Efficiency of Regression, ANN and ANN-algorithm Genetic Hybrid Models in the Evaluation of Wind Erosion

 Wind erosion has occurred in a large part of Iran, which has caused land degradation and reduced fertility along with environmental effects. Identifying erosion-sensitive areas can help natural resource and environmental managers in soil conservation planning.

This study is a step to estimate the erodible component of soil against the wind (EF) from soil accessibility characteristics in Allahabad plain located in the east of Qazvin province. For this purpose, the soil erodibility component, which is closely related to soil erosion versus wind, using multivariate regression (MLR), artificial neural network (ANN), and artificial neural network with genetic algorithm for weight optimization (GA-ANN) were estimated using accessible characteristics. Regarding soil map, soil differences, and environmental characteristics of Allahabad plain, 103 soil samples were collected according to a stratified random pattern of 10 cm of soil surface. In soil samples, some soil properties were measured as inputs of models for estimating erodible soil components against the wind. The inputs of each model included pH, ECe, CCE, SAR, bulk density, sand particles, silt and clay, coarse soil particles with a diameter of more than 2 mm, and organic matter. Accuracy and reliability of the results of the created models were compared with each other according to the criteria of coefficient of determination, square of error, Morgan-Granger-Newbold and Akaike information criterion.

Based on data, the highest correlation between soil erodible fraction to wind erosion (EF) was observed with soil clay content (r = -0.789). Also, soil erodible components showed a correlation with other soil properties including pH, electrical conductivity, SAR, organic matter, and the should be omitted density. This correlation was significant with three properties of SAR, organic matter, and clay at a should be added 1% level. The models created by the three methods were much more capable of predicting EF in the test data series than the training series data. The results also showed that the neural network model had a should be omitted more accuracy and less estimation error compared to hybrid and regression models. The results of sensitivity analysis of the models also showed that the highest sensitivity of the model to input variables in the ANN model, related to organic matter and SAR, respectively, and in the model GA-ANN was related to soil clay content variable.

According to the results, R2 in the regression model of training data was more than 50% in estimating EF, but this value (R2 = 0.56) is not reliable. According to the test data, all three models, including regression, artificial neural network, and its combination with genetic algorithm had not been efficient enough in estimating EF, so that can be omitted the highest R2 in the neural network model in the test data (R2 = 0.43) had an accuracy of less than 50% in estimating the EF, which cannot be an appropriate accuracy in predicting EF.