Integration of Cartographic Knowledge with Generalization Algorithms S. Kazemi, S. Lim, and L. Ge School of Surveying and Spatial Information Systems The University of New South Wales Sydney, NSW 2052, Australia E-mail:
[email protected] Abstract - For the automation of the map generalization process, it is necessary to integrate cartographers’ experience with the generalization operations within Geographical Information Systems (GIS). This paper provides a brief discussion on Expert System (ES) and its application in GIS with a particular emphasis on automation of map generalization. The discussion concludes with the fact that, although solving automated generalization problems could greatly benefit from such a combination of technologies, there have been no attempts to combine ES, GIS, and cartographers’ experience for a comprehensive evaluation of generalization systems and their performance. This will pave the way for developing an expert system that will help cartographers in choosing the appropriate techniques for map and database generalization tasks such as feature displacement.
Keywords: GIS, Cartographic Knowledge, Generalization Tools, Expert Systems, Road Network, Mapping, Database
I. INTRODUCTION During the two last decades, a number of commercial generalization systems have been developed but the literature review indicates that a critical evaluation and comparison of such systems have not been performed yet. In addition, the application of Expert Systems (ES) to generalization systems and cartographers’ experience has not been investigated thoroughly. This section is therefore intended to provide a brief review of ES in the light of Geographic Information System (GIS). ES has played an important role in automatic generalization in different cartographic software such as ArcGIS, DynaGen and LaserScan. Development of ES represents a major commercial application of Artificial Intelligence (AI) designated to enhance the quality and availability of knowledge for automated decision-making [1]. Also, ES is described as a system of software or combined software and hardware capable of competently executing a specific complex problem which has been performed by a human expert but requires significant expertise for their solution ([2] and [3]). ES accommodates a large amount of judgmental interpretation and heuristic knowledge or “rule of thumb”
which specifies a set of actions to be performed for a given situation. This is done through simulating the element of a human specialist's knowledge (e.g. cartographers or image interpreters) and reasoning that can be formulated into knowledge chunks typified by a set of facts and heuristic rules. In other words, ES is a communication device between the knowledge of an experienced user and the computer program in order to solve cumbersome problems. ES tries to reduce cost and time, but increase accuracy, stability and consistency. An example of the use of rule-based systems in mapping and computing field is feature extraction from remotely sensed data, detection of road networks ([4], [5] and [6]), or map generalization ([7] and [8]). Knowledge Based Systems (KBS) assists human with decision-making by applying probabilistic rules within a knowledge base to specific conditions. This is done through two main approaches in developing ES: 1) using programming language (LISP and PROLOG, FORTRAN, C, JAVA), and 2) using ES shell tools which requires minimal knowledge of any high level programming. Many ES tools were originally written in LISP, PROLOG, and FORTRAN, which have been recently recoded in C++ and VISUAL BASIC to improve their speed and increase portability [9]. Fundamentally, there are three knowledge levels for ES including facts, rules and inference engine. These can be translated into four main components of the ES development: knowledge acquisition, knowledge representation, inference engine and user interface that form the architecture of ES [10]. In reality, the use of expert system is the execution domain of AI ([5] and [9]). Basically AI covers three major industries, namely robotics, ES and natural language processing. Historically, AI began during the 1940s and 1950s with one of its objectives to make computers more intelligent and thus effective. One of the most mature areas of AI research and development is ES. Since early 1950s, AI community has focused on two main areas of R&D, i.e. cognitive science and search methods [11]. The cognitive modeling is of interest to map generalization context. The late 1960s witnessed an incremental shift from formal reasoning techniques to knowledge itself and combined with many applications such as engineering and medical science
[12] encompassing development of highly complex search strategies and inference engines. By early 1970s the knowledge base was separated from the inference engine that resulted in maturing of ES field. Around 1980, the AI community realized that ES architecture is built on structured knowledge in which KBS becomes a clear definition for ES developers and emerged into a wellorganized structure [13]. II. COMPONENTS OF EXPERT SYSTEMS ES consists of four main components including: (a) knowledge representation, (b) inference engine, (c) knowledge representation, and (d) user interface. These are briefly discussed here. In connection to this, a conceptual ES for road generalization is shown in Figure 1. The proposed plan of a conceptual ES forms a part of our generalization framework for semi-automated road network generalization.
Figure 1. Conceptual ES architecture for road generalization A. Knowledge Acquisition The knowledge acquisition module formulates storage and collection of facts from knowledge sources (e.g. textbooks, manuals, case studies, cartographers, technical publications, learning from examples, discovery) to software systems. This could include extraction of rules for problem solving from an expert and cartographer, and make it adequate for machine intelligence processing. The process of transferring knowledge from a cartographer to a map generalization software is not natural knowledge transfer process, whereas from lecturers to students, or from parents to children is natural. Several phases are noted in the knowledge acquisition stream such as knowledge elicitation, knowledge
extraction from the knowledge sources, knowledge encoding into symbolic form, knowledge based organizing, and modifying to gain the best performance [14]. In the context of digital cartographic generalization, it involves the examination and interpretation of manual generalization processes. Translating cartographer’s thoughts into a set of explicit and well-defined process is a major challenge [15]. Knowledge extraction approaches are sometimes termed as methods of inference. In [16], a knowledge extraction approach can be categorized into deduction, induction, intuition, heuristic, generate and test, abduction, default, auto-epistemic, non-monotonic, and analogy. This research will mainly build knowledge through interviews and surveys of cartographers to find the basis on which the generalization decisions can be made. This is the framework of knowledge engineering. It attempts acquiring from the cartographers with all the elements of heuristic experience. A major challenge in this process is intensiveness of the knowledge that is not well organized in a proper form. Therefore, a bottleneck in developing such ES is to compile it into a machine-readable format. In another work [8], it is noted the “knowledge acquisition bottleneck” in the field of cartographic knowledge acquisition. However, in [17] this problem is overcome by analyzing different types of knowledge involved in the cartographic generalization process. It is necessary to gauge the depth of knowledge, find the right amount of knowledge and accomplish map generalization via adding some learning abilities to the software and database system. It is noted that cartographic rules are as numerous, contradictory and often not formalized. This is because of knowledge mixture that is not easy to maintain and limits the comprehensibility of the reasoning done by the system. In [18], inadequate knowledge formalization was identified as a problem in successful implementation of generalization in geographic databases. Other researchers [19] also suggested evaluation of computer-aided generalization versus manual generalization is required. Often generalization algorithms ignore the role of cognitive issues through knowledge discovery techniques such as decision trees, fuzzy logic, data mining and neural networks to extract the hidden knowledge. Once knowledge is discovered it can represent in a suitable form to build an ES. Examples of generalization rules are: (a) contours never intersect, water bodies located in the bottom of valleys, (b) roads can crossing each other, and (c) symbology as well as coloring of land use data is best displayed with color-hue for visual perspective. These are used as general guidelines when a cartographer makes maps. But cognitive methods introduce more specific knowledge that is ignored in published generalization algorithms. Establishing cartographic rules dynamically provides a possible solution to automated generalization. Different tasks have different rules and knowledge base. Therefore, creating distributed knowledge base in regard to
symbolization, color schemes, layout, object displacement, etc is an essential part of this process. Our research is currently evaluating generalization tools and their functions in order to develop workflows and guidelines for generalization of road networks. The results need to be compared with maps of similar scales. Current generalization systems such as Intergraph’s DynaGen that formalized the learning process between cartographer and the generalization functions through interactive mode. The data analysis (e.g. cluster analysis) and decision making (e.g. the identification of critical points) are done visually. Drawback of this approach is its subjective nature of generalization, but it has been proved that it is superior to manual generalization in regard to processing time and efficiency. The manual generalization operations are implemented in the generalization systems as functions. An example of workflow for road data generalization is: (a) elimination – very short branches and unimportant roads to be eliminated, (b) simplification – the complexity and the amount of data representing the roads to be reduced, and (c) smoothing – the simplified roads to be smoothed to improve visual impression of the output. B. Knowledge Representation The knowledge representation brings collected knowledge into a suitable form such as decision trees [6]. For example, KnowledgeSEEKER algorithm [20] produces the knowledge from example data and represents it in decision trees, and offers the capability to convert it into both generic rules and programming statements. Rules are formulated as If ... Then, or If-And-Then statements into a knowledge base containing separate Condition using Boolean operators (And, Or, inequalities such as >,