Predicting Mean and Bankfull Discharge from Channel Cross-Sectional Area by Expert and Regression Methods

This study employed four methods—non-linear regression, fuzzy logic (FL), artificial neural networks (ANNs), and genetic algorithm (GA)-based nonlinear equation—for predicting mean discharge and bank-full discharge from cross-sectional area. The data compiled from the literature were separated into two groups—training (calibration) and testing (verification). Using training data sets, the methods were calibrated to obtain optimal values of the coefficients of the non-linear regression method; optimal number of fuzzy subsets, their base widths and fuzzy rules for the fuzzy method; and the optimal number of neurons in the hidden layer, the learning rate and momentum factor values for the ANN model. The GA-based method employed 100 chromosomes in the initial gene pool, 80% cross over rate and 4% mutation rate in determining the optimal values of the coefficients of the constructed nonlinear equation. The calibrated methods were then applied to the test data sets. The test results showed that the non-linear regression, ANN and GA-based methods were comparable in predicting the mean discharge while the fuzzy method produced high errors and low accuracy. The GA-based method had the highest accuracy of 75%. In terms of predicting bankfull discharge, all methods produced satisfactory results, although the fuzzy method had the lowest accuracy of 33%. The results of sensitivity analysis, which is limited to the GA-based and nonlinear regression methods, showed that the GA-based method calibrated with low bankfull discharge values can be successfully applied to predict high bankfull discharge values. This has important implications for predicting bankfull rates at ungauged sites. On the other hand, the sensitivity analysis results also showed that both the non-linear regression and GA-based methods have poor extrapolation capability for predicting mean discharge data. Copyright Springer Science+Business Media B.V. 2011