a ^c» & Q
land cover types have not aggregated in the first several
clusters(table 2) but separately distributed. Having
identified the corresponding land cover type of these
peaks, it is convenient to evaluate and label surrounding
'non-peak' clusters. The extracted rice area accuracy
with strategy B is over 84% compared with the statistics
obtained from the Agricultural Investigation Team
affiliated to the Agricultural Bureau of Hubei province.
4. DISCUSSION
In supervised classification, the spectral characteristics
of the training sites are used to "train" the classification
algorithm for eventual land cover mapping of the
remainder of the image(Jensen, 1986). Once
representative training sites selected, multivariate
statistical parameters calculated from each training site
are used to evaluate every pixel. The pixel is assigned to
the class of which it has the highest likelihood of being a
member no matter it is within or outside the
administrative boundary. As far as strategy A and
strategy B are concerned, there is not much difference
between these two strategies for supervised
classification.
In a unsupervised classification, the computer is allowed
to select the class criteria such as means and covariance
matrices(Jensen, 1986). In strategy A, class criteria are
calculated within the administrative boundary only
while it is done within the whole circumscribing
rectangle in strategy B. Certainly, diverse results will
emerge for unsupervised classification with different
strategy.
In image data, the spatial dependence among pixels is
the fundamental aspect of spatial pattern(Henebry, 1993).
If the satellite image was masked with boundary pixels,
information about this dependence is lost at all(Henebry,
1993). According to this rationale, contrary to the
method used by many authors, strategy B is used in our
project. Statistical unsupervised classification work was
243
performed firstly and then mask out the uninterested
area outside the administrative boundary. By this way,
we hope, the spatial dependence can be retained
especially when it is concerned with the rice theme. Our
results suggest that strategy B is practical in rice
identification and obviously excellent than strategy A
for unsupervised classification.
In the above discussion, no matter which strategy was
used, the deviation induced by the boundary pixels
should be noticed. Readers can refer to Crapper(1984)
on how to calculate the boundary pixels. Rao and
Mohankumar(1994) explained in detail the effect of
spatial resolution and the percentage of boundary pixels
on accuracy of area estimation. In general usage,
boundary pixels when occupying only a little percentage
of the total study area(0.442 96, in our experiment), can
usually be ignored in acreage estimation.
For visual interpretation, uncertainty in assigning a
theme to a grid cell can occur when the grid cell lies near
or on the boundary of a region(Crapper, 1984).
According to this rationale, strategy B should also be
considered in visual interpretation. That's to say, first,
conduct the visual interpretation work and then mask out
the uninterested area and calculate the labeled results.
5. CONCLUSIONS AND FURTHER RESEARCH
For strategy A, we only give the result of pre-classified
50 clusters. Further labeling and interpretation work
haven't been done. Surely, it can be done, but certainly,
it will be more complex than strategy B and the accuracy
won't be satisfying.
In our test, the study area was also classified into 70
clusters for strategy A and strategy B in unsupervised
classification. Similarly, with strategy B, the formerly
classified 70 clusters can be more easily reclustered into
10 types than with strategy A.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996