537 
> cll 
3 4 
>C2 
5 6 
^ n 
y 
3 4 
5 6 
>C3 
3 4 
r 
5 6 
Bands 
Bands 
~ m 
3 4 5 6 Bands 
>C5 
3 4 5 6 Bands 
TM-PC6 
3456 Bands 
2. Loadings or 
Lents for Prin- 
Dmpcnents of 6 
Lve TM bands, 
ït D). 
tion. 
in Principal 
seled Cap Planes 
data when the 
sformation; then 
ed by the TM 
ained with MSS 
ults of Anuta, 
spectrally se- 
21 spectrally 
3 same area. 
bands (Approach 
i the first data 
lasses. The cla- 
nonsidered the 
tion from all 
andard procedure 
ational facili- 
TM bands 
the classifi 
cation . In this 
rded in the Per 
classes selected 
rat band. 
1 bands (Approah 
assifiacticn for 
yLth the best 6 
^formed to assess 
aand cn classi- 
4.5 Classification with the best four TM bands 
(Approach IV) 
A classification with the best four bands was oerfor 
med to compare this results with those of four Prin 
cipal Components. Even though the combination of 
bands 1, 4, 5, and 6 had a greater average ET (1975 
-vs- 1973); the 1, 4, 5 and 7 band combination was 
selected because of the higher minimum ET value ob 
tained with this combination (660 -vs- 457). 
These results confirm that band 7 provides more 
information in the higher ordered Principal Compo 
nents transformation than does the thermal band. 
4.6 Classification with the six reflective bands 
(approach V) 
If digital pattern recognition analysis of remotely 
sensed data is performed with a selected combinatio- 
of spectral bands the training statistics (cluster) 
must be generated with those bands (Swain, 1983). 
A second multispectral analysis was conducted over 
the sane area to evaluate a classification performed 
without the thermal band. The training stattistics 
were derived from the six reflective bands only. 
There were 37 spectrally separable classes as in 
the data set A, but there were differences between 
the training statistics of the two data sets. This 
second set of training statistics had more mixed 
spectral classes than did the data set A. This mixing 
occured mainly in non-water, non-vegetative classes. 
Both the minimum and the average ET values for the 
second data set (B) were greater than those obtained 
in the data set A. 
4.7 Classification with 4 Principal Components 
(Approach VI) 
The multispectral analysis of Principal Components 
was carried out with a slightly different technique 
than that used for the analysis of the TM bands. The 
selection criteria used in the analysis of Principal 
Components was based mainly on the separability 
betweeen pairs of classes and their spatial distribu 
tion on the cluster map. 
The final training statistics for approach VI con 
tained 35 spectrally separable classes. However, the 
number of mixed spectral classes had increased. The 
minimum and average separability values were greater 
in approach VI than those obtained using the best 
four bands (approach IV). 
4.8 Classification with 3 Principal Components 
(Approach VII) 
The first three principal components of data sets 
C and D contained approximately the same amount of 
information, the difference being that the data set 
D had slightly greater cumulative percentage variance 
than did the data set C (Tables 3 and 4). A classi 
fication with the first three Principal Components of 
data set C was performed to be compared with the cla 
ssification from data set D. 
4.9 Multispectral analysis of data set D (Approach 
VIII) 
In data set D the first three principal components 
account for 98.979 % of the total variance in the 
scene (Table 4). These three components were utilized 
in the multispectral analysis, in which 31 spectrally 
separable classes were obtained. 
The minimum ET value obtained in this approach was 
significantly greater tahn that obtained in the data 
set C for the best (first) three Principal Components 
(Approach VII). There was no great difference among 
the average separability of all the eight approaches. 
Table 6. Average and minimum separability values 
(Transformed Divergence Distance, DT) for each cla- 
ssifiaction approach. 
Approach Bands or Minimum Average 
P. Components Separability Separability 
Data set 
A = 37 spectrally 
separable 
classes 
I 
1,2 ,3,4,5,6 ,7 
1625 
1991 
II 
1,2,3,4,5,7 
959 
1983 
III 
1.3.4.5.6.7 
1578 
1990 
IV 
1,4,5,7 
660 
1973 
— 
1,4,5,6 
457 
1975 
Data set 
B = 37 spectrally 
separable 
classes 
V 
1,2 ,3,4,5 ,7 
1659 
1991 
Data set 
C = 35 spectrally 
separable 
classes 
VI 
1,2 ,3,4 
1650 
1986 
VII 
1,2,3 
753 
1970 
Data set 
D = 31 spectrally 
separable 
classes 
‘ VIII 
1,2,3 
1534 
1979 
4.10 Visual evaluation 
The eight classifiactions were displayed on a color 
video display device where they were visually evalua 
ted. This evaluation was performed by assigning a 
different color for each of the spectral classes 
obtained in the 8 classifications and comparing them 
with the low altitude aerial photographs. 
The classifiactions performed with the data set A, 
were considered the best. THe classifications perform 
med with data sets B, C and D were ranked from good 
to bad in that order. 
4.11 Statistical evaluation 
To evaluate the classification accuracy of each appro 
ach, the final spectral classes obtained for each data 
set were grouped into nine major domains: Com, soy 
bean , forest, grass , bare soil, roads, urban, industry 
and water. One hundred pixels of known identity were. 
defined for each of the nine cover types. Those nine 
hundred points were compared with the identification 
label obtained for each of them in the eight classi 
fications . 
Confidence intervals. may be more useful than signia 
ficance test in multiple comparisons. Confidence in 
tervals show the degree of uncertainty in each compa 
rison in an easily interpretable way. Considering this, 
a BOnferroni confidence interval test was adopted to 
evaluate the classification performance of each of 
the eight approaches for the nine coyer types. 
The results of the BOnferroni test are presented in 
Table 7. The eight approaches of classification were 
evaluated for each cover type. 
There was not an approach that could be considered 
different from the others for all the nine cover 
types. 
Approaches I and II were considered non significantly 
different for the nine cover types. Approaches I and 
V were not considered different for cover types in 
dustry, soils and water. Approaches II and III were 
considered different for the cover type roads, and 
approaches VI and VIII where considered significantly 
different for the non-vegetated cover types.