GIS software applications and other mapping tools enable users to correlate data from multiple layers and gain insight from the resulting visualizations. However, most of these applications only feature basic, monolithic layer compositing techniques. These techniques do not always support users effectively in their tasks, as we observed during interviews with GIS experts. We introduce MapMosaic, a novel approach based on dynamic visual compositing that enables users to interactively create and manipulate local composites of multiple vector and raster map layers, taking into account the semantics and attribute values of objects and fields in the composit-ing process. We evaluate MapMosaic's interaction model against that of QGIS (a widely-used desktop GIS) and MAPublisher (a professional cartography tool) using the " Cognitive Dimensions " framework and through an analytical comparison, showing that MapMosaic's model is more flexible and can support users more effectively in their tasks. We also report on feedback obtained from experts, which further confirms the potential of this highly dynamic approach to map layer compositing.