In a dense sensor network, sensors are randomly placed at a density high enough that over-sampling of a physical field occurs. Depending upon the application, very often some regions contain more information than the rest. For optimal energy efficiency, the amount of compression and hence the sensing accuracy should vary with the relative significance of different sub-regions. To achieve this goal, we propose a progressive data collection scheme. First, the more significant low-frequency components, which give a crude profile of the signal, are extracted and forwarded to the destination. Then, based on the rough profile, an application may identify some strategic areas and request additional details, i.e. components of higher frequencies, from them until the desired resolution is reached. This process of sensing a phenomenon is like making observations through a telescope: starting with a panoramic view of low resolution, then gradually zooming into a target region to reveal finer details. Noise in data can also be effectively removed during the process by applying digital low-pass filtering locally within sensor clusters. Simulation results based on the temperature distribution of a fire in one of the engineering buildings on the Columbia campus generated by the NIST Fire Dynamics Simulator show that with a sensor density of 1.2 /m2 and a SNR of 6 dB, using a simple Gaussian low-pass filter for local compression, an average overall improvement of 4.4 dB can be obtained while collecting only 4.8% of the raw data traffic at the processing center, resulting in a reconstructed field deemed accurate enough for extracting crucial information about a fire.