Clustering and spectral interpretation is done for each of the major
cover types independently before the analyst tries to assemble the results
with the help of the LARSYS separability function. Clusters from spectrally
similar major cover types are compared with this function which indicates
statistical separability between all clusters. Some of those that conflict
greatly are deleted at this stage if they represent two different types of
land cover. Those clusters defining the same cover are pooled together if
spectrally similar to form new clusters, the average of their components.
The analysis establishes clusters for as many discrete land uses and
covers as possible realizing that subsuming into broader but well defined
classes is necessary in some instances. Thus, original clusters representing
mostly new multi-family residences and those representing mostly new single
family residences may eventually have to be subsumed into a class called new
residential. The end product contains land-use classes that are quite
discrete, sometimes beyond the Level II proposed by the U.S. Geological
Survey for use with remotely sensed imagery (Anderson, Hardy and Roach,
1972). The degree of detail possible seems to be a function of the geo
graphic environment and the availability of enough sample areas on which
to cluster. It was found that certain broad levels of land use are repli
cable in all cases studied while those more discrete could perhaps be dis
cerned only in areas where they were most pronounced. It was not possible
to map older versus newer residential in Phoenix as it was in Springfield
and Indianapolis only because the indicators (high density and/or accompany
ing mature vegetation) that made such a distinction possible in the latter
areas were absent. For final display, various subsumings of the discrete
classes made the final map more readable by suppressing some of the detail.
These groupings of the classes can be made to fit various user needs.
The analysis proceeds in building block fashion -- clustering, spectral
interpretation, checking separability, deleting, pooling -- until all clusters
remaining have been checked with each other using the separability function.
At this stage the function is utilized as a feature selector, providing the*
analyst with a statistical idea of cluster separability using various subsets
of the available spectral channels. In the Indianapolis work it was decided
from this information that only five of the eight channels were needed:
channels 1, 2 and 4 from January and 2 and 3 from September were selected as
the all-round best combination.
Throughout, test classifications are made to assess cluster suitability
in various areas from throughout the test site. After a final set of clusters
and channels is decided upon, a final check is made. Areas surrounding county
boundary intersections are classified and checked. At the same time the line
and column coordinates of those boundaries are found, providing information
necessary for aggregating land use/cover by county. Computer cards are
punched designating the ending column of each county encountered along each
line as input for a modified version of LARSYS Printresults that then aggre
gates the data by hectares and percentage.