Sensor networks collect data at multiple distributed nodes and transfer the acquired information to points of interest. The raw data collected by each individual sensor is typically not of interest. Instead, a reduced representation of the measured phenomenon is to be generated. Multiple readings, however, add to the information about the phenomenon by providing its description at multiple points in space for distributed phenomena and multiple perspectives for a localized phenomenon. We also note that sensor readings have noise, and multiple readings can help mitigate the effect of this noise. Thus, while all the sensor readings need not be communicated, enough data must be exchanged to reliably reproduce the phenomenon. Considering the above effects, it becomes important to determine how much data should be transmitted from multiple sensors such that only useful information is exchanged and energy or bandwidth are not wasted on redundant data. We address this question using information theoretic techniques. The effects of sensor noise and correlation in the sensor readings are explicitly modelled.