An evaluation model of artificial neural network to predict stable width in gravel bed rivers

Regime width of alluvial channels is a vital problem in river morphology and channel design. Many equations are available in the literature to predict regime width of alluvial rivers. In general, there are many approaches to estimate regime width; however, none of them is widely accepted at present. This is due to the fact that most hypotheses have many constrains which may lead to simplify governing conditions and also lack of knowledge of some physical processes associated with channel formation and maintenance. Intelligent models are a new approach to describe complex problems one of which is artificial neural networks. In this research, initially, gravel bed rivers database was used in bankfull discharge condition to train various dimensional and non-dimensional neural-network schemes with three and four variables as input data, respectively. Then, the same database was applied to fit regression analysis to estimate regime width; this led to drive dimensional and non-dimensional equations. Finally, dimensional and non-dimensional neural-network models and regression equations were compared together based on 50% error bands with other dataset. Results show that neural network can adequately estimate the regime width in gravel bed rivers and multilayer perceptron network with one hidden layer and eight hidden neurons based on dimensional data set was selected as optimum network to predict regime width. A sensitivity analysis also shows that bankfull discharge has a greater influence on regime width of gravel bed channels than the other independent parameters in dimensional scheme of neural network.