Technology and Map-Learning: Users, Methods, and Symbols

Abstract

This article investigates the cognitive processes used by learners (adults and young adolescents) for tasks that require the integration of geographical information across spaces, hierarchies, and geographic scales. An experiment simulated basic GIS functions and contained four experimental conditions (Chunk, Layer, Scale, and Whole). Reaction time, accuracy, and confidence were recorded as dependent variables related to the success of the integration process. The data were used as input for a back-propagation neural-network model. The neural network model was successful in learning patterns and could be used to predict the confidence, reaction time, and accuracy for combinations of learners, experimental conditions, and map-feature categories. A multivariate analysis of variance was used to determine significant relationships among the behavioral variables and characteristics of the learners, experimental conditions related to GIS functions, and map features (points, lines, and areas). The results of the analysis generally indicated that young adolescent learners were slower, less accurate, and more confident than adult learners for all experimental conditions. Overall, subjects were more accurate and confident in tasks that required less integration of geographical information. Learners had the most success recalling information related to area symbols and the least success recalling information related to point symbols.

Key Words:

• GIS
• neural network
• spatial abilities
• spatial cognition
• young adolescents

Notes

^aThe dark shading indicates the scales with same features, while the light shading indicates the scales that do not have the same feature.

^aNeighboring input neurons in the list that have a common font were compared. Dark shading indicates the vector is pointing in the correct direction along that dimension. Light shading indicates the vector is not pointing in the correct direction along that dimension. A shaded name indicates the vector is consistently correct or incorrect on all three dimensions.

^aF-statistics are based on Type III sum of squares.

1. Kulhavy and Stock (1996, 126) list four important task demands associated with learning from maps: “1) explicit instructions about use of the map; 2) methodological factors like repeated map-learning and recall trials; 3) expectations concerning the type of test to be completed; and 4) subtle hints about the kinds of performance the experimenter hopes to produce.”

2. For early discussions of strategies for acquiring spatial information, see Stone (1964) and Thorndyke and Staz (1980).

3. Discussions of scale are frequently confusing. If one defines “scale” by comparing an observer to the size of the environment, then a large scale would have a large space compared to a relatively small person. Discussions centered on maps may use terms such as “cartographic scale” or “map scale,” but they frequently use just the word “scale.” If the ratio of a unit of distance on the map and the units of distance it represents on the earth is used to define scale, a small scale would be related to large space and a large scale would be related to a small space.

4. For technical discussions of this model and other neural-network models, see McClelland and Rumelhart (1989) and O'Reilly and Munakata (2000). For examples of using neural-network models in geographic research, see Hewitson and Crane (1994) and Lloyd (1997a).

5. The neural network was constructed using NeuralWorks Professional II/Plus software (NeuralWare 1996). The learning coefficient for weights connecting the input and hidden neurons was set to 0.30. The learning coefficient for weights connecting the hidden neurons and output neuron was set to 0.15. The momentum coefficient was set to 0.4. The delta rule was used to compute changes in the weights after each trial. A sigmoid transfer function was used to compute neuron activations throughout the model.

6. A binomial test was used to determine if learners in all the experimental conditions were performing above chance levels. True/False can be processed correctly by chance 50 percent of the time. Given the number of times learners evaluated statements, a minimum of 52.2 percent correct was established as a threshold value for being significantly above chance. Only the Young Adolescent Chunk combination category failed to achieve an accuracy value above this threshold (theirs was 51.6 percent).