The efficiency of Artificial Neural Network, Neuro-Fuzzy and Multivariate Regression models for runoff and erosion simulation using rainfall simulator

According to the complexity of environmental factors related to erosion and runoff, correct simulation of these variables find importance under rain intensity domain of watershed areas. Although modeling of erosion and runoff by Artificial Neural Network and Neuro-Fuzzy based on rainfall-runoff and discharge-sediment models were widely applied by researchers, scrutinizing Artificial Neural Network and Neuro-Fuzzy models based on environmental factors has been paid less attention. Therefore, this study aimed at determining the efficiency of different models including Artificial Neural Network, Neuro-Fuzzy and Multivariate Regression for runoff and erosion simulation using rainfall simulator in some catchments of the North-West of Iran selected in terms of the same rain intensity of half an hour with a 10-year return.
THEORETICAL FRAMEWORK: Modeling runoff and erosion relations with environmental factors under prevelant rainfall intensity in a watershed scale are considered as the novel aspect of recognition of these complex relations. In this regard, implementation of determined rainfall intensity in a watershed scale is needed in the utilization of rainfall simulator apparatus. Also, the complexity of runoff and erosion relations with the environmental factors is the reason for the application of different models including Artificial Neural Network, Neuro-Fuzzy and Multivariate Regression. In fact Artificial Neural Network models are able to recognize the complex and unknown relations based on working as human brain. The simulation by these models finds importance when these relation have a non-linear feature. Parallel and Distributive processing of information and interpolation ability are major properties of Artificial Neural Network and Neuro-Fuzzy models characterized in the utilization of these models in the correct simulation of complex relations.
The establishment of rainfall simulator conducted at 86 sites and 21 environmental variables (the characteristics of topography, pedology, vegetation and species diversity) were used as inputs to models. In this regard, Topographic characteristics (including elevation, slope and …) of established sites of rainfall simulator apparatus were first recorded. Then sampling of soil was done from 4 corners of each site and compounded in order to eliminate soil heterogenic effects. After providing one soil sample from each site, all samples were sent to soil laboratory for measurement and analysis of different pedology properties including soil organic matter, total nitrogenous, absorbable phosphorus, available potassium, pH, electrical conductivity, soil moisture, calcareous content, gypsum content, Ca cation, Na cation, soil texture, distribution of clay, silt and sand percentage of soil. Also, vegetation characteristics including canopy cover, pavement and stone percentage and species abundance of each site was investigated in plot of simulator apparatus. Abundance parameter of species in each site was used for determining different species diversity indices (including species number, Simpson, Shannon-wiener and dominance indices) in PAST software package. Implementation of determined rainfall intensity of each site by simulator apparatus was finally performed for the measurement of runoff and erosion variable. Analysis of data was done through Multivariate Regression in SPSS software package, simulation via Artificial Neural Network (multi - layer perceptron and radial basis function methods), and Neuro-Fuzzy models was performed via MATLAB software package. Model validation conducted on 18 percent of the data based on Root of Mean Square Error, Nash–Sutcliffe Efficiency and Mean Absolute Error indices.
The results of Multivariate Regression model of this research showed that variables such as soil moisture, absorbable phosphorus, canopy cover percentage and soil sand percentage caused for runoff content and variables as calcareous content, total nitrogenous, canopy cover percentage, soil organic carbon and land slope determined erosion variable. In this regard, Multivariate Regression model was able to explain 68% and 46 % of changes in the runoff and soil erosion variables and its efficacy was lower in the simulation. As a result, Radial Basis Function neural network model compared with Multi Layer Perceptron as well as Neuro-Fuzzy model with scenarios of cluster (hybrid procedure) compared to grid method was able to predict more accurately. As indicators of RMSE, MAE and NSE were gained on optimum model of neural networks of 0.135, 0.114 and 0.99 for runoff volume, 0.011, 0.009 and 0.98 for the erosion and on optimum model of neuro-fuzzy models of 0.132, 0.111 and 0.92 for the volume of runoff and 0.013, 0.011 and 0.98 for the erosion, respectively.
CONCLUSIONS AND SUGGESTIONS: In general, it can be concluded that according to the presence of the complex environmental relations of erosion and runoff variables, Artificial Neural Network model with Radial Basis Function method and Neuro-Fuzzy model with scenarios of cluster (hybrid procedure) are recommended to be simulated based on ecological factors.