Supervised Classification

This map shows a supervised classification to identify land use and land cover in Germantown, MD for the purposes of tracking changes in population and land consumption.  An image of the area was used to create signatures to be used in classification, and based on those signatures, the image was classified and the signatures merged into 8 classes.  The area for each class was calculated so that, when compared with images from previous or later years, the changes can be calculated.  

Creating signatures requires paying attention to what type of process you use for each signature.  Sometimes it is better to draw a polygon, other times Growing a Seed is the preferred method.  In general, the second method seemed to be best, although for the Fallow Field 1 signature, the polygon was a good choice.  By looking at histograms of each signature, it was possible to see if there was a good enough sample of pixels, and if there was separation in the spectral signature between classes.  By using the Mean Plot it was possible to determine the best band combination to create separation between each spectral signature.

With more time, even better signatures could have been created, but the distance image shows there were not too many poorly classified areas.

Unsupervised Classification

This is an example of unsupervised classification. Here a Landsat image was processed so that clusters of similar pixels were grouped as classes, and then those classes were matched to various feature types in the original image.  The result is a thematic map representing, in this case, 5 classes.  The classes can be analyzed to answer questions; in this case the question was what percentage of the image represented impermeable versus permeable surface.  The necessary information was contained in the Attribute table for the newly created classification image.  One issue that came up was that some pixels were classified in more than one group - for example there were green grass pixels on the roofs of some buildings.  The solution was to create a "Mixed" class that would allow for those situations where classification into a clearly identified group was not possible.