(Portions of this paper to appear in "A Conceptual Framework for Incorporating Cognitive Principles into Geographical Database Representation" by J.L. Mennis, D.J. Peuquet, and L. Qian, forthcoming in International Journal of Geographical Information Science)
There is currently an unprecedented and increasing amount of available digital spatio-temporal data due to the growth of data gathering technologies and organizations. These data hold within them a vast store of information about the earth and its inhabitants, information that can provide insight into a variety of environmental, social, economic, and other processes and systems. The primary means of spatio-temporal data handling is through the use of geographic information systems (GIS). GIS provides spatial data management, analysis, and display functions and facilitates the integration of diverse sources of data, qualities that make GIS an ideal tool with which to explore the vast amount of digital data that is becoming available. However, due to the sheer amount and complexity of these data, as well as the sophisticated analytical demands of today's researchers, the conventional vector and raster GIS database models often prove inadequate as a means to support the types of analyses that are necessary to fully exploit the data resources now available (Burrough and Frank 1995).
Limitations of the conventional GIS data models include the inability to robustly represent spatial and categorical 'fuzziness' (Burrough and Frank 1996), change through time (Langran 1992), and composition and other semantic relationships between geographic entities. The roots of many of these representational deficiencies may be traced to the fact that the vector and raster data models were developed based on the graphic representation elements of the paper map and the traditional grid used in computer graphics, respectively. Subsequently, it has been widely recognized that improving GIS database representation beyond the limitations of the vector and raster data models is dependent on reducing the semantic gap that exists between the GIS database model and the way that people conceptualize geographic domains. In recognition of the need for more robust and human-oriented GIS database models, the University Consortium for Geographic Information Science (UCGIS) and the National Center for Geographic Information and Analysis (NCGIA) have targeted geographic cognition, and the incorporation of cognitive principles into GIS database representation, as primary research foci in the field of geographic information science (UCGIS 1996, Egenhofer et al. 1999, Mark et al. 1999).
While a number of authors have described the use of certain elements of cognition in GIS database modeling (e.g. Usery 1993) or developed domain-specific cognition-informed GIS database representations (e.g. Raper and Livingstone 1995), a generic and non-domain specific conceptual framework for integrating cognitive principles into GIS database representation has yet to be developed. It is the purpose of this research to develop such a framework. Note that the reason for devising this framework is not to create a GIS that 'thinks' like a human; rather it is to provide the basis for developing more robust and intuitive GIS database models and to facilitate GIS human-computer interaction.
Following this introduction, the paper proceeds as follows. Sections 2 and 3 review the literature concerning select elements of cognition and previous research in cognition and GIS database representation. Sections 4 and 5 describe the proposed cognition-informed database representation framework in detail and section 6 provides an example of how the framework may be applied to a real-world environmental application, the representation of storm phenomena. Plans for further research are presented in section 7.
2. Principles of Cognition
The field of cognition is clearly much too broad to be reviewed in depth in the context of this paper. Instead I focus on four principles or processes of cognition that I believe are particularly relevant to geographic database representation: 1) categorization, 2) the cognitive separation of 'where,' 'what,' and 'when' aspects of geographic knowledge, 3) the distinction between percept and concept, and 4) knowledge acquisition and refinement. Categorization concerns how humans group entities that are assumed to share common properties into categories in order to organize and most efficiently store knowledge. Of primary importance here is the idea that categories are cognitively arranged in a superordinate-subordinate hierarchy (taxonomy) that ranges from broad and inclusive categories to those that are more specific and exclusive, as in the categorical hierarchy: furniture, chair, and recliner (Rosch 1978). The most useful level in the hierarchy is termed the 'basic-level' (Rosch 1978). It is at this level that properties attributed to the specific categories most closely mirror those perceived in the real world. These properties are characterized by overall shape and motor interaction and are the most general level in the hierarchy at which a mental image can be formed. Thus, 'chair' is a basic-level category while 'furniture' is not. Basic-level categories are located in the middle of any given taxonomy, and are the first to be learned.
Our stored knowledge is, generally speaking, organized by kinds vertically within the hierarchy (taxonomies), and by parts (partonomies) within categories. At the basic level, our knowledge is mostly concerned with parts (Rosch 1978, Tversky and Hemenway 1984). Parts determine shape, hence shape plays a key role in categorization, e.g. chairs have seats. To increase the distinctiveness and flexibility of categories at all levels vertically in any hierarchy, categories tend to become defined in terms of prototypes that contain the attributes most representative of items within that category. The concept of schema addresses the cognitive use of a priori knowledge to interpret and categorize new observational data. Although there are many theories as to the nature of schema, a useful definition is that it is a set of information about a category that is used as a set of 'fuzzy' rules to discover new instances of the category (Thorndyke 1984).
These principles of categorization have also been demonstrated in a geographic context (Tversky and Hemenway 1983, Lloyd et al. 1996). Certainly, geographic entities also tend to be taxonomically arranged in nested hierarchies. For example, country, region, state, city, and neighborhood are conceptually arranged in a hierarchical order. There is also a container relationship built into this hierarchy. In other words, for any set of specific examples (e.g. the US, New England, Massachusetts, Boston, and Beacon Hill), each entity is spatially contained within the entity next above it in the hierarchy. A nested hierarchical structure of cognitive geographic knowledge has been experimentally confirmed by a number of researchers (Stevens and Coupe 1978, Eastman 1985).
There is also a body of neurological and psychological evidence suggesting that humans cognitively store 'what,' 'where,' and 'when' knowledge in separate categorical hierarchies, or knowledge structures, that capture differing characteristics for differing purposes (Ungerleider and Mishkin 1982, Farah et al. 1988, Kosslyn et al. 1992). Broadly speaking, 'what' knowledge concerns the identity of an entity, 'where' knowledge concerns relative spatial relationships, and 'when' knowledge concerns the detection of change, motion, and process (Peuquet 1988). It is the interaction between these knowledge systems that allows people to function in the world (Kosslyn and Koenig 1992, Landau and Jackendoff 1993). Of particular relevance here is the pioneering work of Marr (1982), who proposed an information processing view of vision in which visual interpretation of the environment is the process of abstraction from raw visual input (a location-based array of color values, i.e. 'where' knowledge), termed the image, to recognition of objects within the visual scene (identification of objects via categorization, i.e. 'what' knowledge). Objects, as concepts, are thus always higher-order information and can be interpreted using pre-existing knowledge, such as that which is stored in a schema.
Marr's theory of abstraction for visual interpretation is built upon the cognitive distinction between that which is perceived by the senses (percept) and that which is conceptualized in the mind (concept). Both philosophers (Kant 1950, Cassirer 1973) and cognitive psychologists (e.g. Neisser 1976) have noted that concepts, as knowledge about entities in the environment, are derived from percepts, sensory observations of the environment. The relationship between percepts and concepts is mediated by schema, which serve as generalized patterns of relationships, or rule-bases, for categories that facilitate the abstraction of observations into conceptual knowledge. As a simple example, consider the recognition of a geographic entity such as a summer thunderstorm. A layperson's schema for a summer thunderstorm may be that it is a brief and local event indicated by a dark and low cloud cover, heavy rainfall, high winds, and the presence of thunder and lightning. When one observes these properties in the environment (percepts), they are together identified as a thunderstorm entity (concept). Of course, each of these individual properties may themselves be considered a conceptual entity that is derived from 'lower-order' percepts (e.g. the conceptual object 'lightning' is recognized in the environment from the visual perception of color), indicating that the percept-concept division is relative to context.
We may consider the abstraction of concepts from percepts within a basic model of geographic knowledge acquisition and refinement (figure 1). Taxonomic and partonomic hierarchies facilitate the recognition of objects from sensory perception of the environment by providing schema for generic types of objects. These schema are used within a cycle of knowledge acquisition in which sensory observations (Marr's image, or percepts) are used to identify objects in the environment (higher order conceptual entities) through abstraction,which may then be used to modify and refine the original schema.
3. Related Research in Cognition and Geographic Database Representation
While the integration of semantics into non-geographic database representation has been an ongoing project since the advent of the relational data model (Hull and King 1987, Peckham and Maryanski 1988), there have been a number of recent studies that focus specifically on the integration of cognitive principles into geographic database representation. For example, Usery (1993) notes that theories of cognitive categorization can provide a useful model for development of a semantic 'feature-based' GIS. Of particular relevance to this research is the idea that database representation can be based on the decomposition of geographic information into space, time, and theme properties (Sinton 1978, Nyerges 1991), analogous to the 'where,' 'when,' and 'what' components of cognitive representation. Of equal importance is the idea that since cognitive representation (and acquisition) of geographic information is divided between sensory information (percept) and derived knowledge (concept), so too can geographic database representation be divided between spatio-temporal observational data and higher level geographic knowledge derived from that data (Rennison and Strausfeld 1995, Freksa and Barkowsky 1996).
Building on these themes, Peuquet (1994) proposed the Triad model for geographic database representation which decomposes geographic entities into interrelated location-based (where), time-based (when), and feature-based (what) information. Raper and Livingstone (1995) take a similar approach to develop a spatio-temporal database model that is specific to the representation of geomorphologic phenomena. Their data model represents a geomorphic phenomenon as a semantic decomposition into attributes derived from the basic-level categories form, process, and material. These attributes are referenced to a four-dimensional spatio-temporal framework with axes of location (in three dimensions) and time.
Other researchers have sought to improve geographic database representation using object-oriented data modeling. The object-oriented abstraction techniques generalization and aggregation are analogous (but clearly not identical) to taxonomic and partonomic cognitive categorization hierarchies, respectively. In geographic database research, these techniques have primarily been applied to the representation of complex geometry of geographic phenomena (e.g. Milne et al. 1993, Worboys 1994, Tang et al. 1996). However, because these approaches to object-oriented GIS are typically applied at the implementational level, as opposed to conceptual modeling level, they do not fully reap the benefits that the object-oriented modeling approach offers for semantic representation (Leung et al. 1999) and the integration of cognitive principles into database representation.
4. A Conceptual Framework for Cognition-Informed Geographic Database Representation
I describe here a conceptual framework within which to develop a geographic database model that is based on the principles of cognition and related research efforts in cognition-based geographic database representation. This framework, named the Pyramid framework, integrates object-oriented conceptual modeling and other knowledge representation techniques. It is intended to be non-domain specific (i.e. inclusive of a variety of geographic phenomena) and is based on reducing knowledge about a geographic phenomenon into perspectives based on the cognitive separation of 'what,' 'when,' and 'where' knowledge. The Pyramid framework also incorporates key cognitive structures including categorization, part-whole relations, and behavioral rules, in order to create a high-level semantic representation of an interrelated web, or system, of geographic phenomena. The Pyramid framework also integrates theories of geographic knowledge acquisition, from sensory perception to conceptualization, in order to separate the representation of geographic data from the representation of geographic knowledge that is derived from that data. Here, data can be defined as observational measurements that have been recorded in some way. Knowledge, on the other hand, refers to any and all contextual meaning that may be derived from data.
Based on the semantic difference between data and knowledge, the Pyramid framework is composed of two distinct, yet interrelated, components: the Data Component and the Knowledge Component (figure 2). The Data Component represents uninterpreted observational data, analogous to the idea of Marr's image or the cognitive percept, while the Knowledge Component represents the conceptual entities of the database user that may be captured by the data and the rules for their identification, analogous to the idea of schema or the cognitive concept.
4.2 The Data Component
Observed geographic data are formally represented within the Data Component by the location, time, and theme perspectives, analogous to the cognitive separation of 'where,' 'when,' 'what' knowledge. The location and time perspectives can be conceptualized using a geometric 'container' metaphor, in which space and time can be thought of as an absolute and enclosed 'space' within which geographic phenomena exist and processes and events take place. This is relatively intuitive - real world geographic entities extend across space and over time. The general notion of the Data Component can be understood by using the 3-D 'space-time cube' metaphor (figure 3) with location extending across the x and y axes and time extending along the t axis. This cube defines a set of spatio-temporal subspaces, or voxels, each with an assigned location and time.
A theme is defined here as a property that can be sensed, measured, and assigned a qualitative or quantitative value. Temperature is a good example of a theme, as it used here, because its measurement yields a numeric value for any point on the earth's surface. Although less intuitive than location and time, a 'theme space' can be conceptualized in a way similar to the space-time cube, such as in a multidimensional statistical space or in a 'feature space,' as is commonly referred to in the remote sensing literature, where each sensed band of electromagnetic energy is defined as one dimension in the feature space (Landgrebe 1998). The integration of the theme space with the space-time cube creates a multidimensional hypercube that is termed the data space. Each multidimensional 'voxel' within the data space is defined by the intersection of its location, time, and multiple theme values.
4.3 The Knowledge Component
The Knowledge Component represents derived knowledge that is analogous to the recognition of objects embedded within the visual image (i.e. the observational data) and the classification of those objects into categories. It is in the Knowledge Component that the three perspectives of observational data come together to form a semantic object that is of interest to the user, hence the graphic pyramid structure and the name of the framework. Note that the term 'object' is used here to refer to a geographic conceptual entity that has a unique and coherent identity, and is related to a specific combination of observational data stored in the location, time, and theme perspectives.
In order to represent semantic properties associated with an object, and to link the object with its observed properties stored within the data space, an object is represented within the Pyramid framework through the use of an object template. To allow maximum flexibility in the types of information that can be recorded, as well as for the sake of uniformity throughout the Knowledge Component, we use a frame type of structure (Minsky 1975).
Figure 4 shows an example of a 'blank' object template where the attributes are defined as follows:
As can be seen, the template for an object stores only those values that are critical in the history of the object, as well as those attributes and attribute values that describe an object as a whole. The function of the object template is not to exhaustively describe the object. Rather, the object template serves two interrelated functions: first, to provide a guide for finding the location, time, and theme history of the object within the Data Component, when desired, and, second, to represent enough information to define the object's unique identity, including its relationships with other objects and its behavioral characteristics. These relationships and how they function are defined below in relation to the storage of classes.
Classes are represented using a class template. This is a frame structure similar to an object template, but the attributes of a class record ranges of values that define the criteria for membership in the given class rather than values that record the observed character of an object. In this way, class templates play a role analogous to that of schema in the cognitive process of object recognition. The ranges of attribute values described in the class template serve as a rule-base for the interpretation of the data stored in the Data Component. These are used in discovering specific instances of that class, i.e. objects. For example, while an object has the attribute 'lifespan', a class has the attributes 'maximum lifespan' and 'minimum lifespan'. All objects that are members of that class are noted to have an actual lifespan within that maximum and minimum range to be a class member. Additional information may be included to note whether the stored attribute value ranges are required or simply expected for members of that particular class.
In addition to the location, time and theme information that describes the attributes of a class, the class template contains slots for attributes that are not directly related to the Data Component but concern the semantics of the object category. This includes information that describes the behavior of an entire class. These include the behavior 'becomes' - what other type of object objects of this class can change into, thereby undergoing a change of identity. An example of this would be that the class 'tropical storm' becomes 'hurricane'. Of course, not all tropical storms become hurricanes, but this establishes a potential behavioral path for that class. The co-occurrences relation defines what other types of objects are typically spatially and/or temporally coincident with objects in this class.
The class template also includes information on the interrelationships of the various classes that mimics cognitive categorization structures; each class may have taxonomic 'is-a-kind-of' and 'has-kinds' relationships. This allows each class to be categorized into a higher-level superclass, which in-turn may be categorized into another higher-level class, and so on, forming a generalization, or taxonomic, hierarchy. In addition to generalization relationships, classes can have aggregation relationships. Some real-world geographic entities are part of other, larger and/or longer-lasting, geographic entities. To capture this, classes can have partonomic 'is-a-part-of' and 'has-parts' relationships with other classes. This relationship property indicates that the objects that are members of that class can have aggregation relationships with objects in other classes.
Generalization and aggregation relationships, and behavioral properties, exist not only for classes but also for specific objects and they are therefore represented in the object template (figure 4). In the object template, the 'is-a-part-of' and 'has-parts' relationships indicate composition between specific objects (as opposed to the general and potential aggregation relationship indicated for classes). An object also has an 'is-a-kind-of' relationship with the class of which it is a member (by definition an object cannot have a 'has-kinds' relationship with any other object). The behavioral attributes listed in the object template, such as 'becomes,' 'evolves-from,' and 'co-occurrences,' also refer to other specific objects.
Note that some objects can be defined entirely from other objects and not defined directly from the data space. Objects that are derived directly from their observation in the data space are termed 'atomic;' those objects that are composed solely of other objects are termed 'composite.' Note that the terms 'atomic' and 'composite' are not used in the conventional sense to refer to primitive and complex data types, such as integers and arrays, respectively. Rather, the terms refer to an object as the term 'object' is used here - a conceptual entity derived from the interpretation of observational data. Like the generalization hierarchy, the aggregation relationship can be extended from the simple atomic-composite object relationship to a hierarchy of composite objects. Many composite objects may have 'part-of' relationships with another composite object, forming an aggregation hierarchy, or partonomy.
So far we have described two separate and parallel hierarchies built upon generalization and aggregation. These two hierarchies are interrelated - each class in the generalization hierarchy may have one or more superclasses and/or subclasses, and also may have a set of objects (instances). Each object, whether atomic or composite, is a member of one or more particular classes in the generalization hierarchy. Class behavior and relationship attributes are passed down to subclasses and member objects through inheritance. This approach ensures that information that is common to an entire class does not have to be redundantly stored in each object template for every member object.
5. Data Exploration
Data exploration takes place through iterative refinement of the template rule-base and relationship structures and the subsequent visualization and statistical analysis of the spatio-temporal and attribute characteristics of the semantic objects (and observational data) at each iteration. Note that the original observational data are never altered; the user explores the data by altering the conceptual model, or semantic representation, that is built 'on top' of the observational data in the Knowledge Component. The Data and Knowledge Components represent levels of abstraction and facilitate a 'cycle' of knowledge development that is analogous to how humans acquire and modify geographic knowledge, as is illustrated in figure 1. Observational data about the world is stored in the Data Component, then classified and interpreted to derive meaningful information that is stored in the Knowledge component. This knowledge may then be used to further interpret the observational data.
6. An Example: The Representation of a Storm System
A representation for a particular type of storm system, a mesoscale convective complex (MCC), is described here in order to demonstrate the representational power of the Pyramid framework. MCCs are regional scale storm systems (as opposed to larger, synoptic scale phenomena) that are composed of many individual storm cells that can act in concert to create particularly violent storms (Maddox 1983). These typically 'egg' shaped storms often occur at night in the Midwestern US during the summer or early fall. They can produce copious amounts of rain and are often associated with flooding and other climate hazards. The representation of MCCs and their derivation from observational data are particularly appropriate for this conceptual modeling task because their identification is based on a set of specific parameters derived from analysis of remote sensing data, analogous to the extraction of objects from a data space described here.
The first step in developing the conceptual representation of an MCC is to define the Data Component. For this purpose, consider a data space with four dimensions. The two location dimensions define the Midwestern region of the US, regularly partitioned into locational cells 1000 meters on a side. The time dimension defines the year-long period beginning January 1, 1998 and ending December 31, 1998, regularly partitioned into hour long intervals. For simplicity, we define only one theme, cloud top temperature. Cloud top temperature is derived from satellite imagery and indicates the altitude of the cloud top (the higher the altitude, the lower the temperature), which in turn can indicate the stage of development or severity of a storm (Carleton 1991).
Development of the Knowledge Component in this particular case proceeds through the use of expert knowledge. As such, we begin by defining the MCC class template and its attributes which then later act as criteria for identifying actual MCC objects within the data space. The MCC class is described by its class template (figure 5) which shows the MCC class attributes and their respective values. The criteria listed here are derived from Carleton (1991):
Note that while the minimum and maximum shape eccentricity attributes may not necessarily be included in every class template (as the attributes minimum and maximum lifespan are), they are included here because eccentricity is essential for the identification of MCC objects. The class template also records the generalization and aggregation relationships of the MCC class: an MCC is-a-kind-of larger category of storm, mesoscale convective system (MCS), and may be part-of an object that represents a specific type of synoptic regime associated with a large-scale front (Maddox 1983). In addition, the class template lists other attributes of MCCs: they can evolve from squall line storms, they typically occur at night, and they are often associated with a lower level jet stream (Carleton 1991).
The criteria captured within the MCC class template acts as a rule-base for a query on the data space to identify individual MCC objects. The query procedure to identify MCC objects steps through each of the class attributes described in figure 5 to select the portion of the data space that satisfies the class membership criteria. The following query procedure (described simply in plain English, as opposed to a formal query language) serves as an example of how MCC objects may be identified from the data space. Note that each selection operation described below is applied only to the objects that were selected in the previous selection operation.
Theme value criteria:
Select only those voxels with a cloud top temperature < -52° C
Group all contiguous selected voxels into individual (temporary) potential MCC objects
Calculate the area of each potential MCC object
Select only those potential MCC objects with an area between 50,000 and 350,000 square kilometers
Calculate the major and minor axes for each potential MCC object
Calculate the eccentricity (minor/major axis) of each potential MCC object
Select only those potential MCC objects with an eccentricity >= 0.7
Querying based on the duration criteria can be complex. In order to calculate the duration of each potential MCC object, it is necessary to maintain the identity of each potential MCC object over time. If an object is stationary, location can be used to identify the same object over time; however, MCCs are not stationary but often move. For the purpose of this example, it is possible to maintain the identity of each potential MCC object through time by simply stating that if the centroid of a potential MCC object is located within 200 kilometers of the centroid of another potential MCC object at an adjacent time interval (i.e. either the prior or following hour), it is recognized as the same potential MCC object that has moved over time. This method assumes that MCCs do not move (planimetrically) at a rate that is greater than 200 kilometers per hour, which is typically the case. Using this approach to maintaining identity over time, the duration criteria may be satisfied by selecting all potential MCC objects that exist (maintain their identity) over a duration of greater than six hours.
Each of the potential MCC objects that meet all the criteria listed above (theme, size, shape, and duration) are then defined (instantiated) as actual MCC objects. All MCCs that occurred over the United States during 1998 are represented as individual MCC objects with associated location, time, and theme data captured as a portion of the data space. Each MCC object also has an object template that captures the important attributes of that particular MCC: its minimum and maximum cloud top temperature, its minimum and maximum size, its duration, etc. As an example, figure 6 shows the object template for a particular MCC object that occurred the early morning of August 23, 1998. Other attributes included in the object template are the MCC object's 'is-a-kind-of' relationship with the MCC class and its 'is-a-part-of' relationship with a particular large-scale front object, as well as its relevant co-occurrences, such as the fact that it occurred during the night. Note the difference between the class template (figure 5) and this object template (figure 6). The two templates are similar, however the class template stores attribute values that are common to the entire class (including ranges of values for certain attributes) while the object template stores attribute values that describe an actual geographic entity.
This example is relatively simple and straightforward. MCCs are modeled here as atomic objects, using only one theme, and with only one higher level of generalization and aggregation. One can imagine, however, that these same principles can be used to develop a much more complex storm system representation if other data, such as precipitation rate and air pressure, and a richer hierarchy of storm types, such as single-cell storms and squall lines, were integrated into the representation. Further, note that the structure of the Knowledge Component is relatively flexible. This representation of MCCs in the Midwestern US during 1998 may be refined from both additional observational data and from declarative knowledge generated by domain experts. By altering the attributes (criteria) of the MCC class, new sets of MCC objects can be generated, providing a variety of different structural interpretations of the climate data for different application contexts.
7. Future Research
The Pyramid framework is intended to be a 'jumping off' point for developing and implementing a cognition-informed GIS database model. Current efforts focus on using the object-oriented visual modeling language UML (Universal Modeling Language) to design and specify the Data Component and Knowledge component. This cognition-informed database model will be developed using an object-oriented database platform. While it is possible to transform the representational devices of the Pyramid framework into a relational or object-relational database model, an object-oriented database best supports the Pyramid framework structure. In order to evaluate the utility of the proposed database model, a case study database implementation will focus on the spatio-temporal analysis of storm events in the Susquehanna River Basin in central Pennsylvania using an approximately 1.7 gigabyte spatio-temporal meteorological data set. User interaction with the database will be provided through the GeoVista Studio, a stand-alone suite of visualization and geocomputation tools being developed at Pennsylvania State University.
Many of the ideas presented here are built upon the research of Donna Peuquet, who contributed a great deal to the development of this paper and whose guidance is greatly appreciated. I would also like to thank Liujian Qian, Agnar Renolen, Monica Wachowicz, and Alan MacEachren for sharing their ideas concerning this research. This research was supported in part by US Environmental Protection Agency grant R825195-01-0.
Burrough, P.A. and Frank, A.U., 1995, Concepts and paradigms in spatial information: are current geographical information systems truly generic? International Journal of Geographical Information Systems, 9(2), 101-116.
Burrough, P.A. and Frank, A.U. (eds), 1996, Geographic Objects with Indeterminate Boundaries (London: Taylor and Francis).
Carleton, A., 1991, Satellite Remote Sensing in Climatology (London: Bellhaven Press).
Cassirer, E., 1973, The Philosophy of Symbolic Forms III: The Phenomenology of Knowledge (New Haven: Yale University Press).
Eastman, J.R., 1985, Graphic organization and memory structures for map learning. Cartographica, 22(1), 1-20.
Egenhofer, M.J., Glasgow, J., Gunther, O., Herring, J.R., and Peuquet, D.J., 1999, Progress in computational methods for representing geographical concepts. International Journal of Geographical Information Science, 13(8), 775-796.
Farah, M.J., Hammond, K.M., Levine, D.N., and Calvanio, R., 1988, Visual and spatial mental imagery: dissociable systems of representation. Cognitive Psychology, 20, 439-462.
Freksa, C. and Barkowsky, T., 1996, On the relations between spatial concepts and geographic objects. In Geographic Objects With Indeterminate Boundaries, edited by P.A. Burrough and A.U. Frank (London: Taylor and Francis), pp. 109-121.
Hull, R. and King, R., 1987, Semantic database modeling: survey, applications and research issues. ACM Computer Surveys, 19(3), 201-260.
Kant, I., 1950, Critique of Pure Reason (New York: Humanities Press).
Kosslyn, S., Chabris, C., Marsolek, C., and Koenig, O., 1992, Categorical versus coordinate spatial relations: computational analyses and computer simulations. Journal of Experimental Psychology: Human Perception and Performance, 18(2), 562-577.
Kosslyn, S.M. and Koenig, O., 1992, Wet Mind: The New Cognitive Neuroscience (New York: The Free Press).
Landau, B. and Jackendoff, R., 1993, "What" and "where" in spatial language and spatial cognition. Behavioral and Brain Sciences, 16, 217-265.
Landgrebe, D., 1998, Information extraction principles and methods for multispectral and hyperspectral image data. In Information Processing for Remote Sensing, edited by C.H. Chen (River Edge, NJ: World Scientific Publishing).
Langran, G., 1992. Time in Geographic Information Systems (London: Taylor and Francis).
Leung, Y., Leung, K.S., and He, J.Z., 1999, A generic concept-based object-oriented geographical information system. International Journal of Geographical Information Science, 13(5), 475-498.
Lloyd, R., Patton, D., and Cammack, R., 1996, Basic-level categories. Professional Geographer, 48(2), 181-194.
Maddox, R.A., 1983, Large scale meteorological conditions associated with mid-latitude, mesoscale convective complexes. Monthly Weather Review, 111, 1475-1493.
Mark, D.M., Freksa, C., Hirtle, S.C., Lloyd, R., and Tversky, B., 1999, Cognitive models of geographical space. International Journal of Geographical Information Science, 13(8), 747-774.
Marr, D., 1982, Vision (New York: W.H. Freeman Co.).
Milne, P., Milton, S., and Smith, J.L., 1993, Geographical object oriented databases - a case study. International Journal of Geographical Information Systems, 7(1), 39-55.
Minsky, M., 1975, A framework for representing knowledge. In The Psychology of Computer Vision, edited by P.H. Winston (New York: McGraw-Hill), pp. 211-277.
Neisser, U., 1976, Cognition and Reality (New York: W.H. Freeman and Company).
Nyerges, T.L., 1991, Geographic information abstractions: conceptual clarity for geographic modeling. Environment and Planning A, 23, 1483-1499.
Peckham, J. and Maryanski, F., 1988. Semantic data models. ACM Computing Surveys, 20(3), 153-189.
Peuquet, D.J., 1988, Representations of geographic space: toward a conceptual synthesis. Annals of the Association of American Geographers, 78(3), 375-394.
Peuquet, D.J., 1994, It's about time: a conceptual framework for the representation of temporal dynamics in geographic information systems. Annals of the Association of American Geographers, 84(3), 441-461.
Raper, J. and Livingstone, D., 1995, Development of a geomorphological spatial model using object oriented design. International Journal of Geographical Information Systems, 9(4), 359-383.
Rennison, E. and Strausfeld, L., 1995, The millenium project: constructing a dynamic 3D virtual environment for exploring geographically, temporally and categorically organized historical information. In Spatial Information Theory: A Theoretical Basis for GIS, edited by A.U. Frank and W. Kuhn (Berlin: Springer-Verlag), pp. 69-91.
Rosch, E., 1978, Principles of categorization. In Cognition and Categorization, edited by E. Rosch and B. Lloyd (Hillsdale, N.J.: Lawrence Erlbaum Associates), pp. 27-77.
Sinton, D.F., 1978. The inherent structure of information as a constraint to analysis: mapped thematic data as a case study. Harvard University, Cambridge.
Stevens, A. and Coupe, P., 1978, Distortions in judged spatial relations. Cognitive Psychology, 10, 422-437.
Tang, A.Y., Adams, T.M., and Usery, E.L., 1996, A spatial data model design for feature-based geographical information systems. International Journal of Geographical Information Systems, 10(5), 643-659.
Thorndyke, P., 1984, Applications of schema theory in cognitive research. In Tutorials in Learning and Memory: Essays in Honor of Gordon Bower, edited by J. Anderson and S. Kosslyn (San Francisco: W.H. Freeman), pp. 167-192.
Tversky, B. and Hemenway, K., 1983, Categories of environmental scenes. Cognitive Psychology, 15, 121-149.
Tversky, B. and Hemenway, K., 1984, Objects, parts, and categories. Journal of Experimental Psychology: General, 113(2), 169-193.
UCGIS, 1996, Research priorities for geographic information science. Cartography and Geographic Information Systems, 23(3), 115-127.
Ungerleider, L.G. and Mishkin, M., 1982, Two cortical visual systems. In Analysis of Visual Behavior, edited by D.J. Ingle, M.A. Goodale, and R.J.W. Mansfield (Cambridge, MA: MIT Press), pp. 549-586.
Usery, L., 1993, Category theory and the structure of features in geographic information systems. Cartography and Geographic Information Systems, 20(1), 5-12.
Worboys, M.F., 1994, Object oriented approaches to geo-referenced information. International Journal of Geographic Information Systems, 8(4), 385-389.