Sensors are typically deployed to gather data about the physical world and its artifacts for a variety of purposes that range from environment monitoring, control, to data analysis. Since sensors are resource constrained, often sensor data is collected into sensor databases that reside at (more powerful) servers. A natural tradeoff exists between resources (bandwidth, energy) consumed and the quality of data collected at the server. Blindly transmitting sensor updates at a fixed periodicity to the server results in a suboptimal solution due to the differences in stability of sensor values and due to the varying application needs that impose different quality requirements across sensors. In order to adapt to these variations while at the same time optimizing the energy consumption of sensors, this paper proposes three different models and corresponding data collection protocols. We analyze all three models with a Markov state machine formulation, and either derive closed forms for the operation point of the data collection application or suggest algorithms for estimating this operating point to achieve a minimal energy consumption. We observe that the operating point depends on environmental characteristics and application quality requirements, which the proposed algorithms aim to accommodate. Our experimental results show significant energy savings compared to the naive approach to data collection.