Deployed in a harsh oceanic environment for mission critical tasks, underwater acoustic networks have been regarded as a special type of wireless sensor networks. The limited bandwidth and power resources as well as the 3-D topology in such networks have made the geographic routing a favorite choice. In underwater sensor networks environment, most of the detouring strategies of the existing geographic routing do not work well. Especially, when greedy forwarding fails, a detour for a packet can be found efficiently by the spanning tree routing detouring strategy. But, the quality of pre-constructed spanning tree plays an important role on the effectiveness of the spanning tree routing. The routing trees constructed by most of the existing spanning tree algorithms are centralized and in top-down approach. Beside these, traffic pattern in the network is not considered. Those are the reasons that the routing trees created by the methods usually have poor routing performance. In this paper, we propose a novel spanning tree, namely Traffic-Aware Routing Tree (TART), which is constructed completely in a bottom-up fashion with the traffic load in mind. Simulation results show that TART have very few conflicting hulls and result in much higher path throughput when compared against other spanning trees, leading to a better routing performance in a 3-D underwater sensor network.