of Landsat MSS pixels in May 1984. Test field has
16km x 12km area and the population is 14768 after
excluding mixels. The test field contains seven cate-
gories with widely varied frequencies (Table 6): agri-
cultural field(A),barren(B), developed area(D), for-
est(F), paddy field(P),residential area(R) and water
surface(W).
Table 6 Pixels of each category (total 14768)
Category A B D F P R W
Number 3398 175 1903 273 8084 566 369
Frequency (96) | 23 1.24 413 1.8 ..55 12 2.5
Training samples were randomly selected in propor-
tional to the frequency, sampling rates was 5 percent
for the whole pixels. The rest 95 percent samples
were used for classification test. Namely, number of
training samples was 738 and number of classification
samples was 14030.
Table 7 Classification accuracy about samples in
determined nodes for error tolerance p
p 396 5% 7%
Size of determined nodes (%) | 42.8 62.5 72.4
Total accuracy(%) 91.5 86.7 84.4
Averaged accuracy(%) 42.4 36.7 36.5
A,B,D,F,P,R,W
The classification result about Landsat MSS'data
are shown in Fig.7 and Table 7.
Fig. 7 shows the triplet tree produced by the pro-
posal method. Categories expected have similar fea-
tures such as {P, W} and {B,D,R} were in the near
node and divided repeatedly in the tree. In this case,
no determined node for category F was appeared.
Characteristics of Forest area in this image is near
other category, especially agricultural field, so clas-
sification with specified reliability could not be exe-
cuted.
Table 7 shows the sizes of determined nodes and
their classification accuracies. When parameter p be-
comes larger, size of determined nodes increases but
classification accuracies decreases.
Table 8 and Table 9 show classification accuracy.
Bayesian method is superior in averaged performance.
Triplet trees is superior in total accuracy, by exclud-
ing indistinct part to undetermined node. Present
design method calculate boundary with sample num-
bers, so in case categories with similar distribution
have large difference in frequency, categories with
small frequency is tend to be ignored.
A,B,D,F,R
Fig.7 Tree for Landsat MSS data in case error tolerance p = 5%.
Table 8 Classification accuracy of triplet tree (p = 5%) using 95% of samples: size of determined
node, accuracy and reliability (= 100 — commission_error) for each category, averaged accuracy
and total accuracy
Classifier A B D F
P R W | Averaged | Total
Size of determined nodes | 1365 42 411 0
6773 . 140 4l (Total 62.5%,8772 pixels)
Classification accuracy 15.9 8.1 53.9 0
Classification reliability 79... 14.3... 68.1. 5 NaN
93.3 - 11,80 01 36.7 86.7
93.0 ^ 12.9 43.9 E =
Table 9 Classification accuracy of Bayesian classifiers using 95% of samples: accuracies for each
category, averaged accuracy, and total accuracy
Classifier | A B D F P R W | Averaged | Total
LDF 33.0 30.3 ‚52.1 386.1. 753 .20.5. 40.1 48.2 59.2
QDF 298 337. 532 639 77.5 182 852.3 47.7 59.8
992
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
m
su
te
al
cl
tl
fic
S:
tr
tl
st