A powerful concept to cope with resource limitations and information redundancy in wireless sensor networks is the use of collaboration groups to distill information within the network and to suppress unnecessary activities. When the phenomena to be monitored have large geographical extents, it is not obvious how to define these collaboration groups. This paper presents the application of geometric duality to form such groups for sensor selection and non-local phenomena tracking. Using a dual-space transformation, which maps a non-local phenomenon, e.g., the edge of a halfplane shadow, to a single point in the dual space and maps locations of distributed sensor nodes to a set of lines that partitions the dual space, one can turn off majority of the sensors to achieve resource preservation without losing the detection and tracking accuracy. Since the group such defined may consist of nodes that are far away in the physical space, we propose a hierarchical architecture that uses a small number of computationally powerful nodes and a massive number of resource constrained motes. By taking advantage of the continuity of physical phenomena and the duality principle, we can greatly reduce power consumption in non-local phenomena tracking and extend the life time of the network.