GISdevelopment.net ---> AARS ---> ACRS 1992 ---> Land Use

Knowledge - based Approach to Update Landuse Layer of an Operational GIS

Amarsaikhan. D
Informatics Center Mongolian Academy of Sciences

Gorte. B, Nieuwenhuis. W, Bakker. W
International Institute for Aeorspace Survey and Earth Sciences. Enschede. The Netherlands


Abstract
A knowledge-based approach was used to extract information of land cover types and update landuse layer of an operational GIS. For this purpose was designed a prototype Knowledge-based System (KBS) connected with the RS/GIS system. Whilst the RS/GIS is used to store spatial and non-spatial information and procedures to be used in data manipulation, the KBS is used for decision making on the basis of the gathered knowledge. For the decision making, Gaussian maximum likelihood classifier is used and prior class probabilities are stored within the KBS.

Introduction
Supervised of multispectral remotely sensed data has assumed an increasing importance as an automatic means for land cover mapping, and a great number of investigations have dealt with the application of diverse statistical procedures for discrimination between the cover types of a territory on the basis of their spectral signatures. One of the widely used supervised classifiers is the maximum likelihood decision rule which gives best results by incorporating prior probabilities. Using this technique for the classification of land cover types the same set of training samples whose conditional probabilities are likely to change depending on what cover type is present in a given season can be used. This approach can successfully be applied within an operational GIS, where livelihoods for the state of objects are stored and prior probabilities are defined from the layer (e.g. landuse) of interest according to the context and time. The main aim of the study is to solve this problem within the domain of a knowledge based approach which will require in this case two kinds of knowledge, that is knowledge about conditional probabilities and analysis tools (i.e. techniques to be used).

For that purpose, a prototype KBS linked to a RS/GIS system (e.g. ILWIS) was developed RS/GIS: The function of an information system is to improve a user's ability to make decisions in research, planning and management. An Information system involves a chain of steps: from observation and collection of data, through their analysis, to their use in some decision making process. GIS is computer-based technology, used to input, store, retrieve at will, analyze and output referenced data. This is normally considered to involve a georeferenced computer database and appropriate hardware and application software Primary GIS were mostly used to store and analyze vector objects which are defined by points, lines. And polygons according to the locations and attribute information referred to their characteristics. Present so call RS/GIS systems effectively handle, process and analyze remotely sensed data and have facilities to integrate with the cartographic information stored in vector structure, thus making a promotion in spatial analysis and decision making.

However, retrieving and analyzing parts of these systems is slow and labour consuming. In order to avoid these problems one can design and add to a RS/GIS a KBS to make a rapid decision. Knowledge - based system : A knowledge - based system is mainly a set of computer programs used to make important decisions in a selected narrow field on the basis of gathered knowledge. It consists of the knowledge Base (KB), the Inference Engine (IE) and the User Interface. The KB contains a collection of specific and general facts and descriptions of relations between them. Knowledge representation is the main step in the development of a KB and widely used structures are frame-based and production rule types. In rule-based system for the demand of a user, who is communicating via the Use interface, the Inference engine implements a search within the Knowledge base by the use of defined rules to find answers. There are two reasoning techniques used in production rule systems: forward -chaining and backward-chaining inferences. Forward chaining starts with the evidence, chaining the rules towards the hypothesis. It is the preferred technique if the observations are to be readily available, for instance in Remote Sensing and GIS. In pattern recognition most classification algorithms are based on the forward chaining principle. Backward chaining starts with the most likely hypothesis and searches in the KB facts which can support that hypothesis. In this way only those observations which are necessary to support the hypothesis are treated. It is very applicable if the domain of observations is large. In the Frame-based system an object is represented by a frame, in which the basis storage unit is the slot. A slot is used to store knowledge and procedures to be used. In comparison with rule-base the frames can be arranged in hierarchy, maintaining inheritance, which is a powerful way of retrieving attribute values and getting the same information from a variety of sources. The introduction of KB approach to the spatially-oriented systems will have a number of advantages, e.g. it will promote rapid decision making and simplify procedures to be used.

GIS and RS Data
As a test site Ulaanbaatar area, Mongolia has been selected. We assumed that there is an operational GIS which stores data in layers. Landuse map of the area has 3 classes; urban, dense vegetation and low vegetation. To update the layer of the GIS through the knowledge-based image segmentation process, multi-temporal remotely sensed data has been used. The first image is the SPOT XS image of 1986 and the second is the image of 1990. Before being used in the decision making error correction and geo referencing procedures have been applied. To update the GIS the overall accuracy of the classification should be as high as possible (4).

Design and implementation of the KBS
For the development of the KBS a frame-based structure has been selected. Whilst RS/GIS (ILWIS) is used to store spatial and non-spatial information and procedure to be used in data manipulation, slots of the KBS store information about priors. Livelihoods to be selected according to the time and attached predicates. FBS is linked with the ILWIS system and it is possible to manipulate all procedures of the RS/GIS to use some of the their results in further analysis just temporarily leaving the FBS. In the basic frame of the FBS slots are used to store likelihood for the objects to be evaluated in the classification process which should be taken from the GIS. The role of the database already stored in a GIS is to provide the best possible prior probability for the class of the object under consideration (10). The slots are organized in an hierarchical way, and in lower levels are stored class prior probabilities according to the time. For each selected object explanation facilities have been provided.

The parameters mean (x) and covariance - matrix (x) to be used in the decision making process are evaluated on the basis of the samples defined from the likelihood vectors stored within the GIS. The output is the landuse map which can be used to update the layer of the GIS. Looking at the classified image of the Ulaanbaatar area illustrated in Fig. 1 one can judge what changes have been occurred during the 4 years period. The results of such classification can be used e.g. for the general planning of the city. To update the layer the following arrangements should be considered: 1. The overall classification accuracy must be as high as possible. 2. To compare the classification result with the previously stored layer and judge whether it is necessary to substitute or not. If one consider a model for 4D GIS. Which in comparison with 2D GIs stores volume and time information, the object classes can be stored in different time domain.


Fig. 1

Conclusions
We have shown that the KB approach can successfully be used to update the layer of an operational GIS by storing likelihood and priors. The developed FBS is an early prototype and is limited in many respects. In further improvement, rules to select best prior probabilities. Class uncertainly definition and other necessary changes must be considered.

References
  • Srinivasan. A., and Ricahrds, J.A. 1990 Knowledge Based Techniques for Multi-source Classification, International Journal of Remote Sensing.
  • Jun Chen, Peng Gong, Newkirk. R.T. and Davidson. G. 1991. A knowledge based GIS for Zoning.
  • Fabio Maselli, Claudio Conese, Ljiljana Petkov, and Raffaelo REsti, 1992. Inclusion of Prior Probabilities derived from a Nonparametric Process into the maximum Likelihood Classifier. PE & RS. Vol. 58, No. 2 201-207
  • Mulder, N.J. and Middelkoop, Hans, 1990, Parametric Versus Non-Parametric Maximum Likelihood Classification, proceedings of ISPRS, Wuhan, China.
  • David Miller, and Jane Morris 1992, An ES and GIS Approach to Predicting Europena Conference on GIS 11-19.
  • Levine, R.I. Drang, D.E. and Edelson. B. 1990 Artificial intelligence and expert systems, New York, McGraw-Hill Publishing Company.
  • Matsuyama. T., 1987 Knowledge-based Aerial Image Understanding Systems and Expert Systems for Image Processing I.E.E.E. Transactions on Geoscience and Remote Sensing. 25. 317-0328.
  • Middelkoop. H., Miltenburg.J.W., 1987 Soil. Classification from R S data using knowledge Engineering Approach, ITC 61p
  • John E. Edwards, 1991. Building Knowledge-based Systems, Pitman publishing 128 Long Acre. London WC2E 9AH 244p.
  • Mulder. N.J. 1992. Thematic Information Extraction and Data Fusion, a knowledge Engineering Approach to the Problem of Integral Use of Data from diverse Sources to be publ. Proc. ISPRS. Washington 1992.
  • Mulder, N.J. and Schutte, K., 1992 knowledge Engineering Based Systems in GIs. To be publ. Proc. ISPRS. Washington. 1992.
  • Karman Parsaye et. al., 1987 Intelligent Databases Object-Oriental Deductive Hypermedia Technologies Wiley, ISBN 0-71-50345-2.