Nonlinear adaptive estimation is applied to terrain-referenced navigation in three dimensions. In this scheme, a bank of parallel filters is initialized with different altitude hypotheses, where each filter represents a discrete (gridded) approximation to the Bayes minimum variance estimator. The importance weight of each filter is recursively updated using the measurement residuals. The altitude bias estimate is found from a weighted sum of the filter hypotheses. Numerical simulations with synthetic data indicate that an altitude bias can be accurately estimated without the use of three-dimensional grids. The computational simplicity and parallel nature of the filter may make it suitable for estimation of additional parameters such as horizontal velocity components.