Training sites and ground cover

You designate training sites based on samples of different surface cover types in your imagery by drawing colored regions or areas over the parts of the image that are likely to be the information classes you want to extract.

You cannot know for certain what the actual ground cover in an image is by referencing only the image; therefore, samples (training sites) must be based on familiarity with the geographical region and knowledge of the actual surface cover types shown in the image.

Training sites are areas in an image that are representative of each of the land cover classes that you want to define. Focus examines the pixel values within the training sites in order to compile a statistical signature for each training site class. The training signatures serve as the interpretation key for each pixel in the image. All pixels in the image are compared to the signatures and then classified.

Each training site has an associated threshold value and bias value. Threshold is a relative measure used to control the radius of the hyperellipse for each class. By changing the threshold values, you can reduce the chances of pixels being classified into more than one class. Bias is a value from 0 to 1, where higher values weigh one class in favor of another. It can also be used to resolve overlap between classes. You can use both of these measurements to test the training site separability.

You can use the Training Site Editor to create training sites to supervise the classification.

For mathematical reasons, all training sites for a class must together have at least one pixel for each channel used in classification. For example, if four (4) channels are selected, the total number of pixels in all training sites for the class must be at least four (4). A class is rejected from computations if this condition is not met.

For statistical reasons that relate to confidence in estimates, it is recommended that all training sites for a class be at least 5*(N*N+N) pixels large, where N is the number of selected channels. If four (4) channels are selected, the recommended number of training pixels for each class is 100. Larger training sites usually yield better statistics and thus better classification results.