artificial neural network model

In this study, the discharge coefficient (Cd) of piano key weirs was estimated using the Support Vector Machine (SVM) model, and its performance was subsequently compared with the Multilayer Perceptron Neural Network (MLPNN) model. For this purpose, the parameters of the upstream head-to-weir height ratio (h⁄P), the inlet-to-outlet width ratio (w_i⁄w_o), the key length-to-width ratio (L_cy⁄w_cy), and the number of keys (N) were considered as inputs, with Cd as the output. The results indicated that the minimum value of the R2 statistical index and the maximum value of the RMSE for the mentioned models during the validation phase were R2=0.99 and RMSE=0.01, respectively. The developed MLPNN model consisted of two hidden layers, with four neurons in the first layer and two neurons in the second layer, both utilizing the sigmoid tangent activation function. The SVM model employed a radial basis function (RBF) kernel. Sensitivity analysis of the models revealed that the most influential parameters in modeling and estimating the discharge coefficient were h_o⁄P.

Soil structure stability refers to the ability of a soil to hold up the solid particle arrangement and the spaces among them when face to different stresses. According to the important role of organic matters in formation and stability of soil aggregates; it appears that the different fractions of them may also have significant effects on soil aggregate stability. Furthermore, separating the different fractions of organic matter helps the identification of the sensitive and sustainable parts and their locations in aggregate structure. The purpose of this study is to evaluate the effect of different fractions of organic matters on the soil macro (>0.25 mm) and micro (
Fifteen soil samples were taken from the surface (0-10 cm) using a supervised random method in three different land uses including forest, range, and agricultural lands (totally, 45 points) in Rabor region, Kerman province. After air-drying the samples and passing them through a 4 mm sieve, mean weight diameter (MWD) of soil aggregates was measured using the wet sieve method for the macro and micro aggregates. Then, different fractionations of organic matters in the macro and micro aggregates were determined using the density method. Finally, the amount of organic carbon in different fractions of organic materials and also the total amount of soil organic carbon were identified using the Walkley-Black method and then the percentage of total organic matter existed in each fraction was calculated by considering the weight and percentage of organic matter in each fraction,. Afterward, the organic matter data were used as input to the artificial neural network model. Besides, the regression relationships among the variables and soil aggregate stability were investigated.
The results showed that the free light fraction of organic matter (F1) in the macro-aggregates was greater than in the micro-aggregates. Also, the F1-amount in the forest was greater than the other two investigated land uses due to the higher organic matter content. The amounts of occluded light fraction of organic matter (F2) in the agricultural lands were lower than other land uses which might be due to tillage operation, soil aggregate destruction, and release of the trapped organic matter inside of them. The parts of organic matter which were associated with mineral fractions (F3) were allocated with the largest percentage of the total soil organic matter as compared to the other two fractions. The amount of last mentioned fraction was higher in the micro-aggregates than the macro-aggregates in all three investigated land uses. Results also revealed that the linear regression was not able to identify the interrelationship between the studied variables and aggregates stability index. In contrast, the diagrams related to artificial neural network model showed that all input variables to the model have influenced the MWD, although the important coefficients of the input variables were different.
Different fractions of soil organic matter were more sensitive to the land use type than the total organic matter. The macro and micro aggregate percentages were different depending to the land use type; however, the micro aggregates with greater organic matter contents were more stable against entered stresses than the macro aggregates. Artificial neural network model had more efficiency in estimating the soil aggregate stability than the linear regression method indicating that there is a non-linear relationship between different fractions of organic matter and MWD. By considering the accuracy and efficiency of artificial network model, it appears that this method can be used to determine the linear and non-linear relationship among different soil properties, with higher precision and lower cost and time.

Knowledge of organic carbon spatial variations in different land uses will help to interpret and simulate the behavior of terrestrial ecosystems facing environmental and climate changes. The purpose of this study is comparing regression, geostatistics and artificial neural network (ANN) methods for predicting organic carbon content in 192 samples of surface soil (0 to 30 cm) of Sistan plain (Miankangi region). In this study, Only 25% of organic carbon variations were explained by variables used in linear regression model in the study area (R2= 0.25). Moreover, simple co-kriging (with clay as co-variable) which was the best geostatistical method in the current study, predicted organic carbon content weakly (R2= 0.23 and RMSE= 0.127). However, using latitude and longitude parameters, ANN performed much better than linear regression and geostatistical methods for predicting organic carbon content (R2= 0.79 and RMSE= 0.044).

Cation exchange capacity (CEC) measurement is a very expensive and time-consuming method in large scale assessments. It can be an appropriate approach to predict CEC from readily available properties via developing nonparametric models. In the present study, a nonparametric technique has been used for estimating CEC and compared with the most common nonparametric models which is based on artificial neural networks (ANN). 683 soils were selected from central Iran that 120 of them were used as target data and the others (563) were the reference data set. The parameters, clay, silt, sand and organic carbon content were the input independent variables (readily available properties) and the CEC was as an output dependent variable in this work. The results showed that the maximum error (MaxE) in K-NN and ANN techniques were 4.81 and 5.26 cmol+/kg, respectively. Root mean squared error (RMSE) for the K-NN and ANN were 1.51 and 1.53, respectively. This indicated that both methods are able to properly and equally predict CEC. The positive values of mean error (ME) showed that both models tended to underestimate CEC values in samples. The results analysis also showed that the efficiency of models (EF=0.88) were high by the estimation of CEC values in target soils.