538 
Table 7. Percent correct classification of nine major cover types by each classification approach. 
Classi 
COVER 
Ì Y P E S 
Overall 
Perfor- 
mance 
fication 
Approach Com 
Forest Grass Industry Roads 
Soils Soybean Urban 
Water 
I 
100 
a * 
100 
a 
79 
a 
90 
a 
84 
a 
93 
a 
100 
a 
97 
a 
100 
a 
93.7 
0. 
0 
II 
95 
a 
89 
ab 
65 
ab 
83 
ab 
76 
ab 
88 
ab 
99 
a 
89 
ab 
100 
a 
87.1 
% 
III 
99 
a 
98 
a 
75 
a 
87 
a 
81 
ab 
80 
ab 
99 
a 
98 
a 
100 
a 
90.8 
0. 
0 
IV 
92 
a 
73 
c 
63 
ab 
83 
ab 
71 
ab 
73 
b 
98 
a 
92 
ab 
100 
a 
82.8 
0. 
0 
V 
53 
be 
78 
be 
22 
d 
85 
a 
62 
b 
80 
ab 
86 
b 
64 
cd 
100 
a 
70.0 
0, 
0 
VI 
49 
be 
88 
abc 
50 
be 
75 
ab 
74 
ab 
74 
b 
86 
b 
60 
d 
100 
a 
72.9 
0. 
0 
VII 
64 
b 
55 
d 
40 
cd 
66 
b 
82 
a 
73 
b 
86 
b 
58 
d 
100 
a 
67.4 
% 
Vili 
44 
c 
77 
be 
67 
ab 
36 
c 
38 
c 
49 
c 
85 
b 
78 
be 
100 
a 
63.8 
a 
0 
MSE (%) 
14 
13 
21 
15 
14 
17 
6 
14 
— 
* Within each cover type, approaches followed by the same letter are not significantly different at =0.05 
level by the Bonferroni T - test. (Degrees of freedom = 792, Critical value of T = 3.13) 
4.12 Computer time evaluation 
Considering classifiaction approach I as the standard 
procedure, the CPU time consumed for the Maximum Li 
kelihood Classifier in this approach (7,783 secs) 
was considered as the reference time to compare with 
the other approaches. 
A reduction in CPU time is result of less channels 
used in the classifications. 
Table 8. Computer time (CPU) consumption for each 
approach. 
Classification 
Approach 
CPU Time Ratio 
I 
1 : 1 
II 
1 : 1.3 
III 
1 : 1.3 
IV 
1 : 1.3 
V 
1 : 2.5 
VI 
1 : 2.6 
VII 
1 : 4.0 
Vili 
1 : 4.5 
5 CONCLUSIONS 
The four data sets examined in this research provide 
a method for evaluatting the effect of the TM thermal 
infrared band in multispectral classifications. A 
Per Point GAussian Maximum Likelihood classification 
was performed with eight different approaches. The 
analysis of the data sets with all seven bands or the 
six reflective bands (i.e., data sets A and B), pro 
vided 37 spectrally separable classes. THe use of 
four or three Principal Components provide fewer 
spectrally separable classes. 
The use of the seven TM bands for the analysis pro 
cedure gave better discrimination among classes and 
fewer mixed classes. This same situation prevails 
between data sets C and D where the use of three 
Principal Components gave more mixed classes than 
set C 
The use of the seven TM bands gave the best minimum 
and average separability values. If the thermal 
band is not included for multispectral classification, 
then it is better to generate the training statistics 
(cluster) without the thermal band. 
Water features show to be equally discriminated with 
all the approaches. Soybean and com were better 
discriminated with classifications of the data set A. 
Urban classifications using statistics generated with 
the seven TM bands (data set A) were significantly 
different from those of the other three data sets. 
Soils and industrial classes in the approach VIII 
(Three Principal Components) were significantly di 
fferent and had the lowest accuracy mean values. 
Classifications performed with data sets B, C and D 
provided fores/com mixed classes because of lower 
separability values between those features. 
In general, classifiactions using the thermal band 
were significantly different from classifications 
without this band. THe separability values between 
pairs of classes were higher when the thermal band was 
used. 
When there is a constraint on computer tine and/or 
hardware, the use of data compression techniques such 
as PRincipal Components may be advantageous due to 
the drastic decrease in CPU time consumed. 
The thermal band itself has great possibilities for 
specific types of research, specially in the areas of 
thermal pollution mapping, detection of vegetation on 
stress situations and mapping of sea currents. 
REFERENCES 
Anderson, J.R.,E.E. Hady, J.T. Roach S R.E. Witnjer 
1976. A Land Use and Land Cover Classifiaction Sys 
tem for USe with REmote Sensor Data. Geological 
Survey Professional Paper 964. Supt. of Docs. No. 
119.16:964, U.S.A. 
Anuta, P.E., L.A. Bartolucci, M.E. Dean, D.F. Lozano, 
E. Malaret, C.D. McGillem, J.A. Valdes S C.R. Va 
lenzuela. 1984. LANDSAT-4 MSS and Thematic Mapper 
Data Quality and INformation Content Analysis, in 
IEEE Transactions on Geoscience and Remote Sensing, 
Special Issue on Landsat-4 , Volume GE-22, No. 3, 
pp 222-236. 
Bartolucci, L.A.,M.E. Dean S P.E. Anuta 1983. Evalua 
tion of the Radiometric Quality of the TM Data 
Using Clustering, Linear Transformations and Multi 
spectral Distance MEasures. in Proceedings, Landsat 
4 Scientific Characterizarion Early REsults Sympo 
sium, NASA Goddard Space Flight Center, Greenbelt, 
MD. February 22-24, 1983. 
Bauer, M.E. 1977. Crop IDentification and Area Esti 
mation over Large Geographic AReas Using Landsat 
MSS Data. LARS Information NOte 012477. LARS/Purdue 
University, West Lafayette, IN. 
Bauer, M.E., P.H. Swain, P.R. Mcrozynski, P.E. Anuta 
5 B.R. MacDonald 1971. Detection of Southern Com 
Blight by Remote Sensing Techniques. Proceedings 
Seventh Symposium on Remote Sensing of Environment. 
Ann Arbor, Michigan, pp. 693 - 704 
Brown, 
tion 
mmet 
1414 
Clum, 
viro 
Soci 
Cris^ 
Tran 
Tass 
Remo 
GE - 
Curtis 
of L 
364. 
Hoffer 
tent 
in R 
Univ 
of G 
Kumar, 
tanc 
With 
Mult 
0724 
Lougea 
Regi 
Sens 
MacDon 
J.D. 
the 
of t 
Sens 
Mich 
MacDon 
ment 
679. 
Mroczy 
LARS 
sity 
Myers, 
mote 
hfeas 
7: 1 
Myers, 
mas. 
vati 
on F 
Ann 
Phi Hi 
Mara. 
Price. 
Date 
togi 
229 
Swain. 
due 
Tiltor 
shoj 
Date 
due 
USDA. 
Stat 
Ser\ 
Valdes 
de / 
diar 
espe 
y de 
rets 
Valdes 
red 
tisj 
Agrx 
yeti 
Wear, 
res1 
dice 
Pac, 
Repc