A Land Evaluation System Using Artifical Neural Network Based Expert's Knowledge and GIS
Joel C. Bandibas1
1Research Center, Cavite State University,
4122 indang, Cavite, Phillppines
Tel: 046-4150020; Fax: 046-4150012,
E-mail:bandibas@gpu.sru.ualberta.ca
Abstract
This study focused on the development a land evaluation expert systems without undergoing the tedious process of rules formulation. An automated land evaluation system with the transformation of land suitability classification expertise, following the FAO framework, into trained ANN. An error back-propagation ANNs model was used. The trained ANNs were stored in a data base, representing the knowledge base. This knowledge base was used in a land system successfully determined the best land suitability class in the area for com and rice cultivation.
Introduction
Land evaluation is a comprehensive assessment of the bio-physical properties of the land, in order to determine the best use of the area. It is a mulit-disciplinary research, which involves the creation of a data base composed of diverse and multi-criteria information. Consequently, and evaluation export systems using the rule based knowledge base is very difficult, inefficient and time consuming to develop because these requies the formulation of numerous rules or chains of rules (Bandibas, 1995). Indeed, the formulation of rules or chains of rules to store knowledge has been acknowledgement as the bottleneck during expert system development and design.
An alternative method to capture experts knowledge without using rules is the use of ANN. ANN can be trained to process information and to represent knowledge in it (Aleksander and Morton, 1989) Indeed, recent studies suggest that ANN can be trained to identify complex patterns of information and relate them to a certain phenomenon or an object.
The main objective of this study is to develop a land evaluation expert systems without using a rule-based knowledge base. Specifically, this study focused on the development of a GIS linked automated land evaluation system using trained ANN as the knowledge base. A methodology of transforming a land evaluation expertise into trained ANN was developed. An ANN architecture and training scheme were formulated.
ANN Training and the Fao Framework
In this study, the ANNs were trained to simulate the FAO land suitability
classification framework wherein the the land's physico-chemical properties (land characteristics) are divided into six groups namely:
-
Climatic characteristics;
- Topographic/rish to erosion factors;
- Factors affecting wetness of the soil;
- Physical soil characteristics;
- Soil fertility characteristics;
- Salinity and alkalinity
An error back-propagation ANN model was used. The neural network training methodology described by Alexander and Morton (1990) and Rumelhart et al. (1986) was adopted. Figure 1 shows the summary of the ANN training scheme to simulate the FAO land suitability classification expertise. This expertise was in the form of crop requirements. After training the trained ANNs were stored in a neural network based data base. These network were queried and retrieved when used for the actual land suitability classification and land evaluation. Each land use type has its own set of six trained neural network corresponding to the six groups of land characteristics.
The ANN used in this study has three layers representing the input, the hidden and the output. The number of nodes in the input layer depends upon the number of a corresponding neural network. All of the six network has five output nodes and five possible output training patterns. These five output patterns correspond to the five land generate one of the output pattern based on the values of the inputs. For example, if the input pattern of climatic characteristic is
| Length of the growing period (days) |
: |
130 |
| Rainfall, growing cycle (mm) |
: |
560 |
| Mean temperature |
: |
22 |
| Mean minimum temperature |
: |
16.5 |
| Sunshine (n/N) |
: |
0.50 |
| Climatic hazards |
: |
1 (none) |
Which are highly suitable (S
1) values for rainfed com production, the training input and output patterns should be like the one shown in the following figure.
Figure 1. Summary of the ANNs training scheme
Table 1. The land suitability classes output training patterns for the ANN training
| Suitable (S1) |
Moderately Suitable (S2) |
Marginally Suitable (S3) |
Potentially Suitable (N1) |
Unsuitable (N2) |
| 1 |
0 |
0 |
0 |
0 |
| 0 |
1 |
0 |
0 |
0 |
| 0 |
0 |
1 |
0 |
0 |
| 0 |
0 |
0 |
1 |
0 |
| 0 |
0 |
0 |
0 |
1 |
Figure 2 shows the data structure used in this study and the flow of information form the GIS to the ANN during training and actual land suitability classification.
Figure 2. The data structure and the flow of information from the GIS to the ANN (the values used are hypothetical).
