Building a geo-expert system integrating Remote Sensing and GIS

Building a geo-expert system integrating Remote Sensing and GIS

Mrs. P. Venkatachalam and C.V.S.S.B.R. Murty
Centre of Studies in Resources Engineering
Indian Institute of Technology, Bombay, 400076, India

Abstract
Ideally GIS and Remote Sensing should be integrated into one Geo-Expert System which can answer questions, take action and give advice in a seemingly intelligent way based on facts contained in GIS and on the procedures and data available in digital Remote Sensing system. A hierarchy of lower level expert systems can work be more specialized and contain knowledge about a specific domain, eg. Geology, soils etc. In a way, the logic rules for interpretation and classification given in the Manual of Remote Sensing can be put into expert system for specific domain of application. Using expert system shells, a domain knowledge engineer can translate the application domain knowledge of a real expert (e.g a hydrologist or a geologist) into a knowledge base so that both Remote Sensing and Geographic Information Systems are integrated optimal for solutions in specific areas. This paper deals with the development of a knowledge based system which can assist a user to find the groundwater potential of region using a segmented Remote Sensing imagery and the related data stored in a GIS. The domain experts knowledge is constructed into a rule base with backward chining control strategy. As the rule base is not exhaustive additional rules can be imbedded into it. The system is built using PC based shell.

Introduction
Human experts in any field are frequently in great demand and are therefore, usually in short supply. Whether for repairing automobiles or drilling for oils or analyzing chemicals, there are times when access to the knowledge experience and judgement of an expert in a field would be an invaluable asset. One solution to the dilemma is the expert system, an AI Computer program specially designed to represent human expertise in a particular domain (area of expertise). An expert system contains knowledge about a particular field to assist human (Experts) to provide information to people who do not have an access to an expert in the particular field. Although both expert systems and data base programs feature retrieval of stored information, the two types of programs differ greatly. A data base programs retrieves facts that are stored, while an expert system uses reasoning to draw conclusions from stored facts, add new facts and gives the possibility to work with uncertain and missing data.

An information system is a data base system providing answers to questions about its domain. It contains facts about the field of specialization and has facilities for maintaining. Extending and correcting these facts. Geographic information systems (GIS) are systems specialized to deal with earth related data. Remote Sensing (RS) is an activity which collects earth related data and information extraction from Remote Sensing data is done either by image processing followed by human interpretation or automated pattern recognition.

Ideally GIS and RS systems should be integrated into one Geo-Expert System (GE) which can answer questions, take action and give advice in a seemingly intelligent way based on facts contained in GIS and on the procedures and data available in a digital RS system. A hierarchy of lower level expert systems can work under the supervision of a master GES. Each lower level system can be more specialized and contain knowledge about a specific domain e.g. geology, soils etc. In a way, the logic rules for interpretation and classification given in the manual of Remote Sensing (Reeves, 1975) can be put into expert system for specific domain of application. In image processing applied to Remote Sensing, large number of subroutines and hardware are available. To make an intelligent use of them an expert advisor can be prepared which can help in translating the user's problem into a strategy so that suitable subroutines can be called and optimal parameters can be set up for analyzing the data. Presently GIS packages are often based on business type data bases. The expert system approach, will allow a more flexible way to implement optimum query language, ease data base maintenance and will add the possibility of working with uncertain or partly missing data. Using expert system approach, a domain knowledge engineer can translate the application domain knowledge of real expert (Eg. a hydrologist or a geologist) into a knowledge base so that both RS and GIS can be integrated for solutions in specific areas.

To demonstrate the integration of knowledge engineering techniques for image interpretation and classification, study on soil mapping using SPOT image has been carried out (Mulder et al., 1988). Input into the system is defined by a segmented image. The system works through a hierarchy of observations, partly following a forward chaining search strategy. Using the information provided by the user, the system concludes the most likely soil class for each region in the segmented image.

In earth resources application, one of the earliest expert systems developed was PROSPECTOR (Duda et al., 1974). It contains the rule based models for different kinds of ore deposits and helps to evaluate the favourableness of a ore deposits and helps to evaluate the favourableness of a geologic district for any kind of ore. A mixed control strategy is followed and it accommodates uncertainly in both evidences and rules. GEOMYCIN (Davis & Nanninga, 1985) demonstrates the possibility of incorporating spatial knowledge for forest management. Knowledge based systems for aerial photo interpretation have been developed (Nagao & Matusuyama 1980; McKeown , 1985). The need for Artificial Intelligence Principle in the integration of remotely sensed data with GIS has been emphasized (McKeown, 1987). A rule base has been built to classify thematic mapper data into Eucalypt forest types been integrating terrain features (Skidmore, 1989).

In this paper, an attempt has been made to build a knowledge based system which can assist a user to find the potential groundwater region using a segmented Remote Sensing imagery and related geo-hydrological data that can be provided by GIS. The system works on backward chaining control strategy and Bayesian theory is applied for supporting the hypothesis based on evidences. The system is prepared on a PC based shell and has necessary explanation facility.

Groundwater Assessment Model
The purpose of this study is investigate how a knowledge based system could be developed which would be able to assist the groundwater resources assessment. The expert system 'AQUIFER' which helps to detect aquifers from Landsat MSS images (Peacegood et al., 1986) became the starting point of the study. In this system, the entire satellite image is considered as a single object and a class label in the form of an 'overall likelihood' is assigned for finding and aquifer in that area. The area covered by a satellite image cannot be treated as a single homogeneous unit for identification on an aquifer. Instead, on the basis of a hydrological model, the image can be subdivided in hydrogeological unit a label of occurrence of aquifer can be assigned to each of these units. Also the weight factors are introduced in the system to express 'strong', 'weak' etc by assigning some numerical values which are not based on the frequencies of co-occurrence between classes and attributes.

The basic goal for the present study is to provide a measure of probability for finding groundwater in an area on the identification and delineation of geohydrological units which may be shallow aquifers. The consultation starts with forward chaining process to identify main landscape types. Then the user is asked to process to identify main landscape types. Then the user is asked to subdivide each landscape type into homogeneous areas (regions) based on the elements relevant for identification of geo hydrologic units within that landscape type. These regions are the object and for each object, the system will run and determine the probability for the presence of gorundwater. This part of the procedure is done following a backward chaining procedure.

At the highest level, the user is asked to segment the image into alluvial and hardrcok areas. These areas must be further segmented into homogenous regions based on relief and stream characteristics. The system tries to find the presence of shallow aquifers for each of these regions. It asks the user the type of depositional environment such as miander belt, back swamp, braided, deltaic etc. under these environments, it looks for the favorable areas for groundwater. For example, in miander belt, the system asks the areal spread of sandy deposits like natural levees, points bars and abandoned channel areas. For each of these deposits, evidence for existence of shallow aquifer is found by the presence of dense vegetation irrigated arable land etc. The data organization in the system is shown in figure 1.

Fig. 1
Even though thus system is a simplification of the real world it helps to know the presence of shallow aquifers by identifying hydrological units.

There is always a measure of uncertainly between landscape units and geo-hydrologic units. For example, the large parcels and presence of arable land may not always indicate natural levee. Hence a posteriori probability for finding a unit on the basis of landscape elements is calculated following Bayes' rule.

The presence of levee based on the occurrences of elevated area, irrigated land, arable land, dense vegetation etc. can be calculated by

It gives the posteriori probability using multiple evidences with the assumption that the evidences are not correlated. The probabilities on the right hand side can be calculated by counting the pixels from the digital map of a training area. In the present study, these probabilities are estimated visually from a map of known area assuming that the observations are not correlated. Although this assumption violates the real situation, it became possible to implement Bayesian formula in such a way that the a priori probability of a hypothesis (presence of levee) is updated after each observation (which can be TRUE, FALSE or UNKNOWN). In a available in a further stage, when a large amount of reliable field data is available in a GIS, these probabilities can be checked and implemented in a large table.

Conclusion
An expert system for ground water resources assessment could be successful if one does not over estimate its possibilities but applies it for relatively simple but time and cost saving items. It is difficult to build an expert system which gives detailed information about groundwater availability on site scale. Instead, it could be useful it can indicate the presence of groundwater in a subbasin. Areas with low probabilities can be excluded from further exploration. In groundwater assessment problem, the crucial parts is the classification of objects into various geohydrological units and estimation of corresponding probabilities for the presence of aquifer. Also the statistical relationships between the elements and classes may vary from place to place. Hence the system must be tested by local experts in representative areas and must be used in a restricted domain. In the current system, the objects are formed by the manual segmentation of remotely sensed images together with maps available in GIS. For each of these objects, the system determines the probability for the presence of groundwater. The system will be further modified to incorporate automated segmentation of digital Remote Sensing data, calculation of prior probabilities from maps available in GIS and creation of maps marking the probable areas for groundwater. The domain knowledge will also be modified taking into consideration the local expertise.

Reference

Davis,J.R. and Nanninga, P.M. (1985) "GEOMYCIN : Towards a Geographic Expert System, for Resource Management", Journal of Environmental Management, 21, 377-390.
Duda, R.O. and Gasching, J.G. (1981) "Knowledge Based Expert System Come of Age" BYTE Publications Inc. September.
McKeown, D.M, Harvey, W.A and McDermott, J. (1985) "Rule Based Interpretation of Aerial Imagery" IEEE Tran. On Pattern Analysis and Machine Intelligence , PAMI-1 (5), 510-585.
McKeown, D.M. (1987) "The Role of Artificial Intelligence in the Integration of Remotely Sensed Data with Geographic Information Systems", IEEE Trans. On Geosceince and Remote Sensing, GE-25 330-348
Mulder, N.J., Middelkoop, H. and Miltenburg, J.W. (1988) "Progress in Knowledge Engineering for image interpretation and classification ", ISPRS Congress, Kyoto.
Nagao, M. and Matsuyama, T. (1980) "A Structural Analysis of Complex Aerial Photographs", Plenum Press,, New York.
Peacegood, G. Fisher, P.F. and Wilkinson, G.G. (1986) "an Expert system to assist the Recognition of Aquifers in Remotely Sensed Images", Kingston Polytechnic, U.K.
Reeves R.G, Anoson, A. and Lande, D. 91983) "The Manual of Remote Sensing ", American Society of Photograpmmetry, Falls Church, Virginia.
Skidmore, A.K. (1989) "Expert System Classifies Eucalypts Forest Types Using Thematic Mapper Data and a Digital Terrain Model", Photogrammetric Engineering and Remote Sensing, 55, 1449-1464.