Usually, classification is performed with a set of target classes in mind. Such a set is called a classification scheme (or classification system). The purpose of such a scheme is to provide a framework for organizing and categorizing the information that can be extracted from the data (Jensen et al, 1983). The proper classification scheme includes classes that are both important to the study and discernible from the data on hand. Most schemes have a hierarchical structure, which can describe a study area in several levels of detail.
A number of classification schemes have been developed by specialists who have inventoried a geographic region. Some references for professionally-developed schemes are listed below:
- Anderson, J.R., et al. 1976. "A Land Use and Land Cover Classification System for Use with Remote Sensor Data." U.S. Geological Survey Professional Paper 964.
- Cowardin, Lewis M., et al. 1979. Classification of Wetlands and Deepwater Habitats of the United States. Washington, D.C.: U.S. Fish and Wildlife Service.
- Florida Topographic Bureau, Thematic Mapping Section. 1985. Florida Land Use, Cover and Forms Classification System. Florida Department of Transportation, Procedure No. 550-010-001-a.
- Michigan Land Use Classification and Reference Committee. 1975. Michigan Land Cover/Use Classification System. Lansing, Michigan: State of Michigan Office of Land Use.
Other states or government agencies may also have specialized land use or land cover studies.
It is recommended that the classification process is begun by defining a classification scheme for the application, using previously developed schemes, like those above, as a general framework.
A process is iterative when it repeats an action. The objective of the ERDAS IMAGINE system is to enable you to iteratively create and refine signatures and classified image files to arrive at a desired final classification. The ERDAS IMAGINE classification utilities are tools to be used as needed, not a numbered list of steps that must always be followed in order.
The total classification can be achieved with either the supervised or unsupervised methods, or a combination of both. Some examples are below:
- Signatures created from both supervised and unsupervised training can be merged and appended together.
- Signature evaluation tools can be used to indicate which signatures are spectrally similar. This helps to determine which signatures should be merged or deleted. These tools also help define optimum band combinations for classification. Using the optimum band combination may reduce the time required to run a classification process.
- Since classifications (supervised or unsupervised) can be based on a particular area of interest (either defined in a raster layer or an .aoi layer), signatures and classifications can be generated from previous classification results.
Supervised vs. Unsupervised Training
In supervised training, it is important to have a set of desired classes in mind, and then create the appropriate signatures from the data. You must also have some way of recognizing pixels that represent the classes that you want to extract.
Supervised classification is usually appropriate when you want to identify relatively few classes, when you have selected training sites that can be verified with ground truth data, or when you can identify distinct, homogeneous regions that represent each class.
On the other hand, if you want the classes to be determined by spectral distinctions that are inherent in the data so that you can define the classes later, then the application is better suited to unsupervised training. Use Unsupervised training to define many classes easily, and identify classes that are not in contiguous, easily recognized regions.
Supervised classification also includes using a set of classes that is generated from an unsupervised classification. Using a combination of supervised and unsupervised classification may yield optimum results, especially with large data sets (for example, multiple Landsat scenes). For example, unsupervised classification may be useful for generating a basic set of classes, then supervised classification can be used for further definition of the classes.
Classifying Enhanced Data
For many specialized applications, classifying data that have been merged, spectrally merged or enhanced—with principal components, image algebra, or other transformations—can produce very specific and meaningful results. However, without understanding the data and the enhancements used, it is recommended that only the original, remotely-sensed data be classified.
Dimensionality refers to the number of layers being classified. For example, a data file with 3 layers is said to be 3-dimensional, since 3-dimensional feature space is plotted to analyze the data.
Feature space and dimensionality are discussed in Math Topics.
Using programs in ERDAS IMAGINE, you can add layers to existing image files. Therefore, you can incorporate data (called ancillary data) other than remotely-sensed data into the classification. Use ancillary data to incorporate variables into the classification from, for example, vector layers, previously classified data, or elevation data. The data file values of the ancillary data become an additional feature of each pixel, thus influencing the classification (Jensen, 1996).
Although ERDAS IMAGINE allows an unlimited number of layers of data to be used for one classification, it is usually wise to reduce the dimensionality of the data as much as possible. Often, certain layers of data are redundant or extraneous to the task at hand. Unnecessary data take up valuable disk space, and cause the computer system to perform more arduous calculations, which slows down processing.