Radio communication is an integral part of wireless sensor networks. This dissertation focuses on improving the energy consumption of radio communication in sensor networks by proposing novel approaches in two key aspects of low-power wireless communication, namely, wireless link quality estimation and low duty-cycle data forwarding protocol. I first motivate the research with a comparative study of the routing performance and energy consumption with respect to existing link quality estimation protocols. Then, I propose 4C, a data-driven approach to build link quality prediction models based on empirical data collected from the deployment site in order to address the problems of link quality estimation in low-power sensor networks. Furthermore, I improve this novel data-driven approach to predict short temporal link quality variations without prior collected training data by employing online learning techniques. Analytical and empirical evaluations show that the proposed link quality prediction approach can significantly reduce the cost of radio transmissions by utilizing long links with variable quality. Moreover, I proposed SAF, an energy-efficient data forwarding protocol that can effectively utilize the short term link quality prediction models in duty-cycled networks. With the help of link quality prediction models, SAF not only minimizes the energy consumption spent on idle nodes, but also leverage the spatial diversity of wireless links via opportunistic routing. The dissertation concludes with a discussion of the potential improvements of the proposed approaches in low-power sensor networks as well as in other wireless networks.