Spatial Knowledge for Image Understanding

Abstract

ABSTRACT: Both computer-assisted and fully automated image understanding in remote sensing require the ability to extract and merge spatial knowledge in a variety of forms. In this article some of the processes involved in extracting, representing and interpreting spatial knowledge in images, and in maps as they relate to images, are examined. In particular, a 'naming' of image units perceived as single entities is found to be the preferred mode of describing shape information when both map and image interpreters communicate their findings. Furthermore, the interpreter appears to construct a top-down hierarchical structure of embeddings and relations, each node of which consists of a single 'sentence'. The structure of this cartographic sentence is determined by the attentional field of the interpreter, and consists of an attentional frame, an object/texture combination (subject) and a set of spatial and topological referents (predicate) to other objects. The analysis of this nested structure provides insight into appropriate strategies for automated image and map interpretation as well as providing clues to building natural language interfaces with GIS