462 
are either mature (non-green) or harvested in 
March. Possible spectral confusion classes for 
boro in March are forested areas, aquatic 
vegetation such as water hyacinth in ponds or 
reservoirs and homesteads. 
The best available Landsat scene of sufficient 
quality during the time of greatest boro green 
biomass was 3 March (scene number E-2406-03404). 
Maps and statistical information on the location 
and amount of boro rice were obtained from the 
Landsat MSS data using both digital and visual 
analysis strategies. 
3 DIGITAL PROCESSING 
A variety of digital processing techniques for 
mapping boro rice were investigated as a part of 
this project. These techniques ranged from rather 
simple and inexpensive, to more complex and/or 
costly procedures. The objective was to determine 
the technique with the best combination of 
accuracy and cost effectiveness. The various 
techniques examined for mapping the boro are 
presented following. 
3.1 Single date single band level slicing 
Graymaps from individual MSS bands 5 and 7 
produced for the study area using system derived 
level slicing did not discriminate features of 
interest. Manual level slicing maps improved on 
the maps and provided indications of some of the 
boro locations but also confused water and boro 
areas. 
MSS 5 
30 I 
29 I 
28 I 
27 I 
26 I 
25 J 
24 ! 
23 I 
22 I 
21 I 
20 I 
19 I 
18 I 
17 I 
16 I 
15 I 
14 I 
13 I 
12 I 
WWWWWWWW00000000000000000000000000000000 
wwwwwwwwoooooooooooooooooooooooooooooooo 
wwwwwwwwoooooooooooooooooooooooooooooooo 
wwwwwwwwoooooooooooooooooooooooooooooooo 
WWWWWWWW00HHHHH0000000000000000000000000 
WWWWWWWWWHHHHHH0000000000000000000000000 
WWWWWWWWWHHHHHHOBBBBBBBBOOOOOOOOOOOOOOOO 
WWWWWWWWWHHHHHBBBBBBBBBBBOOOOOOOOOOOOOOO 
WWWWWWWWWHHBBBBBBBBBBBBBBBOOOOOOOOOOOOOO 
WWWWWWWWWWBBBBBBBBBBBBBBBBOOOOOOOOOOOOOO 
WWWWWWWWWWBBBBBBBBBBBBBBBBOOOOOOOOOOOOOO 
WWWWWWWWWWBBBBBBBBBBBBBBBBOOOOOOOOOOOOOO 
WWWWWWWWWWBBBBBBBBBBBBBBBBOOOOOOOOOOOOOO 
WWWWWWWWWWBBBBBBBBBBBBBBOOOOOOOOOOOOOOOO 
WWWWWWWWWWBBBBBBBBBBBBOOOOOOOOOOOOOOOOOO 
wwwwwwwwoooooooooooooooooooooooooooooooo 
wwwwwwwwoooooooooooooooooooooooooooooooo 
wwwwwwwwoooooooooooooooooooooooooooooooo 
wwwwwwwwoooooooooooooooooooooooooooooooo 
0000000001111111111222222222233333333334 
1234567890123456789012345678901234567890 
MSS 7 
Figure 1. Table look-up classification for boro 
rice(B;, water (W), homestead (H), and 
other (0). 
signature map had a boro recognition similar to 
the look-up table but with a slightly lower amount 
of boro acreage. 
3.6 Multidate processing 
3.2 Ratio 
Ratio graymaps using bands 5 and 7 provided 
reasonable recognition of boro pixels but confused 
other cover types and did not adequately delineate 
the water features. 
3.3 Look-up table 
A spectral scatter plot of the study area was 
generated in MSS bands 5 and 7. The plot was 
spatially divided on the basis of known spectral 
signatures to recognize four cover types; boro, 
water, homestead and other, see Figure 1. 
Emphasis was then placed on examining those 
boundary pixels between the known signatures. In 
these spectral boundary areas there were 
considerable numbers of reflectances involved 
where the signatures might be from boro, from 
homesteads or very frequently, mixed pixels. The 
classified map generated by the look-up table was 
compared to available ancillary data and ground 
truth information. A reasonable map showing the 
four cover types was obtained by this technique. 
3.4 Unsupervised classification 
Unsupervised signature extraction was accomplished 
with an automatic clustering algorithm that used a 
sample of the pixels and all four Landsat MSS data 
channels. Initially 24 clusters were obtained and 
then reduced to 17. The clusters were identified 
using available ground truth and a four category 
map was generated. The percentage of mapped area 
identified as boro was slightly less in the 
cluster map than in the table look-up map. 
3.5 Supervised classification 
Twenty-six sites were selected for signature 
extraction for boro, homestead, water and other to 
represent the scene classes. Using these 
signatures, a maximum liklihood classifier was 
employed to produce a map for comparison to a map 
generated the look-up table. The supervised 
Since the classes for possible spectral confusion 
with boro rice in March, forests, aquatic 
vegetation and homesteads, are largely spectrally 
invariant, one way to separate them may be by use 
of multidate Landsat data in which boro areas 
appear non-green on one of the dates. Multidate 
clustering was accomplished on a registered data 
set including bands 5 and 7 for March and December 
(when many of the boro fields are non-green) and 
bands 5,6 and 7 for the same months (Colwell, 
1977). Cluster classifications were attemped by 
use of a zoom transfer scope to merge the cluster 
maps and available ancillary maps. The cluster 
spectral signatures were also examined. Multidate 
clustering and classification indicated an 
improved recognition of homesteads, but this was 
accompanied by poorer recognition of other land 
classes due to decreased resolution caused by 
slight misregistration of the merged data. 
Misregistration is particularly troublesome in 
areas such as Bangladesh where many of the 
features, such as homesteads and water areas, are 
small and often linear. 
For mapping boro rice, a table look-up procedure 
has the greatest cost-effectiveness. The 3 March 
Landsat data were used, and Landsat MSS 5 and 7 
data space was divided into boro, water, homestead 
and other categories. The data was then processed 
using the subjectively established non-linear 
boundaries between the various classes in this 
data space to produce maps. By using the table 
look-up, as in any traditional multivariate 
recognition processing not based on resolution 
decomposition, some pixels that contain less than 
100 percent boro will be called entirely boro, and 
some pixels in which a fraction of the pixel is 
occupied by boro area will be called non-boro. 
The effects should be partially compensating 
(Horwitz, 1971). 
4 BORO STATISTICS 
The average field size in Bangladesh is quite 
small, usually less than the spatial resolution 
(0.4 h 
single 
invento 
expecte 
product 
adjacen 
hence 
cluster 
pumps 
hectare 
Howev 
many pi 
so sm 
disting 
This 
irrigat 
methods 
based 
would p 
Boro 
officer 
units n 
Landsat 
the ver 
Landsat 
indeper 
on Lar 
maps 
betweer 
the Lar 
Initi 
indicat 
boro we 
not ur 
factors 
those 
small 
confuse 
situati 
process 
produce 
boro t 
small f 
of the 
total 
variab] 
of it 
to be 
It is 
consist 
but the 
aerial 
of Lar 
better 
A c 
impleme 
fairly 
procede 
inventr 
estimal 
estimal 
within 
total 
betweer 
sampler 
For 
which ' 
the di 
indica- 
were i 
When i 
perfori 
was foi 
constai 
Therefi 
assumei 
estima 
the doi 
Landsa 
determ 
for ani