TM-PC1 
Table 5. Study data sets used for multispectral analysis 
and the eight different classification approaches. 
Data 
Classif. 
Statistics 
ClaS&ification 
Set 
Approach 
Generation 
Bands 
A 
I 
7 TM bands 
7 
A 
II 
7 TM bands 
6 reflective 
A 
III 
7 TM bands 
6 best 
A 
IV 
7 TM bands 
4 best 
B 
V 
6 reflective 
6 
C 
VI 
4 Princ. Coimp. 
4 
C 
VII 
4 Princ. Comp. 
3 best 
D 
VIII 
3 Princ. Comp. 
3 
If a pair of classes had a value of DT of 1850 or 
greater, these classes were considered different and 
spectrally separable. Then the analysis was focused 
cn those with DT values less than 1850. 
Final cluster classes selected to train the computer 
for classification are those which were considered 
totally discriminable within the cluster classes and 
representative of the land cover/land use features 
present in the study set. The cluster classes that 
were not used for classification were "deleted" from 
the stattistics deck. 
The multispectral classifiaction was performed using 
a "Per - Point" Maximum Likelihood Classifier. 
4 RESULTS 
4.1 Principal Components evaluation 
To evaluate the importance of the thermal band for 
classification purposes, two Principal Components 
transformations were performed, one utilizing the se 
ven original TM bands and other with the six reflec 
tive bands. The coefficients of the high ordered 
Principal Components describe which of the TM bands 
contains most of the significant variance of informa 
tion for this data set. 
The fourth Principal Component was almost entirely 
loaded with the thermal band'(98.22%) and accounts 
for 1.019 % of the scenen variation. This result 
show that the thermal band is highly correlated with 
the fourth Principal Component (Table 1) as first 
reported by Bartolucci, et al. ,1983. Even though the 
thermal data provided only one percent of the total 
scene variation, (Table 3), thermal information or 
variance may be distinctly unique from the rest of 
the bands. 
The use of the thermal band in linear transformations 
of TM data creates a fourth dimension or Principal 
Component which is highly correlated with the thermal 
band (Table 1). This plane or Fourth Principal Compo 
nent does not appears in the transformation performed 
using the six reflective TM bands only (Table 2). 
Figures 1 and 2 show graphically the loadings or 
coefficients for both principal components data sets. 
Principal Components 1,2, and 3 of both data sets 
had mare or less the same shape as did the last three 
Principal Components of both data sets. 
The results of the linear transformaton performed in 
the data set B containing the six reflective TM bands 
were compared with the results obtained by Crist and 
Cicone (1984) with a scene over North Carolina. They 
did not use the thermal band for the "Tasseled Cap 
Transformation". The'found that with six TM reflective 
bands there are only three components or features. If 
the thermal band is employed in the transformation, 
the result will be four planes of information in 
which the use of a fourth component will account for 
1 234 576 Bands 
1 234 56 Bands 
1 
TM-PC2 1 
TM-PC2 
V\ vn V\ ^ 
W § 
m 
-1 
1234576 Bands 
1 2 3 4 5 6 
1 
TM-PC3 1 
TM-PC3 
^ vm Vi F3 ^ n 
YX m 03 f£! 
i § 
-1 
1 2 3 4 5 7 6 Bands - 1 _1 
1 2 3 4 5 6 
Bands 
Bands 
’234 576 Bands 
TM-PC5 
— m 
1 2 3 4 5 7 6 Bands 
TM-PC6 
1234576 Bands 
__ TM-PC7 
1 2 3 4 5 7 6 Bands 
Figure 1. Loadings or 
coefficients for Prin 
cipal Components of 
original“? TM-'bands 
(Data set C). 
TM-PC4 
-car-m gg 
1 234 se Bands 
TM-PC5 
E3 ^ 
1 2 3 4 5 
TM-PC6 
Bands 
1234 se Bands 
Figure 2. Loadings or 
coefficients for Prin 
cipal Components of 6 
reflective TM bands. 
(Data set D). 
over 99 % of the cumulative data variation. 
Considering that MSS data have two main Principal 
Components (Anuta, et al.,1984) or Tasseled Cap Planes 
and that TM data have four features of data when the 
thermal band is considered in the transformation; then 
the uncorrelated planes of data provided by the TM 
can be considered twice that those obtained with MSS 
data. This results agrees with the results of Anuta, 
et al. (1984), where they obtained 42 spectrally se 
parable classes with TM data and only 21 spectrally 
separably classeswith MSS data from the same area. 
4.2 Classifiaction with all available bands (Approach 
I) 
The multispectral analysis performed in the first data 
set produced 37 spectrally separable classes. The cla 
sses selected for classification were considered the 
most representative of the scene variation from all 
the spectral cluster classes obtained. 
This type of classification is the standard procedure 
when there are no constrains in computational facili 
ties (Anuta, et al. , 1984). 
4.3 Classifiaction with six reflective TM bands 
(Approach II) 
This approach was performed to compare the classifi 
cation results with the first classification. In this 
approach the thermal band was not included in the Per 
Point classification, but the training classes selected 
were generated with the inclusion of that band. 
4.4 Classifiaction with the best six TM bands (Approah 
III) 
To evaluate the possible changes in classifiaction for 
the second approach, a classification with the best 6 
bands (Bands 1,3,4,5,687) was performed to assess 
the effect of elimination of a single band on classi 
fication .