In-network surface simplification for sensor fields

Recent research literature on sensor network databases has focused on finding ways to perform in-network aggregation of sensor readings to reduce the message cost. However, with these techniques information about the state at a particular location is lost. In many applications such as visualization, finite element analysis, and cartography, constructing a field from all sensor readings is very important. However, requiring all sensors to report their readings to a centralized station adversely impacts the life span of the sensor network. In this paper we focus on modeling sensor networks as a field deployed in a physical space and exploiting in-network surface simplification techniques to reduce the message cost. In particular, we propose two schemes for performing in-network surface simplification, namely (1) a hierarchical approach and (2) a triangulation based approach. We focus on a quad tree based method and a decimation method for the two approaches respectively. The quad tree based method employs an incremental refinement process during reconstruction using increasingly finer levels of detail sent by selected sensors. It has a guaranteed error bound. The decimation method starts with a triangulation of all sensors and probabilistically selects sensors not to report to prevent error accumulation. To demonstrate the performance, the two simplification techniques are compared with the naive approach of having all sensors report. Experimental results show that both techniques provide substantial message savings compared to the naivealgorithm, usually requiring less than 80% as many messages and less than 50% for some data sets. Furthermore, though the decimation algorithm does not provide a guaranteed error bound, for our experiments less than 4.5% of the interpolated values exceeded the given bound.