Steep mountain streams exhibit shallow waters with roughness elements such as stones and pebbles that are comparable in size to flow depth. Owing to the difficulty in measuring fluid velocities at the interface, i.e., from the rough permeable bed to the free surface, experimental results are rare although they are essential to improve models. Using a novel experimental procedure, this thesis attempts to improve predictions of the vertical structure of turbulent flows over rough permeable beds. To explore flows at the bed interface, I devised an experimental set-up where a fluid flowed over glass spheres (8 mm < dp < 14 mm) in a narrow flume (W = 6 cm) with slopes varying from 0.5 % to 8 %. The Refractive Index Matching (RIM) technique has been employed. This involves matching the refractive index of the fluid with that of the glass spheres, thereby allowing the interior of the medium to be examined and velocities to be measured by Particle Image Velocimetry (PIV). Vertical profiles are retrieved by employing the spatiotemporal double averaging method. In the course of this manuscript, flow processes are studied at the mesoscopic scale, i.e., by averaging quantities over distances ranging from 5 to 10 grain diameters. For open-channel flows over rough permeable beds, the spatial averaging procedure yields a continuous porosity profile. When applied to the Navier-Stokes equations, it produces a momentum equation with several terms including drag forces and three stresses: the turbulent, dispersive, and viscous stresses. The momentum equation was employed to devise a one dimensional (1D) model describing the vertical structure of unidirectional turbulent flow. A turbulent boundary layer over the rough bed was observed while experiments were performed at intermediate Reynolds numbers, i.e., Re = O (1000). In such conditions, viscosity plays a critical role through the van Driest damping effect. To model vertical profiles, the Darcy-Ergün equation is well suited to the prediction of friction forces in the permeable bed, i.e., in roughness and subsurface layers. Based on the \textit{Prandtl mixing length theory}, turbulent stress is predicted from a mixing length distribution that considers dispersive effects and assumes a continuous porosity profile. This alternative contrasts with most existing boundary layer models which postulate a discontinuous porosity profile for permeable or impermeable walls. Finally, hydraulic conditions collected by an Unmanned Aerial Vehicle (UAV) and classical flow resistance equations (Chézy, Keulegan, ...) were compared with profile simulations and demonstrate a good agreement between predictions and observations. It reveals the crucial role of fluid depth definition in equations in small submergence conditions. Furthermore, incipient sediment motion conditions have been estimated and compared to empirical results showing the importance of turbulence and lift force for grain entrainment. With regard to fluid dynamics, mountain streams are a case study of the larger scientific family of turbulent flows interacting with porous structures. Insights and developments acquired in the course of this thesis are likely to be transferable to other domains working with these phenomena such as flows over buildings, vegetal canopies or rough wings.