Terrain relative navigation (TRN) offers a means to constrain absolute vehicle position and attitude without using a GPS receiver or star camera, such as for unmanned aerial vehicles or planetary landing spacecraft, using a pre-computed terrain database of distinctive landmarks in the operational environment. However, depending on the length of the planned trajectory, these terrain landmark databases may grow prohibitively large for the onboard data storage, processing, and communication capabilities of these types of vehicles.In this paper, we introduce a definition of landmark utilitywhich encapsulates the observation probability of a landmark and its spatial distribution relative to nearby landmarks in the database -as a measure of the value of potential line-ofsight measurements to that landmark. We additionally present an efficient algorithm to sort all potential landmarks in an environment based on their relative utility without access to the actual sensor measurements or trajectory. This enables pre-determination of a limited-size terrain landmark database containing the N -best landmarks in the environment, which is applicable to any landmark-based TRN system.To evaluate our landmark selection algorithm, we additionally introduce an incremental smoothing-based approach to TRN using a Bayesian factor graph representation. This system is capable of overcoming several challenges associated with TRN systems, including relinearization of past measurements, simultaneously utilizing pre-mapped and opportunistic features in a common estimator, and efficient sensor fusion in a common, probabilistic framework. We provide results showing the benefits of our utility-based landmark selection compared to alternative selection approaches, using evaluation metrics independent of any particular TRN estimation framework and the navigation performance of our graph-based TRN system.