Unsupervised Classification

Unsupervised classification is one of the two basic approaches to digital image classification with the goal of producing land cover maps from remotely sensed data. The other is supervised classification. Unsupervised algorithms evaluate the spectral properties of image pixels and segregate them into naturally occurring statistical groups with little or no guidance from the analyst. The underlying assumption is that a remotely sensed landscape comprises a mosaic of land cover classes, each of which is relatively uniform in terms of spectral pattern. Following image classification, a significant challenge for the analyst is to apply meaningful land cover labels to the spectral classes identified. This entry summarizes the unsupervised classification logic, briefly examines two commonly used algorithms, and discusses the primary advantage and disadvantage of the approach.

Unsupervised classifiers are based on the multi-spectral data space location of spectral class means, the measurement of distance between class means and individual pixel locations, and the establishment of boundaries delineating the individuality of spectral classes. The amount of analyst-supplied input at the onset of the unsupervised classification varies by algorithm. In general, though, the algorithms work by assigning a number of image pixels, either systematically or at random, as initial class means. The distance between individual image pixel locations and the class means is calculated, and pixels are assigned to the nearest class mean. Once the entire image has been evaluated, new class means are calculated based on the addition of pixels to the initial mean value. The process then repeats, or iterates, [Page 2915]until there is a statistically insignificant change in the location of class means or in the number of pixels that switch class membership or until a maximum number of iterations is reached.

The two most commonly used unsupervised classification algorithms are K-means and the Iterative Self-Organizing Data Analysis Technique, or ISODATA. Both are iterative processes that, in general, operate as described above. K-means, however, requires the analyst to specify a fixed number of desired classes and then minimizes the within-class variability by optimizing the sum-of-squares error between individual class means and member pixels. ISODATA is more flexible, in that spectral classes can be merged or split. Classes are merged when two or more class means are located in close proximity. A class is split when its standard deviation exceeds a threshold value.

Relative to supervised approaches, a primary advantage of unsupervised classification is that the potential for human error is reduced. That is because there is less input required from the analyst and because of the high likelihood that each unique spectral class in an image will be identified. Unlike supervised classification, no prior knowledge of the study area is required to initiate an unsupervised process. However, detailed knowledge of the study area is needed to accurately label the resultant spectral classes. The primary disadvantage is that matching spectral classes to the desired set of land cover classes can be difficult.