Estimation of soil erodible fraction using artificial neural network models and integration of artificial neural network with genetic algorithm in the part of Qazvin province

Erosion is one of the main factors restricting the soil fertility and dust production, in several parts of the world, including Iran, has effects on climate agriculture, and human health. Controlling wind erosion would be more effective once sufficient information concerning the effective factors is available. Soil Erodible Fraction (EF) is one of the soil properties that shows the sensitivity of soil particles to wind erosion. The current research aimed to utilize ANN methods and integrating it with GA in order to estimate the soil erodible fraction to wind erosion. Allahabad plain in the southwest of Abiek city in Qazvin province is considered as one of the areas sensitive to wind erosion with strong wind direction from southwest to northeast. The drying up of Allahabad wetland will intensify wind erosion in the region and turn it into a crisis. Determining the extent of land erodibility and identifying its factors affecting can be the basis of a comprehensive plan for soil protection and land sustainability and prioritizing its implementation steps. The present study was conducted to use artificial neural network methods and combine it with genetic algorithm to estimate the soil erodible factor.

In the study area, which was part of the Allahabad plain in Qazvin province, between the coordinates of 50°15 ́- 50°57 ́ east longitude and 35°53 ́- 35°57 ́ north latitude, 95 samples were taken from 10 cm of soil surface. In the samples, the percentage of aggregates with a diameter of less than 0.84 mm as an indicator of EF and percentage of clay, sand and silt, soil saturation capacity, pH, EC, SAR, equivalent calcium carbonate (CCE) and organic matter were measured as input to the models. In this paper, to model the EF using early available characteristics, two methods of artificial neural network (ANN) and its integration with genetic algorithm (GA-ANN) were employed in order to optimize the weights. In this regard, the data were primarily divided into three categories as follows: 60% of the data series was allocated to training, 20% to validation, and 20% to network testing. In this study, MLP networks were used to model the artificial neural network in estimating the values ​​of soil erodible Fracion. In this structure, each artificial neural network includes inputs and hidden and output layers. During the learning process, the degree of network learning by the objective functions was regularly evaluated and networks with the lowest error rate were accepted. To determine the optimal network with the highest level of performance of all stimulus functions defined in the software (axon hyperbolic tangent, axon sigmoid, axon linear hyperbolic tangent, axon linear sigmoid, axon bias, linear axon and axon) by trial and error The most results were used. Levenberg-Marquardt training functions were used to teach defined networks. In this study, genetic algorithm was used to find the optimal point of complex nonlinear functions in combination with artificial neural network (GA-ANN). The genetic algorithm optimizes the weights of the artificial neural network. In fact, the objective function of the genetic algorithm is a function of the statistical results of the artificial neural network.

The results showed that the erodible fraction of soil with five soil properties including pH, electrical conductivity, SAR, clay and organic matter, had a significant correlation at the level of one percent. The models used did not have an appropriate accuracy in estimating EF in both training and testing stages, so that the highest R2 was obtained in the artificial neural network model (0.49) with test series data. Both models were slightly overestimated and the GMER values ​​for the ANN and GA-ANN models were 1.15 and 1.08, respectively, but according to the AIC index, both models had similar predictive power. Sensitivity analysis of the data showed that the greatest effect on EF in the ANN model was related to organic matter (4.07) and in the GA-ANN model was related to clay (8.14).

In the current research, the relationship between soil chemical characteristics and EF might be attributed to their previous effects on vegetation in the region. Additionally, regional evidence indicates the same finding. The highest correlation was observed between EF and soil organic matter. Based on the sensitivity analysis, in the neural network model, the greatest effect on erodible fraction was related to organic matter, pH, and EC, respectively. The effect of pH and salinity on EF could be interpreted due to their effects on vegetation and consequently, the effect of vegetation on aggregates.  An important issue in the research was that the proposed models, which were ANN and its integration with GA for estimating the soil erodible fraction, were not efficient enough for obtaining the highest coefficient of determination (R2) in the model in the neural network in the test phase (R2 = 0.49), which has an accuracy of less than 50% for estimating EF.