Supervised training requires a priori (already known) information about the data, such as:
- What type of classes need to be extracted? Soil type? Land use? Vegetation?
- What classes are most likely to be present in the data? That is, which types of land cover, soil, or vegetation (or whatever) are represented by the data?
In supervised training, you rely on your own pattern recognition skills and a priori knowledge of the data to help the system determine the statistical criteria (signatures) for data classification.
To select reliable samples, you should know some information—either spatial or spectral—about the pixels that you want to classify.
The location of a specific characteristic, such as a land cover type, may be known through ground truthing. Ground truthing refers to the acquisition of knowledge about the study area from field work, analysis of aerial photography, personal experience, and so forth. Ground truth data are considered to be the most accurate (true) data available about the area of study. They should be collected at the same time as the remotely sensed data, so that the data correspond as much as possible (Star and Estes, 1990). However, some ground data may not be very accurate due to a number of errors and inaccuracies.
Training Samples and Feature Space Objects
Training samples (also called samples) are sets of pixels that represent what is recognized as a discernible pattern, or potential class. The system calculates statistics from the sample pixels to create a parametric signature for the class.
The following terms are sometimes used interchangeably in reference to training samples. For clarity, they are used in this documentation as follows:
- Training sample, or sample, is a set of pixels selected to represent a potential class. The data file values for these pixels are used to generate a parametric signature.
- Training field, or training site, is the geographical AOI in the image represented by the pixels in a sample. Usually, it is previously identified with the use of ground truth data.
Feature space objects are user-defined areas of interest (AOIs) in a feature space image. The feature space signature is based on these objects.