The objective of the present study was to determine if spatial
features could be employed for Level I land use classification using satellite
imagery. In this paper, classification performance using spatial features,
spectral features, and spectral average features, which refer to the mean
spectral response of an array of image data, are compared. In addition,
a test of the extendability of spatial features as compared to spectral
features in the temporal domain is reported.
METHODS
ERTS frame numbers 1352-17134 and 1388-17131 which imaged the
Colorado Front Range on July 10, 1973, and August 15, 1973, respectively
were selected for classification analysis using spectral and spatial informa
tion. Four Level I land-use categories as defined by Anderson (1973)
including Urban, Agriculture, Rangeland, and Forests were identified on the
ERTS transparencies by manual photointerpretation. The agriculture category
was further divided into irrigated farmland, characterized by small farm
plots and strong return in the reflective infrared, and dryland farming,
generally characterized by large field plots.
Figure 1 is a grayscale printout generated directly from the
digital magnetic tape onto a microfilm plotter of Multispectral Scanner
(MSS) band 7 for the study area. There is a contrast reversal on the figure
introduced by the photo-reproduction process. Six training fields are
indicated on the Figure, ranging in size from 24,576 ground resolution ele
ments or pixels to 55,296 pixels. All subsequent results reported in this
paper refer to the results obtained on these training areas. As nearly as
possible, -these same areas were identified on the August 15 imagery. For
example, Figures 2a and b show a closeup of the dryland farming category
and of the irrigated farming category in July; Figure 3 shows a close-up of the
urban category. Table I summarizes the Level I land-use categories investi
gated in this study and indicates the sample size in each category for each
of the classification tests performed.
Spatial Features : One way of summarizing the spatial variation in
spectral response over a two-dimensional array of image points is to expand
the array in a two-dimensional Fourier transform and examine the power spectrum
of the individual spatial frequencies occurring in the image array or cell.
Digitally, this is accomplished by employing the fast Fourier transform. A
detailed discussion of the techniques involved in these two processes is given
in an earlier published report by the authors (Hornung, 1974). In applying
the digital techniques to a classification problem, one must determine which
spectral band, which cell size, and which sampling strategy of the resulting
Fourier Power spectrum to utilize. In the present study, it was determined
by successive calculation and classification that digitally scanning MSS band
7 with a cell size of 32 by 32 pixels and using a concentric circular ring
sampling of the Fourier power spectrum yield optimum results. This scheme