2. EXPERIMENTAL METHODS 
Thematic Mapper digital data with path row annotation 
223/76 were acquired on Jan. 19, 1985. Within this 
area, a segment, of 512x512 pixels, which locates 
on the margin of the Ivai river and about 50 km 
south-west of Maringá, Paraná State, was selected 
for investigation. This segment was chosen not only 
due to its representativeness for the crops in the 
region, but also the cooperation provided by an 
agrobusiness company, where ground information 
such as crop type and variety, planting/harvesting 
dates, crop conditions, field practice, yield etc 
were available. On the date of imaging, predominant 
cover types in this subscene were soybeans, 
sugarcane, com and gallery forest. Pasture and a 
small amount of reforestation and bare soil (i.e. 
plowed fields) were also presented. Soybeans were 
at the flowering stage, com plantations were 
entering senescence, while sugarcane fields of 
various varieties were at different phenological 
stages and percentages of ground cover depending 
whether the field under consideration was planting 
or ratoon crop. Information contents of this segment 
was extracted from CCTs and stored on disk file 
for subsequent analyses. Based on the field 
information, 25 samples varying from 36 to 200 
pixels,were chosen for the above cited cover types, 
and located on the image monitor of an interactive 
image analysis system in INPE. These sampled areas 
were used for band combination study and served later 
as training areas for supervised digital analysis. 
2.1 Triplet band selection for color image 
composition 
For color composition, a subset of 3 bands should 
be selected from the available TM bands and for 
this purpose we used the criterion of entropy. The 
technique of principle component analysis was not 
included because the color image resulted from the 
first three eigenvectors is 
scene dependent and the unknown color-surface 
relationships make the understanding of the 
physical meaning difficult. Any three TM bands form 
a three-dimensional feature space and the 
associated variance-covariance matrix defined an 
ellipsoid within the space. According to the 
entropy criterion, the triplet with the ellipsoid 
of the maximum volume should be chosen (Young and 
Calvert, 1984). The advantage of this criterion 
over other feature selection methods, based on the 
maximum total variance, is discouraging the 
inclusion of highly correlated bands in the 
selection. Theoretically, the maximum ellipsoid 
volume represents the maximum variation in tonality, 
thus it should be a proper criterion in band 
combination selection for color image production. 
In this study the entropy with and without a priori 
of normal distribution were both tested. Once the 
band triplet was chosen, the color assignment was 
nade as suggested in Sheffield's study (1985) : 
green, which is most sensitive to hunam eyes, is 
assigned to the band with the highest variance, red 
to the band of the second largest variance, and 
blue to the band of the smallest variance. The 
program implemented in our image analysis system 
gives the first six-band triplets, the prime colors 
were then assigned, slides were taken from the 
image monitor and visual evaluations were made on 
these slides for their interpretabilities. 
2.2 Band selection for digital analysis 
The best TM band combination for digital analysis 
should take into account the accuracy of the 
classification results and the computer time 
consumed. The discrimination function, "Jeffereys- 
Matusita distance" or the J-M distance, was used as 
the criterion for band combination. In multiclass 
classification problem we can chose the band 
combination, which maximizes the mean J-M distance 
between two classes or that maximizes the minimum 
J-M distance. In this study both criteria were 
applied to select the best hand combination if 2,3, 
4 or 5 TM bands were used for digital analysis. 
Plotting the separabilities of the best 2,3,4 and 5 
bands the optimal band combination was chosen. For 
classification, training statistics of the 25 samples 
were used to charecterize the spectral responses of 
the cover types considered. These training statistics 
were then utilized to classify the whole area using a 
maximum likelihood decision rule of the image 
analysis system. Comparisons were made on: (1) 
alphanumeric print-outs (2) the classification 
matrix, (3) the computer time consumed, and (4) 
the upper bound of the probability of error by the 
J-M criterion. After these comparisons the best 
band combination for crop discrimination was 
selected. 
3. RESULTS AND DISCUSSION 
3.1 TM color image composite 
Table 1 shows the basic statistics and coefficients 
of correlation for the six TM bands investigated. 
As expected, intercorrelations were found among 
visible bands, between bands 4 and 5 and between 
5 and 7 for this highly vegetated area. The NIR 
band 4 has the highest variance and a wider spread 
of grey level indicating more information content 
than the other bands. Decreasing variances are 
observed from NIR to middle IR and then to the 
visible bands. The ranked first six band triplets 
by the entropy criterion, with and without the 
Jaussian priori, are listed in table 2. Independent to 
whether the priori was used or not, band 4 was the most 
important TM band for color compositing. The second 
band included in the triplet was band 5 and the 
last was a visible band or band 7. Once the band 
triplet was decided color coding was assigned 
according to the amplitudes of their variances; in 
our case green to the NIR band, red to middle-IR 
band and blue to the visible band. If both middle- 
IR bands were included in the combination, then red 
was assicmed to band 5 and blue to band 7. Visual 
comparisons of slides taken from the image monitor 
for the selected color capos ites showed that when 
bands 4 and 5 were included in the band triplet 
together with any one of the visible bands or band 
7, color appearances of vegetations were similar. 
Thus, we could not claim preference over any one 
of the combinations. Coparing the conventional 
false color caposite (FCC) ; assigning blue to band 
2, green to band 3 and red to band 4, to any of the 
ranked combinations no significant improvement in 
visual crop discrimination was noted. Under this 
condition the conventional FCC is favoured over the 
entropy selections because no additional training 
of photointerpreter on the color-surface relations 
is required. This conclusion, drawn from the 512x 
512 pixels subscene, may be too local-specific. 
Acquisition of FCCs for the quadrant using 
conventional and the first ranked band combination 
will be requested for a further investigation. 
3.2 Band selection for digital analysis 
The ranked first six band combination, when 2,3,4 
or 5 bands were used in digital analysis, are 
shown in table 3. Plotting the separabilities of 
the best combinations (Fig.l) we note that the 
statistical structure for crop discrimination had 
three or four dimensionalities. These results are 
in agreement with that obtained in previous studies 
by Townshend (1984) and Anuta et al. (1984). No 
matter whether the criterion of Max. J-M niean. or 
Max. J-M min was used, the best three-and four- 
band combinations, which should be included in digital 
analysis, were the same; they were bands 2-4-5 and 
Table 1 - 
1 
2 
0.94 
3 
0.70 
4 
-0-07 
5 
0.22 
7 
0.30 
x 
67.81 
a 2 
10.12 
*n 
= 512x! 
Table 2 - 
Rank 
En- 
Gai 
B* 
1 
2 
2 
3 
3 
1 
4 
7 
5 
2 
6 
7 
* Colors ; 
B = Blu< 
2-4-5-7 re 
both comb: 
based on i 
average o 
upper bout 
error was 
three-banc 
(Table 4), 
showed the 
soybeans, 
time const 
this 512x1 
than the i 
difference 
frame clae 
another ac 
considerai 
three-banc 
example, i 
compactab] 
analysis ; 
that for 
bands 2,4 
analysis. 
4. CONCLUÍ 
IANDSAT Tb 
to select 
copositic 
conclusior 
- Accorc 
5 and one 
should be 
However, r 
discrimine 
observed c 
criterion, 
2,3, and 4 
color-suri 
photointei