Surface water quality variables with the common nature of complex, nonlinear, multivariable and high variability need the application of alternative techniques to define optimum input variables and to develop effective models. The present paper describes the systematic or hierarchical development and validation of artificial neural network (ANN) and multiple linear regression (MLR) models for the purpose of predicting chlorophyll a (Chl-a) concentration from nine surface water quality variables in order to investigate the significant input variables and their contribution order. A multi-layer feed-forward network (FFN) trained by back-propagation algorithm was used as ANN approach. Both FFN and MLR techniques were first calibrated with the same three-fourth of the water quality data (304 samples) and then tested with the remaining one-fourth of the data. The performances of hierarchical models of both techniques were evaluated by using two statistical parameters such as coefficient of determination (R(2)) and root mean square error (RMSE). The systematic or hierarchical analysis results showed that only four input variables such as biological oxygen demand (BOD), water temperature (T), dissolved oxygen (DO), and phosphorus (P) among nine variables explained 75 and 88% of the variability in Chl-a compared to the full models of FFN and MLR techniques, respectively. Using these four substantial input variables and adding their powers and possible interaction terms as inputs increased the prediction ability of MLR technique by 13%. The best model of MLR was better than the full model of FFN for predicting Chl-a based on R(2) of 0.69 and 0.61%, and RMSE of 14.99 and 16.35, respectively. The study results show that both FFN and MLR techniques are capable of simulating Chl-a with acceptable accuracy and suggest that the abilities of them should be further investigated in different environments and under different conditions with hierarchical model development.