(2) Results
Of the results of the 15 trials with different
number s of units in their hidden layers, networks
with 11 units were the best ones for both degree
and extent. The accuracies which w ere evaluated
the ratio of points estimated within 0.25
estimation error to supervisor, were 86.4% and 7
9.4% . They approximately satisfied the accuracy
necessary for evaluation.
Figure 6 and Figure 7 are results of map
calculation using network parameters for degree
and extent of | and degradation. Figure 8 shows t he
hazard ranking as a whole. It was produced by
reclassifying outputs of the multiplication.
S
Severe — — Moderate
Figure 6. The estimation for the degree
of land degradation.
(*Color Printed)
3 aE SS RRR M
5 Rank3 à
SS Rank4 ud
From Southeast
Char acteristics related to natural factors of the
hazar dous zones are shown in Tables 3 to 7. They
are occupied area of different ranks in each
category. Specific features were not clear for
elevation, slope, direction of aspect and vegetation
coverage, because they were distributed in
proportion to the ratio of each category over the
whole ar ea. On the other hand, features r elatedto
soil category shown in Table 6 were unique. Though
most of test sites were Calcic Xerochrept (Xero.
deep, Xero. slop) and Litic Xerorthent (Xerot.), the
hazardous sites were in Gypsic Xer ochr ept (Gy ps.)
or Litic Xerochrept (Liti.) categories. Since they
were stony soil, soil structures with stone and
rock appear to influence land degradation.
d
Severe Qo Moderate
Figure 7. The estimation for the extent
of land degradation.
(*Color Printed)
Rank 2
Rank 3
#88 Rank 4 From North
Figure 8. The hazard map for land degradation. (*Color Printed)
792
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
dM oe ja wa- 8