Accordingly, a need was expressed to cluster and classify on a data set
composed of two contrasting season ERTS-1 passes, overlain and registered.
Clustering, as explained above, was done using eight channels of data. In
the first test application, the Phoenix metropolitan area, results using
the overlain dates of 16 October 1972 and 2 May 1973 showed a marked im
provement in rural-urban separation over a single date classification done
earlier (figure 1). Success in the Phoenix area prompted use of overlain
temporal data for both Springfield and Indianapolis. Two dates are also
to be employed in planned work on Washington, D.C. and three dates have
been overlain for a test in San Jose, California.
Selection of optimum dates requires consideration of the crop calendar
and the vegetative cycle found in each test site. In humid areas of the
east, with relatively short and distinct growing seasons, one summer date
reflecting mature crops and one winter season date (non snow covered) with
defoliated trees makes for an optimum selection. In Indiana, for example,
the disparity in values in ERTS Band 6 (Channel 3) between corn and soy is
greater in mid-growing season than it is at the end (July rather than late
September). An earlier date with more exposed soil ambiance is equally
unsatisfactory.
Even with multi temporal data, some confusion between rural and urban
persists. Some remains as a real problem whose solution may have to await
introduction of a new tool such as texture analysis, but other confusion is
simply an expression of the generalization within the resolution element,
the character of the sampling where mapping is done at less than full reso
lution, or, as in the case of sparse urban-fringe settlements, just too
rural-like to make the same sort of conceptual generalizations made by the
mind as the eye looks at an air photograph.
Seeking a solution to rural-urban separation by using multitemporal
data resulted in the serendipitous discovery that sharpening the distinction
between troublesome spectral clusters often resulted in the uncovering of
numerous class sub-types. For instance, distinction between some discrete
crop types (corn from soy) and some urban sub-types (older from newer resi
dential) were made. Thus, in the process of making distinctions between
generalized classes -- agriculture versus urban -- more subdivisions of
each were identified, meaning that penetration of a whole new level of
discreteness has been made, one far beyond any land-use applications of
ERTS-1 envisioned earlier.
The maximum and optimum number of land-use classes to be identified
through spectral interpretation is as yet unknown. While the photo inter
preter uses all the traditional aides of shape, color, and areal association,
the spectral interpreter must make all his identifications from spectral data,
digital surrogates for reality. The art of spectral interpretation is not yet
well enough developed to know the limits of the maximum number of classes
possible. Promise is held for such distinctions as the quality of housing and
the nature of open space.