THE APPLICATIONS OF NEURAL NETWORKS TO GIS 
IN THE CONSTRUCTION OF LAND EVALUATION MODELS 
Yukiyo YAMAMOTO 
National Grassland Research Institute, JAPAN 
Commission VII, Working Group 7 
KEY WORDS: Agriculture, Land, Estimation, Neural, Modeling, GIS, Application 
ABSTRACT 
Since the Geographic Information System (GIS) is useful for conducting spatial data analysis and integration, it is often 
used in land evaluations. To use GIS, logical and accurate criteria must be used to develop evaluation models. An 
evaluation model should express the entire relationship between all factors related to the evaluation. Though weighting 
or ranking methods are often used as GIS evaluation models, weight and ranking values are subjectively determined by 
each evaluator, so it is difficult to objectively prove their validity. 
This study uses aneural network as a GIS evaluation model. The neural network is an example of artificial intelligence 
technology, and it adjusts itself to fit the supervisor. The network, which lets factors relate to results, is represented by 
a mathematical function like the Sigmoid function. 
In this study, neural networks are used in two applications, they are to estimate the suitability of grassland development 
in Tochigi Prefecture, Japan and to estimate the hazards of land degradation in northeastern Syria. 
In the first assessment, the relevant evaluation f actors were topography, slope, elevation and soil productivity. An 
administrative investigation is used to make a supervised data sets for construction of neural network. As the result of 
trials to construct the network, the neural network, which has 11 units on hidden layer, is highly accurate classified of 
supervised data s ets. Using this network in conjunction with GIS, the land evaluation map for grassland establishment 
is produced. 
In the second application, two networks are constructed to estimate the degree and extent of land degradation. They 
are evaluated by topography, vegetation, soil category, and present land state. 
INTRODUCTION OUTLINE OF NEURAL NETWORKS 
Since GIS has spatial analysis modules, it is used for NN is a computer system that determines relationships 
urban planning or regional and environmental analysis. between factors and results as a problem of pattern 
Similarly, GIS is useful for land evaluation for agricultural recognition. Throughout repeated calculations called 
development and for land-use planning. "learning", the c omputer adjusts its parameters to reduce 
To make land evaluations on s pecific area or f or s pecific estimation errors on data sets automatically. As a result 
purpose, al factors should be considered of this process, a network is constructed. The data s ets 
comprehensively, and evaluation models s hould express using learning are called "supervisor". 
the whole relationship between factors related to In this study, " NEURO92 " which is the computer program 
evaluation. Ranking methods or weighted calculations are that constructs NN, developed in National Grassland 
often used to make GIS models. Thought they reflect Research Institute in Japan, was used. NEURO92 
influence and limitation by the factors, the criteria on constructs multi-layer networks with back-propagation 
these methods are decided by each experimenter. algorithms. It is described in Clanguage, so it is available 
Therefore, it is difficult to testify the objectivity of models. to personal computers and EWS. 
The author remarks Neural Networks(NN) can be used as Figure 1 illustrates a three-layered neural network and 
a method to express the overall relationship between Sigmoid function. The first layer is called input layer, the 
various factors objectively, and for GIS modeling for land second or middle layer is the hidden layer and last layer is 
evaluation. In this paper, t wo applications are introduced the output layer. Each layer has some units. Units 
as GIS models using NN, t he land-suitability assessment between layers are interconnected, just as neurons in the 
for grassland/pasture establishment in central Japan; human brain. The data or information is propagated next 
and evaluation of land degradation in northeastern Syria. units in forward layer by Sigmoid function. 
NEURO?92 automatically adjusts network parameters, 
offsets on each unit and linkage coefficient between 
788 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996 
CA 
In 1 
gras 
con 
tom 
Pref 
The 
usin 
Este 
Fore 
for « 
esta 
74 
inves 
stud 
four 
Elev: 
Outpi 
chart 
92, C 
from: 
the ci