With the development of artificial intelligence (AI) based technologies in recent years, new opportunities have emerged to enhance the tools we use to process spatial data. When combined with the data storage and processing capabilities of Geographic Information Systems (GIS), artificial intelligence tools promise to provide landscape planners with new analytical and modelling capabilities. This merging of technologies is still in the early stages. Successful connections are being developed between geographic information systems and a variety of artificial intelligence based packages including Artificial Neural Networks, Genetic Algorithms, and Object Oriented Modelling programs. But little is known of the potential benefits and pitfalls of these technologies in the context of landscape planning applications.
This paper traces the development of an AI tool known as the Artificial Neural Network, a computational system based on the information processing mechanism that is thought to occur in biological nervous systems. The utility of this technology as a tool for landscape planners or architects is addressed, including a demonstration of its applicability in illustrating dynamic landscape processes and human/environment relationships in an educational setting.
Landscape Planners and others, who perform geographic analysis in support of a decision making process, are required to interpret a wide variety of data in spatial or non-spatial formats. When conducting planning exercises involving suitability analysis, planners may utilize a variety of techniques, such as the interpretation of spatial data from air photographs and maps or non-spatial data from census rolls, to discern significant issues surrounding proposed uses of the landscape. Such manual techniques, when combined with a knowledge of the processes of nature and the increasing requirements of our society, constitute an important analysis tool within landscape planning theory.
One common activity in the field of landscape planning is the development of land-use suitability maps. A suitability map shows the spatial pattern of requirements, preferences, or predictors of some activity (Hopkins 1977). Suitability maps may convey information regarding the market, non-market, or nonmonetary costs or impacts of a wide range of activities such as residential development or agricultural land use. The analysis process requires that the planner both identify parcels of land which can be considered homogeneous within the context of the study and develop a procedure for rating these parcels. Hopkins (1977) provides a comprehensive review of the different methods that have been used to combine factors, or themes, during suitability analysis. These methods can be grouped into a number of categories which include, Gestalt, Mathematical Combination, Identification of Regions, and Logical Combination. However, frequently these methods are utilized in suitability analysis exercises without proper consideration of their validity in a particular application.
Hopkins identifies three methods of mathematically combining spatial variables expressed in different factors as ordinal, linear, or nonlinear. Under the ordinal combination method, different factors are identified and individually classified as to the relative suitability of their types for a particular land use. Once each factor has been ranked individually, the factors are combined, or overlaid, to produce a final composite suitability map. The overlaying operation is mathematically equivalent to addition. However, addition is an invalid mathematical operation in this case because the assumed properties do not hold (Hopkins 1977). Ordinal combination does not address the potential interdependence of different factors. Suitability may be a nonlinear and nonseparable function of the combination of types and not simply the sum of suitabilities of individual types. The linear combination method, in which the measurement problems of the ordinal combination method are overcome by assigning weights to the different factor values, still has a problem handling the interdependence of factors. This problem of interdependence can be handled if the combination equation is not linear, as is the case with the non-linear combination method (Hopkins 1977). But in order to execute the non-linear combination method, the relationships between different factors must be known and must be able to be expressed mathematically, Hopkins reports a number of cases in which non-linear combination has been utilized successfully for components of a suitability exercise, such as in the calculation of potential soil loss. But according to Hopkins, developing an overall land suitability approach using non-linear combination suffers from the difficulty of operational implementation.
According to Hopkins, the identification of regions method overcomes problems of interdependence by first identifying the homogeneous regions that result from each combination of factors, and then assigning a suitability rating to each combination. But as the number of factors and variables that are used in the analysis process increase, it becomes increasingly more difficult to understand and model for all possible combinations when performing a task such as suitability analysis (Hopkins 1977). For example, 7 factors with 4 types each produce 16,384 possible combinations. Although the number of distinct regions that would occur in a real world analysis is only a fraction of this number, even a few hundred categories would be unmanageable. Simplifying the process by aggregating combinations of variables, as is done in cluster analysis, still assumes knowledge of all potential combinations (Gimblett et al. 1994).
Burley and Brown (1995) outline a method for reducing the complexity associated with combining a large number of data layers. They utilized a data reduction method called principal component analysis (PCA) to search for simplified explanations indicating the latent structure of the landscape. By utilizing PCA, fifteen suitability overlays were reduced to seven dimensions containing 65% of the original data. Although PCA did not render a simple elegant solution in this application, it did reduce the complexity associated with combining many suitability maps (Burley and Brown 1995). Additional exploration of the capabilities and limitations of this approach is certainly warranted.
As new technologies for handling large and diverse geographic data sets emerge, the difficulty of manually developing rules to handle a large number of factors renders many traditional techniques unmanageable. What is needed is a modelling technique that is capable of adaptive autonomous rule generation for handling a large number of different combinations of interdependent factors. Neural Network technology, in combination with geographic information systems, provides this capability (Gimblett et al 1994). Later in the paper we will detail how this technology can be applied.
The recent explosive development of landscape planning related technology, from remote sensing and digital database development to spatial analysis, has created new opportunities for geographers and planners. Computer based Geographic Information Systems (GIS) provide the user with a variety tools with which she can manipulate large digital databases to execute simple models and display the results. Other technological advances in computer science, such as Artificial Intelligence (AI) are providing new tools with potential applications in spatial data analysis and theory development. These technologies not only promise to add accuracy and speed to the existing landscape planning process, but also to provide a medium for the development of new theory and techniques.
We have been witnesses to the incredible changes that computer related technologies have brought to our personal and working lives in the last twenty years. These technologies have greatly increased the speed and efficiency of many work related tasks. Artificial Intelligence promises to further improve the effectiveness of these tools by bringing a degree of intelligence to many computer based technologies. Initial steps in this direction have been made in such fields as robotics, natural language processing, and medical diagnosis (Charniak and McDermott 1985).
Artificial Intelligence has been defined as the study of mental faculties through the use of computational models (Charniak and McDermott 1985). The word intelligence can be misleading because it implies the creation of a computer that thinks and acts as we do. But the creation of a computer which acts like a human is only a theoretical goal which has little bearing on most current artificial intelligence research. Much of the current AI research is devoted to making computers do things that are automatic for us, such as vision or voice and object recognition. At this point in time, computers and humans are capable of performing quite different tasks very efficiently. While computers can perform mathematical calculations much more efficiently than we can, humans are far better at such tasks as vision and pattern recognition. By developing these human capabilities even in a limited way and combining them with the powerful computational capabilities of computers, it is possible to create AI tools with useful applications.
Research in AI has progressed in numerous directions and resulted in the development of a variety of commercially available software packages including Artificial Neural Networks, Genetic Algorithms, Expert Systems, and Cellular Automata. Some of these packages have already been utilized in spatial modelling applications. For example, cellular automata have been utilized in combination with a GIS database to model changes in residential settlement patterns over time (Gimblett 1989, Deadman et al 1993). All of these new applications have promising applications in landscape planning and analysis. However in the rest of this paper, we will focus on one of these applications which offers a great deal of potentialfor the analysis of spatial data, Artificial Neural Networks.
Artificial Neural Networks, also known as connectionist models or simply neural nets, refer to a collection of computational systems whose architecture is inspired by the structures and processes that are thought to occur in biological nervous systems. The first theories of neurocomputing were published over fifty years ago by such authors as Donald Hebb, Frank Rosenblatt, and Jon von Neumann who were interested in the development of brain inspired computers. A flurry of early theoretical investigations led to the development of the first neurocomputers in the 1950's (Openshaw 1993). In 1959 Bernard Widrow developed a linear element model called adeline. Widrow utilized adeline, and the second generation model madeline, in such applications as weather forecasting and adaptive control (DuBose and Klimasauskas 1989).
Research continued on these early neural models until the publication of a book by Marvin Minsky and Seymour Papert in 1969 which criticized these early neural networks. In their book, they proved mathematically that the neural network known as the perceptron could not implement the exclusive-or logical function (Openshaw 1993). Following this criticism, neural networks were temporarily discredited. Research was placed on the back burner throughout the 1970's as scientists investigated the utility of other forms of artificial intelligence such as expert systems.
Neural Nets were revived in the early 1980's by a number of researchers who developed more advanced multi-layered networks which overcame the criticisms put forth by Minsky and Papert in 1969 (Openshaw 1993). Since that time neural network research has progressed steadily. The development of commercially available, user friendly neural net packages in the late eighties led to an explosion of studies which investigated their utility in a wide variety of applications. Neural networks have been applied in such diverse fields as voice recognition and astronomy. Applications in landscape related research have included; modelling scenic beauty from extracted landscape attributes (Bishop 1994), suitability analysis for forest management (Gimblett et al 1994, 1995, and suitability analysis for development (Sui 1992). Numerous publications and conferences have addressed theoretical and applications based issues surrounding neural networks.
The basic processing element of the Neural Network is known as the neuron or node. This processing element corresponds in theory to the individual neurons of the nervous system. The processing element is broken into two components, the internal activation function and the transfer function (DuBose and Klimasauskas 1989). Internal activation can be calculated in a number of ways. But it typically operates through a summation type function which adds up the values of incoming messages multiplied by a specific connection weight. The resultant output of the internal activation is sent to the transfer function which determines whether or not the processing element will send an output message.
Processing elements or nodes are usually grouped together in a network of layers for the processing of data (Openshaw 1993). This architecture typically consists of an input layer, to which data to be interpreted is sent, one or more hidden layers, and an output data, onto which the correct interpretation is mapped for later use by the user. Typically, every node in a particular layer receives input from all the nodes in the layer below it and/or sends its output, it makes one, to every node in the layer above. The net may be set up so that competition between the nodes of a particular layer determines that only one node will send a message to the next layer.
Neural Networks must be trained by being shown a series of examples which include samples of potential input values and the appropriate output values. The training data may be sent through the network one time only or over and over again until a calculated error value drops below a certain value. The user may select one of a number of learning rules which adjust the weights of the individual nodes during training until the output of the net matches the proper output in the training examples. It is important to ensure that the training examples given to the net both cover the range of possibilities and are representative of the frequency of occurrences of the data (DuBose and Klimasauskas 1989).
Once the network has been trained on the training set to respond with the correct output to a variety of inputs, then the net can be fed the full data set for interpretation. The network will go through the complete data set only once, imputing each set of data and providing the appropriate output. Neural Networks can be trained to interpret data in a variety of forms including, categorical data, discrete data or continuous data. Neural networks are able to handle data which is noisy or incomplete.
Neural networks have not always enjoyed the widespread interest that they do today. Only in the last ten years or so have researchers recognized the inherent advantages of neural networks. For a period of time in the sixties and seventies, neural nets sat largely undeveloped while AI research proceeded in other fields such as Expert Systems. The Expert System, or knowledge based system, is a computer program that mimics the human reasoning process, relying on logic, belief, rules of thumb, opinion, and experience (Plant and Stone 1991). In the rule-based expert system, the knowledge and experience of the expert are captured in a series of if-then rules which are used to solve problems (Plant and Stone 1991). In the expert system program a central inference engine draws on the rules that reside in a knowledge base to interpret information provided by the user and develop a solution. These systems were developed and used widely in such applications as disease diagnosis in medicine and pest management in agriculture. Expert systems have been utilized in landscape research applications such as the interpretation of Scenic Beauty based on the attributes of photographs (Buhyoff et al. 1994).
As expert systems developed, literature began to appear discussing the limitations that researchers had encountered. In particular, expert systems have a difficult time handling noisy or incomplete data. It was also found that some human knowledge is inexpressible in the form of rules and sometimes may not be understandable even though it can be expressed (Hoffman 1987). In addition, most human experts have difficulty expressing the tacit knowledge that they have acquired explicitly and completely. The difficulties associated with the development and implementationof expert systems has lead many researchers to begin exploring other techniques (Sui 1992).
Contrasted with the top down knowledge based approach of expert systems, neural networks take a bottom up data based approach to pattern-information processing. When relationships are understood and can be clearly articulated and computed, expert systems are appropriate. But when relationship requires a complex mathematical model that has not been developed yet, or when the relationship can be stated in general terms but is difficult to compute, neural network based models are appropriate (DuBose and Klimasauskas 1989).
The early stages of the landscape planning process require the collection and interpretation of landscape data in a number of themes. Such themes often include data pertaining to slope, aspect, soils, vegetation, and a variety of human modifications to the landscape. These diverse themes must often be interpreted in relation to one another in order to determine patterns in the landscape that are of interest to the goals of the planning exercise.
Traditional planning techniques such as map overlay analysis provide planners with a framework for relating these diverse themes to one another in reference to a specific question. Modern Geographic Information Systems even provide rudimentary modelling procedures, such as overlaying, for use with digital data. These early techniques provided planners with valuable if limited tools for landscape analysis. But the advent of AI technologies now presents us with the opportunity to perform more sophisticated analysis. In landscape research, neural networks have been employed in a variety of applications. Generally, the neural network is trained to respond to a variety of input variables, of spatial or non-spatial form, with an appropriate output which may take the form of a suitability rating, a vegetation classification, an estimate of recreation use,or a scenic beauty score. Examples of these applications are outlined below.
Sui (1992) utilized a back-propagation network to analyze the suitability of a number of land parcels for development. He compared the neural net based approach to a more traditional cartographic modeling based technique. Sui concluded that the neural net based approach was, inherently more objective and better able to handle noisy or missing data than the traditional approach. However, some questions regarding the neural net technique, particularly relating to the selection of an appropriate net architecture for the problem at hand, still remain. Wang (1992) utilized a neural network in a similar application to determine agricultural land suitability. The network interpreted input data in the form of numerical values, ordered textual symbols, and unordered textual symbols to determine the correct suitability class. Wang found the neural net to be an effective tool for agricultural suitability analysis in a GIS context.
Researchers with the Forest Service are utilizing neural networks in models to forecast occurrences of overnight stays in the National Park Service system on a monthly basis, and the USDA Forest Service system on an annual basis (Pattie 1992). The neural network is trained on data which includes such variables as monthly climatic data, stream flow, U.S. Forest Service timber production, and historical trends in threatened or endangered species. In the future, the generalization of knowledge about backcountry recreation patterns to a broader class of management situations such as optimizing recreation carrying capacity, maximizing biological diversity and wildlife habitat represents an important process in learning from examples presented to neural network models (Pattie 1992).
Neural Network applications are being increasingly utilized by researchers in landscape analysis applications. But if Neural Networks or other AI technologies are to be widely adopted for use in landscape planning, the results of this research must be effectively disseminated to practitioners and students. This will require not only communication of the techniques themselves, but also a clear understanding of the role that AI based analysis can play in the landscape planning process.
At the University of Arizona, AI applications for geographic analysis are taught as a course in the School of Renewable Natural Resources, separate from the Landscape Architecture curriculum. At the present time the course is focused primarily at graduate students in Landscape Architecture or Renewable Natural Resources. The course covers the theoretical and application based issues of a variety of different AI techniques including, neural networks, genetic algorithms, and expert systems. Students are exposed to these technologies through a hands-on approach in which at least half the course time is spent in the computer lab. In previous years, students have applied AI technologies in a wide range of studies including, processing remotely sensed vegetation data with a neural network, determining patterns in pest vegetation growth over time based on a neural net based analysis of environmental variables, hydrological modelling, and the development of an expert system for assessing wildlife habitat.
In the educational environment, AI applications can be taught on there own, or as a component of GIS based planning courses. A course of this nature requires that students have previous exposure to a Geographic Information System. When incorporated in a Landscape Architecture curriculum, students should also have prior exposure to landscape planning concepts. The nature of the prerequisites and the advanced nature of the subject matter means that a course of this nature is better suited to upper level undergraduates, with a specific interest in artificial intelligence, or to graduate students.
The introduction of AI based geographic analysis into the practice of landscape planning or research will add yet another straw to the camels back of landscape architecture education. The Landscape Architecture curriculum is already crowded with a diverse range of courses from fine arts to site engineering. With this problem in mind, can we justify the addition of another area of study? For the students with a particular interest in landscape planning and research, there is an inevitable requirement to understand GIS and its related modeling capabilities. As AI increasingly continues to be incorporated into a growing variety of modeling applications, students will be required to understand the fundamentals of this new technology as well as its applicability in landscape planning.
Applications which merge the data handling capabilities of geographic information systems with the analytical capabilities of artificial intelligence promise to provide resource managers and landscape planners with new automated decision support capabilities. This technology does not guarantee more accurate decision making (Gimblett et al 1994), but it should foster more informed decisions about how natural resources should be managed. According to Gimblett et al (1994), this combination of technologies provides the following benefits:
Models which link AI technologies to existing GIS databases provide the capability to develop rules for more complex land allocation models without the added expense of new data acquisition.
Identification of component interactions in a neural network model can be achieved by examining the rules developed during the simulation or alternatively the learning weights after the simulation.
The inclusion of artificial intelligence provides a means by which biophysical models can be linked with behavioral models such as when modelling recreational behavior within forest harvesting areas. This innovation allows the modeler to include public values directly with the biophysical models in a multiple network model (See Gimblett et al 1993).
When combined with a genetic algorithm, neural net based models are capable of neighborhood analysis, thereby solving the problem of modelling on an isolated cell by cell basis.
This modelling approach assists in isolating conflicts in the results and provides the manager with a tool to experiment with alternative strategies and solutions to these areas.
Once trained, neural networks can be applied and reapplied in similar situations. In addition, neural networks are adaptive by nature and able to handle fuzzy or missing data with real-time capabilities.
Currently, neural networks, and other AI systems are not easily set up to be interfaced with a GIS. Those exploring AI-GIS applications are frequently required to create their own interface links, often by writing programs in C or another programming language. The current lack of easily utilized connections between GIS and AI presents a formidable barrier to people without the necessary computer programming skills. Overall, the advanced concepts of neural networks and the work required to link current AI systems to geographic information systems, render these approaches inaccessible to most resource managers and practitioners. Perhaps because of these and other problems, it appears that academics and professionals in landscape planning and other related fields have only begun to explore the capabilities of these new technologies. Geographers have been slow to appreciate both the potential and impact of these new developments on their areas of study (Openshaw 1993). A recent survey of landscape architects indicated that only six percent of the offices that responded are using computer technology for geographic analysis (Palmer and Buhmann 1994). Palmer and Buhmann surmise that landscape architects may not be able to justify the necessary investment in computer technology for the spatial analysis phases of a large planning project. Practitioners seldom have the funds to devote to the investigation of new computer modelling techniques, particularly when the outcome of the investigation may be uncertain. These practitioners will need to see examples of how these new technologies can be applied in cost effective ways before they adopt them.
Clearly, before we will see the widespread adoption of these technologies, researchers and educators must be prepared both to lead the way in the investigation of new AI-GIS based techniques and theories and to disseminate the results of this research to professionals and students. Future research endeavors should be directed towards the development of easily understood, user friendly, applications. Artificial Intelligence based modelling applications will continue to develop as computer technology grows. As landscape planners and researchers, we have the opportunity to take advantage of these technologies to improve the tools of landscape planning and research.
Bishop, I. 1994. Comparing Regression and Neural Net Based Approaches to Modelling of Scenic Beauty. Centre for Geographic Information Systems and Modelling. The University of Melbourne. Melbourne,Australia.
Buhyoff, G.J., Miller, P.A., Roach, J.W., Zhou, D., and L.G. Fuller. 1994. An AI Methodology for Landscape Visual Assessments. AI Applications. 8:1-13.
Burley, J.B. and T.J. Brown. 1995. Constructing Interpretable Environments from Multidimensional Data: GIS Suitability Overlays and Principle Component Analysis. manuscript accepted for publication in Journal of Environmental Planning and Management.
Charniak, E. and D. McDermott. 1984. Introduction to Artificial Intelligence. Addison-Wesley.
Deadman, P., Brown, R.D., and H.R. Gimblett. 1993. Modelling Rural Residential Settlement Patterns with Cellular Automata. J.of Environmental Management. 37: 147-160.
DuBose, P. and C. Klimasauskas. 1989. Introduction to Neural Networks with Examples and Applications. NeuralWare Inc.Pittsburgh.
Gimblett, H.R. 1989. Modelling in GIS: A cellular automaton to modelling the growth of urban and rural development. GIS National Conference 89. Ottawa, Canada.
Gimblett, H.R., Guise, A.W., and D.P. Kroh. 1993. A Spatial Model for Assessing Conflicting Recreation Values in State Park Settings. MAGIC-Arizona Geographic Information Council, Phoenix, Az, 19-20 August 1993.
Gimblett, R.H., Ball, G.L. and A.W. Guise. 1994. Autonomous rule generation and assessment for complex spatial modeling. Landscapeand Urban Planning. 30:13-26.
Gimblett, R.H.., Ball, G.L. Neural Network Architectures for Monitoring and Simulating Changes in Forest Resource Management. AI Applications. Vol. 9, No. 2, 1995.
Hoffman, R. 1987. The Problem of Extracting the Knowledge of Experts from the Perspective of Experimental Psychology. AI Magazine, 8:53-67.
Hopkins, L.D. 1977. Methods for generating land suitability maps:a comprehensive study. J. Am. Inst. Planners. 43: 386-400.
Openshaw, S. 1993. Modelling Spatial Interaction Using a Neural Net. in Geographic Information Systems, Spatial Modelling, and Policy Evaluation. M.M. Fischer and P. Nijkamp eds. Springer Verlag. Berlin. pp.147-166.
Palmer, J.F. and E. Buhmann. 1994. A Status Report on Computers. Landscape Architecture. 84(7):54-55.
Pattie, D.C. 1992. Using Neural Networks to Forecast Recreation in Wilderness Areas. AI Applications 6:57-59.
Plant, R.E. and N.D. Stone. 1991. Knowledge-Based Systems in Agriculture. McGraw-Hill.
Porter, M.L. 1992. Neural Nets Offer Real Solutions. AI Applications 6(1): 32-40.
Sui, D.Z. 1992. An Initial Investigation of Integrating Neural Networks with GIS for Spatial Decision Making. Proceedings of GIS/LIS 92. San Jose California.
Wang, F. 1992. Incorporating a Neural Network into GIS for Agricultural Land Suitability Analysis. Proceedings of GIS/LIS 92. San Jose. California.