Classify

Functionality

The Classify operation performs a multi-spectral image classification according to training pixels in a sample set. Before classification, a sample set thus has to be prepared with Sample.

The following classification methods are available:

As Sample was the training phase, where classes of pixels with similar spectral values are defined, Classify is the decision phase, where each output pixel is assigned a class name if the spectral values of that pixel are similar enough to a training class; if this is not the case, an output pixel may be assigned the undefined value. As Classify uses the training pixels selected by the user, Classify is a supervised classification.

A supervised classification foremost depends on the spectral values of the pixels that you selected to serve as training pixels in Sample. Relevant information on the classes for which training pixels have been selected in the sample set, can be viewed in the Sample Statistics.

The manner in which the spectral values of input pixels are compared to the known values and statistics of the training pixels, to decide on the class that should be assigned to a pixel, furthermore depends on the classification method that you choose, and the parameters you use for that method. In general, to each output pixel, the class will be assigned of which the spectral values are most similar to (or 'nearest') to the spectral values of an input pixel.

The bands that you wish to classify should be combined in a map list. This map list is part of the required input sample set. There is no limitation for the number of bands that can be classified.

Training pixels should be assigned a class name; this is done in Sample. There is no limitation for the number of classes that a sample set can contain.

Tip:

During Sampling, you can already create some dependent output maps with the Classify operation (various methods, various parameters) using the current sample set as input. To create a dependent output map, click the Define button in the Classify dialog box; you will just create the definition for an output map. In this way various dependent output maps may exist only by their object definition file and are not yet calculated and stored on disk.

If you then add these dependent maps to a pixel info window, you can already see the results of classifying while still busy with the sampling process.

General introduction to classification methods:

All classification methods calculate the means per band for each class of training pixels as defined in the sample set. The set of means per class is called a class mean below; this is an n-dimensional vector in the feature space, formed by n bands.

 

For the Box classifier a multiplication factor has to be specified. This factor is multiplied with the standard deviations of the classes to make the boxes around the classes a bit wider. The larger you choose the multiplication factor, the easier a pixel will be assigned to a class.

For the other classifiers a threshold distance may be specified. You should use a threshold value when pixels with a spectral signature that is not similar to any of the training classes, should not be classified. The threshold value is, for a pixel to be classified, the allowed spectral distance towards a class mean; the method of calculation spectral distances towards class means depends on the selected classification method.

When, for a pixel to be classified, the spectral distance towards a class mean:

The larger you choose the threshold, the easier a pixel will be assigned to a class.

If no threshold value is specified, all pixels of the image will be classified.

Examples of the meaning of a certain threshold value for map lists containing 3 or 7 bands are presented in Classify : algorithm.

In general, a 'good' threshold values does not exist prior to the classification; a 'good' value has to be found interactively while classifying, or estimated from the training pixels in the input sample set.

Box classifier:

For each class, multi-dimensional boxes are drawn around class means based on the standard deviation of the training pixels in each band. The user can insert a multiplication factor (usually > 1) to make all boxes a bit wider.

 

The default multiplication factor is Ö3 for 3 bands.

Minimum Distance to Mean classifier:

For the spectral values of a pixel to be classified, the distances towards the class means are calculated.

 

Minimum Mahalanobis Distance classifier:

For the spectral values of a pixel to be classified, the distances towards the class means are calculated as Mahalanobis distance. The Mahalanobis distance depends on the distances towards class means and the variance-covariance matrix of each class.

 

Maximum Likelihood classifier:

The Maximum Likelihood classification assumes that spectral values of training pixels are statistically distributed according to a 'multi-variate normal (Gaussian) probability density function'.

For each set of spectral input values, the distance towards each of the classes is calculated using Mahalanobis distance. Another factor is added to compensate for within class variability.

 

Maximum Likelihood classifier including Prior Probabilities:

This classification method is similar to the Maximum Likelihood classification but includes a user-defined multiplication factor for each class: the prior probabilities.

Prior probability values can be used to favor certain classes above others, e.g. in case of spectral overlap, on the basis of a priori knowledge. For example, classes which are known to cover large parts of the study area (high frequency), can be given a higher prior probability value.

Prior Probabilities should be listed in a value column of a table, where the table should use the same domain as the sample set. The column then contains for each class, a prior probability value. Each probability value must be ≥ 0. Theoretically, the sum of the prior probability values of all classes should equal 1; however, if the sum of prior probability values in your column is not equal to 1, the program will internally normalize the prior probability values from 0 to 1.

 

For each set of spectral input values, the distance towards each of the classes is calculated using Mahalanobis distance, including a factor to compensate for within class variability, and the prior probabilities.

 

For more information on classification methods, see Classify : algorithm.

Input requirements:

A sample set is required. The standard deviation of each class of training pixels should be > 0 in every band.

For Prior Probability classification, also a table (same domain as sample set) is required; the table should contain a value column that contains for each class a prior probability value (values ≥ 0).

Domain and georeference of output map:

The output map uses the same domain as the sample set.

The output map uses the same georeference as the input maps of the sample set.

Tips:

See also: