WSP07147 - Laserfiche WebLink

Home Browse Search

. . . Classification Q C) C.:> ~ t.t.) to"" Digital image analysis techniques have several advantages compared with visual techniques. They are generally faster, provide objective decisions, and capitalize on the multiple bands and high radiometric resolution of TM data (Meyers, 1983). T~e digital nature of the source image data is also directly compatible with GIS databases. The overall objective of image classification procedures is to automatically categorize all pixels in an image into land cover classes or themes. This is accomplished by identifying different feature types based on their inherent spectral reflectance or emittance properties (Ullisand and Kiefer, 1987). One type of spectrally oriented classification procedure is the unsupervised classification technique. Unsupervised classifiers involve algorithms that examine the unknown pixels in an image and aggregate them into a number of classes based on the natural groupings or clusters present in the image values (Lillis and and Kiefer, 1987). Clustering allows . the analyst to quickly identify many classes and to detect classes that are not in contiguous, easily recognized regions (ERDAS, 1991). All image processing for the Upper Colorado Irrigated Lands Project is being accomplished with ERDAS software. The particular algorithm used for this studY is an iterative technique referred to as ISODATA (Iterative Self-Organizing Data Analysis Technique). The ISODATA process begins with a specified number of arbitrary cluster means and processes them repetitively to shift the arbitrary means to the means of the clusters in the data (ERDAS, 1991). As an iterative technique, it is not biased toward the top of the file, as in one-pass algorithms. Upon completion of the clustering program, a set of spectral signature files is produced. Each signature is a set of statistical data that defines an individual spectral cluster. These signatures must be evaluated to determine the validity of the clustering results. One evaluation technique plots the normal distributions for each two band combination in two-dimensional spectral space. When plotted in this way, each potential class results in an ellipse. When the ellipses have extensive overlap, the spectral characteristics of the pixels represented by the two classes cannot be distinguished (Figure 4a). A plot with no overlap between ellipses would indicate that each class is spectrally distinct (Figure 4b). Too much separation however, can result in portions of the image remaining unclassified. The image analyst must determine the maximum number of classes the data can support and to relate those spectral classes to land cover types. Parameters such as the number, size, and separation of clusters can be specified by the analyst. 8