1 Image classification

In image classification we generate a thematic map from a (set of) input channel(s). These input maps are usually aerial or satellite data. Multispectral data can be considered as a stack of raster maps with identical spatial reference. During the image classification procedure the spectral response of objects is analysed and assigned to classes. The resulting map contains a set of classes which may represent landuse and landcover.

GRASS supports multiple channels, they can be grouped together with i.group. Then either an automated statistical analysis is done on the input channels (unsupervised classification) or training areas have to be digitized by the user to define known landuse/landcover areas (supervised classification). GRASS then derives spectral signatures for the desired classes and runs the final analysis on all pixels of all input channels, assigning each pixel to a class. In the case of unsupervised classification the classes are just numbered, in the case of supervised classification they correspond to the names of the training areas.

While the more sophisticated supervised classification is explained in the literature, we will show here a simple unsupervised classification here (Maximum Likelihood algorithm):

     i.group group=lsat subgroup=lsat in=tm.1,tm.2,tm.3,tm.4,tm.5,tm.7

     i.cluster group=lsat subgroup=lsat sig=sig.cluster $\backslash$

               classes=15 sep=1.5

     i.maxlik group=lsat subgroup=lsat sig=sig.cluster $\backslash$

               class=tm.class rej=tm.class.rej

     d.rast.leg tm.class

     d.rast.leg tm.class.rej

The tm.class map holds the result, the tm.class.rej map the confidence level for each pixel.