Supervised classification revolves around the creation of training sites to train the software in what to look for when conducting the classification. This is accomplished by creating a polygon type area of spectrally similar pixels. Examples would be dense forest, grassland, or water. Each area has a distinct spectral signature. These signatures are used to evaluate the whole of the image and allow the software to automatically reclassify all matching spectral areas. The overall process is usually in four steps: get your image, establish spectral signatures, run the classification based on the signatures, then reclassify or identify rather what your class schema is.
The process is fairly straightforward with only a couple things to watch out for when establishing the spectral signatures. Being sure to avoid spectral confusion. This is where multiple features exhibit similar spectral signatures. This usually occurs most frequently in the visual bands, and can be avoided by doing a good check using tools such as band histograms or spectral mean plotting which shows you the mean spectral value of one or more bands simultaneously. We can see some of the results of this below, such as the merging of the urban/residential and the roads/urban mix.
The image below shows a Land Use derivative for Germantown, Maryland. It was created using a base image and supervised classification looking for the categories displayed in the legend. This map shows the acreage of areas as they currently exist and is intended to provide a baseline for change. As areas get developed, the same techniques can be used on more and more current imagery to map the change and gauge which land uses are expanding/shrinking most and by how much. This was an excellent introduction to one method of supervised classification, there are many types and reasons to conduct it, but those are for another class.
No comments:
Post a Comment