support for using bands 2, 3, 4, and 5 for generating 
spectral statistics on urban areas (Sheffield 1985; 
Chavez, Guptill, & Bowell 1984; Wang 1985). 
Spectral classes were derived using an ELAS module 
named "CLUS" (cluster), a hierarchical clustering 
process based on Ward and Hook's clustering algorithm 
(Ward 1963) . This process works on a pixel-by-pixel 
basis to build a similarity matrix. This matrix is 
computed by suitniing the squared differences in 
spectral value between each possible pair of pixels 
or groups of pixels. The two groups with the most 
similarity are merged into one group at each stage of 
the clustering process and a new "Stat" (mean 
spectral value of class) is computed. In all, there 
were 67 "Stats" generated from the hierarchical 
clustering algorithm. A classification map (named 
CLUS67) was made for the study area using a minimum 
distance classifier which assigned each pixel to the 
class in the "Stat" file that had the nearest mean 
value to that pixel in feature space. The algorithm 
used Euclidean distance measurements to assign pixels 
to their particular classes. 
2 CHARACTERIZATION OF LAND COVER BY SPECTRAL CLASSES 
In many instances, the mean spectral values of 
satellite data are not sufficient for characterizing 
urban land cover conditions alone. It is valuable to 
examine the relationship between spectral classes 
through statistical analysis. In order to better 
understand the relationship between spectral classes 
and the information contained in those classes, 
several statistical analyses were performed on the 
file of 67 "Stat's" generated for this study. 
The first analysis derived principal component 
factor scores from the statistical means. The first 
component accounted for approximately 67 percent of 
the variance found within the data, and the second 
component accounted for 28 percent. The first 
component was very highly correlated to the visible 
TM channels (bands 2 and 3) and the second component 
was very highly correlated to the near infrared 
channel (band 4). Channel,5 was spaced between the 
two factors. 
The principal component factor scores for each 
class were then entered into a clustering analysis 
which grouped spectral classes according to a 
similarity index. This clustering algorithm printed 
a tree-linkage pattern showing which spectral classes 
had means that were similar and calculated their 
amalgamated distances. From studying this tree- 
linkage diagram the groups of spectral classes that 
were most similar were assigned group numbers. 
Discriminant analysis used these spectral group 
numbers and factor scores from the principal 
components to determine canonical discriminant 
function scores for each class. 
The most useful product of the discriminant 
analysis was a scatter diagram which plotted a symbol 
for each spectral class onto a graph according to 
their discriminant scores. The feature space within 
this two-dimensional graph may be divided into 
regions or groups of signatures that correspond to 
particular ground cover types. This procedure allows 
the analyst to concentrate on particular classes of 
interest while signatures of lesser interest may be 
grouped or discarded. The distribution of classes on 
the discriminant function scatter diagram show two 
distinct axes (Figure 1). The first axis displays a 
range of varying brightness from dark signatures to 
light signatures, spreading from class 54 to class 
41. This is ccmmonly referred to as the "brightness 
axis." The second axis, which is highly correlated 
with the near infrared band, stretches from class 61 
to class 40, and is called the "greenness axis." The 
"Greenness axis" is related to percent vegetation 
cover and plant vigor. 
The next step in the land cover classification 
process involved the description of specific land 
cover that was characterized by individual spectral 
Figure 1. Canonical discriminant function scatter 
diagram for CLUS67 spectral classes. 
classes. This was accomplished by sequentially 
highlighting individual classes on an image display 
manifesting variations in cover density or 
brightness. It should be noted that there is a 
gradation between surface cover materials, so 
different land cover classes may display very similar 
spectral curves. Several of the classes contained 
pixels that could be placed into more than one cover 
category, especially when there was confusion caused 
from different surficial materials yielding similar 
signatures. By looking at the shapes of the 
signature curves, the tree-linkage cluster diagrams, 
the discriminant scores, and the aerial photography, 
it was determined which cover category was best 
suited for each spectral class. 
In an attempt to demonstrate characteristics of 
urban land cover categories, the spectral classes the 
for CLUS67 map are represented in Figure 2 by 
"families" of similar spectral curves. Ground 
investigation gave descriptions of surficial 
materials that were manifest in the spectral 
signatures. Figure 2(a), for example, represents the 
curves associated with open water (bottom), and light 
inert materials, such as bare soil, concrete, metal, 
and glass (top). Differences in cover conditions 
contributed to the reflective variations in both of 
these categories. Spectral classes 49, 2, and 48 
displayed a more pronounced horizontal component 
between channels 3 and 4 than the rest of the light 
inert classes. These classes mostly represented 
fields that had been plowed, yet had a small amount 
of crop stubble or vegetation ranaining. Classes 52, 
32, and 61 demonstrated a very steep drop between 
channels 3 and 4 and a flatter curve between channels 
4 and 5. These classes represented many of the flat 
gravel roofs in ccmnercial areas as well as saline or 
mineral soils which were devoid of vegetation. 
Classes 41, 31, and 29 represented bare soil with the 
cover vegetation scraped or plowed off, and were 
generally found in construction areas of dry farms 
(non-irrigated fields in fallow condition, usually 
used for wheat). Classes 16 and 42 were mixed 
between bare soil and some newly completed concrete 
sections of the 1-215 freeway. Classes 50 and 42 
contained combinations of roof tops, construction 
sites, and transportation corridors. Many of these 
classes seemed to confuse roof tops in trailer parks 
and commercial areas with plowed or scraped land. 
This is understandable, since the light colored soils 
are similar components to the bright sand covered 
asphalt shingles on the roofs. 
Water 
in all f 
water-ab 
however, 
dark cov 
classifi 
pattern 
treatmen 
though, 
water an 
pixels i 
noted th 
and near 
while th 
slightly 
however, 
the two 
clusteri 
showed a 
spectral 
displaye 
characte 
water an 
The ve 
primaril 
with occ 
As menti 
difficul 
types fr 
alone, 
differen 
represen 
smelter 
represen 
scattere 
of Salt 
Asphal 
area and 
transpor 
between 
often re 
differen 
inert ma 
areas an 
railroad 
this cov 
characte 
were als 
surplus 
generall 
parking 
also rep 
areas co 
especial 
vegetati 
a higher 
evidence 
and 5 in 
Classes 
and road 
transpor 
grass ar 
There 
mixed pi 
resident 
heteroge 
closely 
the spat 
response 
asphalt 
hybridiz 
f) . Whi 
region o 
as other 
reflecte 
spectral 
Figure 2 
very bri 
courts, 
the shin 
asphalt