be forced (mis-assigned) into an existing category. It is also labour-intensive up front.

Unsupervised classification — advantages and disadvantages

Advantages. It requires no prior training data and little foreknowledge of the scene, so it is fast to launch and ideal for reconnaissance or unfamiliar areas. It is objective — free of the analyst's up-front labelling bias — and it reveals the natural spectral structure of the data, often exposing distinct classes (or sub-classes) the analyst would not have anticipated. The spectral clusters produced are, by construction, internally homogeneous.

Disadvantages. The clusters are spectral, not informational: they still must be labelled afterward, and a single land-cover type may split across several clusters while two different cover types with similar spectra may merge into one — both awkward to interpret. The analyst has little control over which classes emerge, results depend on the chosen number of clusters and algorithm parameters, and the labelling/merging step can be as laborious as collecting training data would have been.

In practice the two are complementary — a common workflow runs an unsupervised pass to reveal the scene's spectral structure and inform the choice of training sites for a subsequent supervised classification.