This paper presents a new technique for information fusion. Unlike most previous work on information fusion, this paper explores the use of instance-level (extensional) information within the fusion process. This paper proposes an algorithm that can be used automatically to infer the schema-level structure necessary for information fusion from instance-level information. The approach is illustrated using the example of geospatial land-cover data. The method is then extended to operate under uncertainty, such as in cases where the data are inaccurate or imprecise. The paper describes the implementation of the fusion method within a software prototype. Finally, the paper discusses several key topics for future research, including applications of this work to spatial-data mining and the semantic web.