Abstract The topic of automatically assigning geographic coordinates to Web 2.0 resources based on their tags has recently gained considerable attention. However, the coordinates that are produced by automated techniques are necessarily variable, since not all resources are described by tags that are sufficiently descriptive. Thus there is a need for adaptive techniques that assign locations to photos at the right level of granularity, or, in some cases, even refrain from making any estimations regarding location at all. To this end, we consider the idea of training language models at different levels of granularity, and combining the evidence provided by these language models using Dempster and Shafer’s theory of evidence. We provide experimental results which clearly confirm that the increased spatial awareness that is thus gained allows us to make better informed decisions, and moreover increases the overall accuracy of the individual language models.