Evaluation of Several Classification Methods for

Land Development Constraint Parameters

A. Murni, W. Setiawan, D. Hardianto and B. Kusumoputro

Faculty of Computer Science, University of Indonesia

PO Box 3442, Jakarta 10002, IndonesiaFax: (62 21) 786-3415E-mail:

ABSTRACT

This paper evaluates two classification approaches for land development constraint parameters. The number of constraint parameters are limited to six parameters which include the quality measures of labor power, economic growth, natural resources accessibility, degree of environment sustainability, market orientation, and investment opportunity. The data measurement is in the form of nominal type and the values range from the lowest quality which has a measure value of 1 to the highest quality which has a measure value of 5 value. The model of data distribution is uncertain and we have used both unsupervised and supervised approaches for doing the classification of constraint types. Two methods have been selected which include a hybrid neural network system and a knowledge-based system. A hybrid system of self-organizing map and back propagation architecture is used as the neural network classifier and both non-hierarchical and hierarchical rule-based system is used as the knowledge based system. This paper evaluates the performance of the two methods based on the table of matching accuracy between the method used and the expert. The result shows that the hybrid neural network system performs the best with the correct classification of 92.2%.

  1. Introduction

Multistage, multitemporal and multisensor remote sensing data together with census data are interpreted into thematic maps and used as input data for a geographical information system. Potential region maps were created using multilayer of land information attributes analysis and activity development suitability formula.

In regional planning, it is important to identify potential regions which have among others the following criteria of quality: (i) superior land potential; (ii) dominant factors for triggering economic growth; (iii) specific prime commodity; and (iv) supported regional environment. The successfully of regional activity development could be assessed by evaluating the constraint parameter structure [1-3].

Each region activity development has its own specific effective constraint type. Before assessing the constraint parameter structure, firstly we have to group the constraint parameters into several types of constraint. An expert who knows the observed region very well usually has an intuitive in deciding the number of constraint types. In the absence of an expert knowledge, the number of constraint types and the assessment of constraint structure become an iterative process. Further analysis of a potential region map and its related constraint structure has to be done to obtain the regional planning map [4].

This paper is organized as follows. Section 2 describes the proposed framework for classifying and assessing the land development constraint types, while Section 3 discusses the methods. Section 4 shows the experimental results. Finally, a few concluding and closing remarks are in Section 5.

  1. CLASSIFYING AND ASSESSING CONSTRAINT TYPES

The diagram in Fig. 1 illustrates the framework for classifying and assessing constraint types. It includes the conditions of both the presence and the absence of expert knowledge. In the case of available expert knowledge, a supervised approach is used. Training samples of each constraint type can be used as input data for assessing the constraint logical structure. We have used and expert system tools for both non-hierarchical rule and hierarchical rule approaches for this purpose.

In the absence of expert knowledge, an unsupervised approach has to be used to evaluate the number of constraint classes (or types). We have used a hybrid of self-organizing map and back propagation neural network system for this purpose. After the classes of constraint types is obtained from the unsupervised process or from the knowledge of an expert, then a sample data of the constraint types can be derived for constraint structure assessment. A supervised approach and an expert system tool are used to assess the constraint logical structure based on both the non-hierarchical and hierarchical rules. If the obtained constraint logical structure is acceptable then the combination of the constraint types and the potential land information can be used for further regional planning process. In the next Section 3, both the unsupervised and supervised approaches are discussed.

Multistage, multitemporal, multisensor remote sensing and census data

GIS analysis

Constraint DataSample Data of Constraint Classes

Classification / Clustering Constraint Assessment

Constraint Classes Constraint Logical Structure

No Acceptable?

Yes

Potential Region Map

Final Suitability Analysis and Assessment

Regional Planning Map

Fig. 1. A framework for land development constraint assessment.

  1. Methods for Classifying and Assessing Constraint Types

This section discusses the unsupervised approach for clustering the land development constraint types and the supervised approach for assessing the constraint logical structure. The unsupervised and supervised approaches are used in the hybrid neural network system [5,6]. The result will give the classified constraint data. From the classified constraint data, we can derive a sample data to be evaluated based on the trial and error approach with the aid of an expert system tool to finally find the constraint logical structure. In the trial and error process using the expert system tool both the non-hierarchical and hierarchical rules are used. And in the presence of an expert knowledge, the expert can provide the sample data. The experimental results of using the hybrid neural network system and the knowledge-based system are compared to the expert knowledge and discussed in Section 4. The following paragraphs discuss the hybrid neural network system and the knowledge-based system.

An impact matrix approach was used to define the best set of weighting factors for the constraint parameters in order to increase the discriminating power of the features. Furthermore, using a hierarchical rule-based approach, the features were used to establish the constraint logic structure. The best decision rules are shown in Fig.2 and are used to in the knowledge-based system. The same weighted data features are used as input data for the hybrid neural network system discussed in the next paragraph.

IF(environment_quality is good)

THENIF(natural_resource_accessibility is good)

THEN constraint D

ELSE constraint E

ELSEIF(labor_quality is good)

THEN IF(market_orientation is good OR

Economic_growth is good)

THEN constraint B

ELSE constraint A

ELSE not_potential_region.

Fig.2. The constraint logic structure.

The hybrid architectural network of self-organized module with a supervised network is shown in Fig 3 [6]. The self-organized module performs an iterative process to obtain an optimal number of neurons. The number of output neurons represents the number of existing clusters. These output neurons become input neurons for the supervised network module. The output of this supervised network module has a definite number of output neurons, and these output neurons represent the number of constraint types. The neural network architecture becomes a hybrid of SOM and BP network.

SOMBP

NeuronNeuron Pola atau Neuron

MasukanNeuron Klaster Keluaran

Neuron Tersembunyi

Fig.3. The hybrid architectural network of self-organized module with a supervised network.

  1. EXPERIMENTAL RESULT

There are six constraint parameters used in this experiment. They include the quality measures of labor power, economic growth, natural resources accessibility, degree of environment sustainability, market orientation, and investment opportunity. The labor power was measured based on the regional population and mobility, while the economic growth was estimated based on the gross margin and regional revenue. The natural resources accessibility was estimated based on the availability of area for development and the natural resources that support energy, raw material and water. The degree of environment sustainability was measured based on the existing natural disaster factors and on the quality of environment. The market orientation was estimated based on the availability of infrastructure such as road network while the investment opportunity was measured based on the availability of government, private and community investments.

The quality measures of the parameters are represented with the number between 1 to 5, where the lowest quality has a measure of 1 and the highest quality has a measure of 5. The weighting factors for the labor power, economic growth, natural resources accessibility, degree of environment sustainability, market orientation, and investment opportunity features are 0.15, 0.1, 0.2, 0.3, 0.1, and 0.15 respectively. The impact matrix of a sample of data is shown in Table 1. The number of training samples is 777. The classification results using the two methods are compared to the expert knowledge and represented in the following Table 2 and Table 3.

Table1. The impact matrix of the weighted condition (constraint parameter)

and the corresponding action (constraint type).

Constraint Type / Labor Power / Economic Growth / Resources Accessibility / Environment Quality / Market Orientation / Investment Opportunity / Total Score
A
B
C
D
E

Table 1. Impact matrix and weighting factors of constraint parameters.

SampleConditionConstraint

DataLaborEconomic Accessibility Environment Market InvestTypes

Weight 0.3 0.1 0.2 0.15 0.05 0.2Score

Wilayah-15 1.5 3 0.3 7 1.4 8 1.2 3 0.15 8 1.66.15 (A)

Wilayah-25 1.5 5 0.5 9 1.8 9 1.35 6 0.3 7 1.4 6.85 (B)

Wilayah-3 5 1.5 4 0.4 8 1.6 10 1.5 6 0.3 7 1.4 6.7 (C)

Wilayah-45 1.5 4 0.4 13 2.6 11 1.65 4 0.2 8 1.6 7.95 (D)

Wilayah-54 1.2 3 0.3 10 2.0 10 1.5 4 0.2 6 1.26.4 (E)

Wilayah-64 1.2 3 0.3 7 1.4 8 1.2 3 0.15 8 1.65.85 (O)

Wilayah-74 1.2 3 0.3 7 1.4 9 1.35 6 0.3 7 1.45.15 (O)

Table 2. Correct classification result of the hybrid neural network system.

Constraint type / Hybrid Neural Network / Expert / Matching percentage
A / 119 / 119 / 100.0%
B / 214 / 260 / 82.3%
C / 212 / 228 / 92.9%
D / 55 / 58 / 94.8%
E / 102 / 112 / 91.0%
Average matching percentage / 92.2%

Table 3. Correct classification result of the knowledge-based system.

Constraint type / Knowledge-Based System / Expert / Matching percentage
A / 115 / 119 / 96.6%
B / 305 / 260 / 85.2%
C / 86 / 228 / 37.7%
D / 56 / 58 / 96.5%
E / 46 / 112 / 41.0%
Average matching percentage / 71.4%
  1. CONCLUSION

Both unsupervised and supervised approaches have been used for classifying and assessing land development constraint types. Besides an expert knowledge, a hybrid SOM and BP neural network has been used to assess the number of constraint types. A knowledge-based approach is used to evaluate the result of constraint logic structure assessment. It can be concluded that for this application the adaptive-learning hybrid neural network system performs the best over the knowledge-based system with matching percentage to the expert of 92.2% and 72.8% respectively.

  1. REFERENCES
  1. K. Pedra, L.Winkelbauer and V.R. Pantulu, Expert Systems for Environmental Screening, International Institute for Applied Systems Analysis, Laxenburg, Austria, 1991.
  2. T.J. Kim, L.L. Wiggins and J.R. Wright (editors), Expert Systems: Applications to Urban Planning, Springer-Verlag, New York, 1990.
  3. P.G. Luckman, R.C. Thompson, M.R. Jessen and R.G. Gibb, Some applications of expert systems in environmental sciences, Land and Soil Sciences Division, DSIR, Palmerston North, 1988.
  4. A. Murni, D. Hardianto and S. Nurbaya, The Use of Remote Sensing Techniques and Expert System in Regional Planning, IEEE 1999 International Geoscience and Remote Sensing Symposium, Hamburg, 1999, pp. 619-621.
  5. Y. Miyanaga et al., An Adaptive Learning With Self-Organized Network, IEEE ISCAS’95, 1, 1995, pp. 482-485.
  6. B. Kusumoputro, Development of Self-Organized Network with a Supervised Training in Artificial Odor Discrimination System, Computational Intelligence for Modeling, Control & Automation, M. Mohammadian (editor), IOS Press, 1999, pp. 57-62.