en
An
ng
by
ips
orn
led
uit
ets
am
and
)92
ion
ble
and
ris
its
the
ext
NS,
units, to minimize t he estimation errors c ompared output
layer in the supervisor. After learning, network
parameters are saved to files in the computer.
I :Input data
H;-fx)-1/ (1+exp(-2x 1/0) H:Output from Hidden Layer
x1=(£ Wii TI)*A; O:Output from Output Layer
J
Op, =f(x)=1/(1+exp(-2x 2/4))
x7=(2 Wy HprBy W:Linkage Coefficent
u :Inclination of Sigmoid
Figure 1. The model of multi-layers neural
network and Sigmoid function f(x).
CASE STUDY 1 : Land Suitability Assessment
for Grassland/Pasture Establishment in
Central Japan
In this study, an NN to evaluate land suitability for
grassland and pasture based on natural factors is
constructed. This constructed network is applied to GIS
to make an evaluation map of test site, which is Tochigi
Prefecture, located in central Japan.
(1) Preparation for the Supervisor
The supervisor for learning by NEURO92 was prepared
using the data of " the Land Inquiries for Grassland
Establishment (1983 - 1986, The Ministry of Agriculture,
Forestry and Fisheries) ". Investigations were c onducted
for all of sites where it is possible to develop and
establish grassland and pasture in the whole of country. 1
74 data sets in Tochigi were extracted from the
investigation data sets as the Supervisor. Since this
study aims at an assessment based on natural factors,
four elements were selected as input data, Slope,
Elevation, Topography and Soil Productivity. As an
output, suitability rankings as evaluated by the person in
charge of each investigation were used. To use in NEURO
92, data for the supervisor was standardized, ranging
from 0.0 to 1.0. Input and output data were modified using
the criteria shown in Table 1.
789
A:Unit Offset on Hidden Layer
B:Unit Offset on Output Layer
Table 1. The criteria for Supervisor of NEURO92.
Natural Factors Categories
Slope ES 8~15 15° <=
Area (ha) /100 77
Elevation Average elevation (m) /1000
Topography Foot of Mt. Moutainous Hill
a T b RR EU n
debis Plats Low-land
0.8 ra
Soil Productivity Excellent Standard Inferior
En SE Ba
Evaluation Excellent Good Possible
by Person 0875 0.625 0.375
(2) Neural Network Construction using
NEURO92
The network was assumed that it had three layers, the
input, hidden, and Output layers. The network is
completed incrementally through learning to reduce
estimation errors. However, the effect of learning is
influenced by learning times and the number of units in
the hidden layer. For this reason, trials altering the
number of units in the hidden layer were conducted in
order to select the most accurate network.
Figure 2 shows the results of trials after 10000 Cycles of
learning with 1 to 15 units. It illustrated the accuracy
which has been calculated the ratio of classified data
correctly to s upervisor. Inthe results, the most accurate
ratio was 85.196. It was marked by two networks with 1 1
and 15 units in their hidden layer. Though both were
considered to be effective as an evaluation model for t he
assessment, the f ewer units network was preferable for
the model because of map calculation costs. For this
reason, the network which has 11 units in its hidden layer
was selected as the evaluation model.
(3) Drafting an Evaluation Map for Grassland
Establishment
To draft and produce an evaluation map for grassland
establishment, it is necessary to inputmap data to the
GIS, which corresponds to the input layer's information of
supervisor. Consequently, contour, topography and soil
productivity maps were digitized and corrected
georeferentially with GIS. Attribute files of each map were
made, in accordance with Table1.
Figure 3 and Table 2 show the network model and
parameters determined by NEURO92. These parameters
were used for map calculations. As a result of map
calculation, the evaluation map shown in Figure 4 was
drafted.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996