Abstract: Cognitive categories in the geographic realm appear to manifest certain special features as contrasted with categories for objects at surveyable scales. We have argued that these features reflect specific ontological characteristics of geographic objects. This paper presents hypotheses as to the nature of the features mentioned, reviews previous empirical work on geographic categories, and presents the results of pilot experiments that used English-speaking subjects to test our hypotheses. Our experiments show geographic categories to be similar to their non-geographic counterparts in the ways in which they generate instances of different relative frequencies at different levels. Other tests, however, provide preliminary evidence for the existence of important differences in subjects ’ categorizations of geographic and non-geographic objects, and suggest further experimental work especially with regard to the role in cognitive categorization of different types of objectboundaries at different scales.

Abstract. Categories are essential to human reasoning, yet the tools we have at our disposal to discover, propose, apply and evaluate categories are still primitive. Current GIS often assume, a-priori, that the required categories for a specific exercise have been constructed by some external process. However, within the field sciences, it is often the case that categories must be proposed and refined in-situ, and through use. A further complication is the need to compromise between the needs of the analyst⎯who wants to be able to study some phenomena using existing categories defined by existing taxonomic knowledge, and the capabilities of the data gathered and algorithms used⎯that must be able to differentiate the required categories reliably. In this paper, we describe our latest progress towards developing an environment that supports the development and evaluation of categories. We show how a 'mixed initiative system', driven both by domain expertise (theory) and by data-oriented analysis tools (visualization and machine learning), can support the creation of categories that are both robust and useful. Theory is injected via a visual interface to the human expert, category separation is explored using visual and machine-based inductive analysis. 1.