In recent years, gliding robotic fish have emerged as promising mobile platforms for underwater sensing and monitoring due to their notable energy efficiency and maneuverability. For sensing of aquatic environments, it is important to use efficient sampling strategies that incorporate previously observed data in deciding where to sample next so that the gained information is maximized. In this paper, we present an adaptive sampling strategy for mapping a scalar field in an underwater environment using a gliding robotic fish. An ergodic exploration framework is employed to compute optimal exploration trajectories. To effectively deal with the challenging complexity of finding optimum three-dimensional trajectories that are feasible for the gliding robotic fish, we propose a novel strategy that combines a unicycle model-based 2D trajectory optimization with spiral-enabled water column sampling. Gaussian process (GP) regression is used to infer the field values at unsampled locations, and to update a map of expected information density (EID) in the environment. The outputs of GP regression are then fed back to the ergodic exploration engine for trajectory optimization. We validate the proposed approach with simulation results and compare its performance with a uniform sampling grid.