An empirical model of hydraulic roughness for overland flow

Abstract

This research has developed a method for estimating hydraulic roughness coefficients for overland flow models in a dynamic approach, to more effectively simulate runoff on natural, agricultural and urban slopes. The hydraulic roughness coefficients are then generated with a series of neural networks. First, a laboratory experiment was designed to explore the effects of soil microtopography, slope and Reynolds number on the magnitude of Darcy-Weisbach, Manning and Chezy roughness coefficients. It was found that three parameters were necessary to describe the soil surface microtopography. Neural networks developed in a preliminary phase were able to reproduce the roughness coefficients obtained in the laboratory experiment by using five predictor variables: bed slope, Reynolds number, and the three parameters used to describe the microtopography. However, these networks failed to generate roughness coefficients for different input variables (generalization). Second, more complex algorithms were developed as combinations of neural networks in parallel. The algorithm output, the sought hydraulic roughness estimate, was estimated with the arithmetic average of the individual network outputs. Results presented in this study demonstrate that combining multiple neural networks reduced the prediction error and improved on the generalization ability of the neural networks. It was also observed that the estimate accuracy was influenced by the characteristics of the dataset, and especially by the relationship between the roughness coefficient and Reynolds numbers. Finally, a field experiment was performed to explore the applicability of the algorithms. A numerical model based on the 1-D diffusion approximation to the Saint Venant equations was constructed, and two surface irrigations were performed to collect data to test the model estimates. The model was used under two scenarios: (1) with constant hydraulic roughness coefficients, and (2) using variable hydraulic roughness predicted with the algorithm. Discharge at the end of the plot and irrigation front advance estimated using both models matched the observations well. However, when using a variable hydraulic roughness, the front was initially delayed until there was a sufficient surface storage to push it forward. The methodology described in this research should be useful for 2-D overland flow models applied to natural slopes with unsteady rainfall.