Introduction
Focusing on the problem of estimating sediment transport over a fluvial system is one of important aspects in environmental management and applied geomorphology. Artificial neural networks (ANNs) have been applied to runoff-sediment modeling and flood forecasting. One of the most important parameters in runoff-sediment process is geomorphic characteristics. Generally, two methods exist for modeling: function-driven and data driven modeling. Function-driven modeling is based on fitting a suitable function such as regression curves but data-driven modeling such as neural networks work by giving weight to each data in a try and test in a frequently algorithm process.Therefore, neural networks are alternative and complementary set techniques to traditional models. The purpose of this paper is to apply and compare both regressions and neural networks for runoff-sediment modeling in a watershed scale by geomorphic parameters and without them. Therefore, this study evaluates performance of two artificial neural networks-a geomorphology based artificial neural network (GANN) and a non- geomorphology based (ANN); and two regression models, power relation (PR) and multivariate adaptive regression spline (MARS), for prediction of suspended sediment.

Material and Methods
The study area comprises of the Plasjan River, in Eskandari watershed, north east of Zayandeh roud basin, Esfahan, Iran. The watershed has an area of approximately 1640 km2. The recorded data of runoff and suspended sediment values are available, measured at one station in outlet of the Eskandari station. Data set of flood runoff and sediment at the same time for the years of 1995 to 2006 were provided by the Esfahan Water Agency. Several geomorphic parameters were used to create geomorphic models such as relative relief, form index and drainage density.
There are several stages processes to develop artificial neural network for simulation application as follow:
1. Data selection: gathering an appropriate data set.
2. Selection of an appropriate predictand: to decide what is to be modeled
3. Artificial neural network selection: to select an appropriate type of network and choose a suitable training algorithm.
4. Data preprocessing: to process the original data in terms of identifying suitable network inputs (predictors) and perform data cleansing as appropriate, for example, if necessary, remove trends or seasonal components. In addition, one must normalize and split the data into training, validation and testing data sets.
5. Training: to train a number of networks using the chosen training algorithm and preprocessed data.
6. Using appropriate assessment criteria evaluate the model produced and select the best solution for subsequent implementation.
It was revealed that the feed-forward ANN model with back propagation algorithm performed well for both the GANN and ANN models. The sediment loads predicted by these models were compared with observed data for the same watershed and compared with regression models including power regression and multivariate adaptive spline with evaluation indexes of root mean square of errors and determination index.

Results and discussion
Development of regression and neural network and using geomorphic parameters besides to runoff and sediment showed noticeable results. However, the GANN predicted better with highest coefficient of determination (R2) of 0.98, root mean square error (RMSE) of 4.49 in comparison to ANN (R2 = 0.96, RMSE = 5.35). The regression model performance was inferior (R2=0.89, RMSE=8.66) for MARS and (R2 = 0.81, RMSE = 15.05) for PR to the ANN models. Therefore ANN technique especially GANN is a powerful tool for real-time prediction of sediment transport in a complex network of rivers.

Conclusion
Neural network (NN) is a suitable tool for simulating the behavior of sediment transport in a river system. One major advantage the NN approach has over traditional input-output modeling is that it makes fewer demands of data. Unlike multiple regressions, where the constraints preparation the number and distributions of data, are often simply used, NN do not make assumptions about the statistical properties of a data set. Data for several variables can be use flexibility on temporal and spatial scales. Therefore NN find a non-linear pattern carefully that it does not with traditional methods. An advantage of the results is the effective rules of geomorphology parameters in modeling procession fact which illustrate importance of them in vision of river’s behavior in a catchment.