Although the processing of data streams has been the focus of many research efforts in several areas, the case of remotely sensed streams in scientific contexts has received little attention. We present an extensible architecture to compose streaming image processing pipelines spanning multiple nodes on a network using a scientific workflow approach. This architecture includes (i) a mechanism for stream query dispatching so new streams can be dynamically generated from within individual processing nodes as a result of local or remote requests, and (ii) a mechanism for making the resulting streams externally available. As complete processing image pipelines can be cascaded across multiple interconnected nodes in a dynamic, scientist-driven way, the approach facilitates the reuse of data and the scalability of computations. We demonstrate the advantages of our infrastructure with a toolset of stream operators acting on remotely sensed data streams for realtime change detection.