Pattern recognition is the science—and art—of finding meaningful patterns in data, which can be extracted through classification. By spatially and spectrally enhancing an image, pattern recognition can be performed with the human eye; the human brain automatically sorts certain textures and colors into categories.
In a computer system, spectral pattern recognition can be more scientific. Statistics are derived from the spectral characteristics of all pixels in an image. Then, the pixels are sorted based on mathematical criteria. The classification process breaks down into two parts: training and classifying (using a decision rule).
First, the computer system must be trained to recognize patterns in the data. Training is the process of defining the criteria by which these patterns are recognized (Hord, 1982). Training can be performed with either a supervised or an unsupervised method, as explained below.
Supervised training is closely controlled by the analyst. In this process, you select pixels that represent patterns or land cover features that you recognize, or that you can identify with help from other sources, such as aerial photos, ground truth data, or maps. Knowledge of the data, and of the classes desired, is required before classification.
By identifying patterns, you can instruct the computer system to identify pixels with similar characteristics. If the classification is accurate, the resulting classes represent the categories within the data that you originally identified.
Unsupervised training is more computer-automated. You can specify some parameters that the computer uses to uncover statistical patterns that are inherent in the data. These patterns do not necessarily correspond to directly meaningful characteristics of the scene, such as contiguous, easily recognized areas of a particular soil type or land use. They are simply clusters of pixels with similar spectral characteristics. In some cases, it may be more important to identify groups of pixels with similar spectral characteristics than it is to sort pixels into recognizable categories.
Unsupervised training is dependent upon the data itself for the definition of classes. This method is usually used when less is known about the data before classification. It is then the analyst’s responsibility, after classification, to attach meaning to the resulting classes (Jensen, 1996). Unsupervised classification is useful only if the classes can be appropriately interpreted.
The result of training is a set of signatures that defines a training sample or cluster. Each signature corresponds to a class, and is used with a decision rule (explained below) to assign the pixels in the image file to a class. Signatures in ERDAS IMAGINE can be parametric or nonparametric.
A parametric signature is based on statistical parameters (for example, mean and covariance matrix) of the pixels that are in the training sample or cluster. Supervised and unsupervised training can generate parametric signatures. A set of parametric signatures can be used to train a statistically-based classifier (for example, Maximum Likelihood) to define the classes.
A nonparametric signature is not based on statistics, but on discrete objects (polygons or rectangles) in a feature space image. These feature space objects are used to define the boundaries for the classes. A nonparametric classifier uses a set of nonparametric signatures to assign pixels to a class based on their location either inside or outside the area in the feature space image. Supervised training is used to generate nonparametric signatures (Kloer, 1994).
Use ERDAS IMAGINE to generate statistics for a nonparametric signature. This function allows a feature space object to be used to create a parametric signature from the image being classified. However, since a parametric classifier requires a normal distribution of data, the only feature space object for which this would be mathematically valid would be an ellipse (Kloer, 1994).
When both parametric and nonparametric signatures are used to classify an image, you are more able to analyze and visualize the class definitions than either type of signature provides independently (Kloer, 1994).
See Math Topics for information on feature space images and how they are created.
After the signatures are defined, the pixels of the image are sorted into classes based on the signatures by use of a classification decision rule. The decision rule is a mathematical algorithm that, using data contained in the signature, performs the actual sorting of pixels into distinct class values.
Parametric Decision Rule
A parametric decision rule is trained by the parametric signatures. These signatures are defined by the mean vector and covariance matrix for the data file values of the pixels in the signatures. When a parametric decision rule is used, every pixel is assigned to a class since the parametric decision space is continuous (Kloer, 1994).
Nonparametric Decision Rule
A nonparametric decision rule is not based on statistics; therefore, it is independent of the properties of the data. If a pixel is located within the boundary of a nonparametric signature, then this decision rule assigns the pixel to the signature’s class. Basically, a nonparametric decision rule determines whether or not the pixel is located inside of nonparametric signature boundary.
When classifying an image file, the output file is an image file with a thematic raster layer. This file automatically contains the following data:
- class values
- class names
- color table
The image file also contains any signature attributes that were selected in the ERDAS IMAGINE Supervised Classification dialog.
Set the class names, values, and colors using Signature Editor or Raster Attribute CellArray.