d of 60
ne in the
5, emplo-
of detail,
ind truth
nd stored
2. Digital methods
The Hardware: The digital analyses of landsat MSS da7gand the geographical information
system (GIS) interplays were done at the Michigan State University Center for Remote Sensing
on the ERDAS 400 self-contained image processing microcomputer.
Ihe LANDSAT Data set The study area occupies the extreme southeastern corner of the
LAFIA scene (E—-2353 090125) of January 10 1976, with a zero percent cloud cover and qua-
lity of 5588.
A window of 600 x 200 pixels, corresponding approximately to the area covered by the
soil map was extracted from this corner of the scene for digital processing.
Supervised classification: A false colour composite (FCC) of this subscene was displayed
using bands 4, 5 and 7. Making use of suitable enlargements, training sites were selected for
signature records of the various physiographic—soil units. Training sites were either near the
points of soil examination (auger bore—hole or profile) or an area known to have similar soil
characteristics. Since the soils were not exposed at the time this Landsat data were acquired,
the reflectances recorded would be of soil reflectances diluted by the vegetation reflectances
or of entirely vegetation reflectances.
Each physiographic—soil unit was sampled at least once, Where multiple training sites were
taken, the signatures were pooled and the mean statistics — reflectances of the pixels on the
training sites in each of the four MSS bands and covariance matrices calculated. These stati-
stics are stored in a signature file, the subscene data were subjected to two types of Super
vised classification: Maximum likelihood and mi nimu m distance.
The maximum likelihood algorithm calculates the probability of a given pixel belonging
to the same classas that of a training site. The assumption is made that there is a normal
distribution of the cloud of points forming.the class, i. e. GAUSSIAN (Lillesand and Kieffer,
1979). This 600 x 200 pixel window took 8 to 8% hours to run.
The Minimum distance classifier, on the other hand, mathematically more simplistic, was
at least 2% times faster than maximum likelihood approach. In this classifier, the Euclidean
mean is calculated for each pixel. The mean spectral value is also calculated for each known
category. An unknown pixel is classified by computing the distance between the unknown
pixel and each of the mean spectral values for the categories. The unknown -pixel is assigned
to the closest class, i .e. the class with the shortest distance to the pixel. In either case, the
spectral curves have been plotted and each class has a known spectral response curve.
Two pairs of classes — the Western sand island levee and the levee of the River Benue banks
and the Awgu valley bottoms and the valley bottoms of the remaining uplands were merged
and the signature file data subjected to maximum likelihood and minimum distance classifica-
tions.
Unsupervised classification: Unlike the supervised classification, the user does not interact
with the computer to supply it with information classes. Rather, the pixels to be classified
are unknown and the computer is to evolve algorithms that group the pixels into a priori classes
on the basis of the spectral values of the image or data. The user supplies a few parameters such
as the maximum number of clusters. The rule is that values in a given class are close together and
the fairly distinct from other classes.