Comparison of the Performance of Artificial Neural Networks and Gene Expression Programming in Estimating the Forest Soil Water Characteristic Curve

One of the most important and practical physical parameters in studies of soil water flow is Soil Water Characteristic Curve (SWCC). Measuring the soil moisture characteristic curve through the direct method is expensive and time-consuming. For this reason, a variety of indirect methods including intelligent models have been developed. In this study, the performance of three models included multilayer perceptron neural networks (MLP), cascade neural network (Cascade-NN) and gene expression programming (GEP) were evaluated and compared to estimate of SWCC. The measured data from 108 soil samples, including soil particle size distribution, soil moisture in different suctions and the bulk density were used. In all models, three different input data combinations were used. Comparison of predicted and observed values of soil moisture showed acceptable performance of all three models, however, the Cascade-NN neural network model was relatively superior. The R2 values of test phase for the best structure of the neural networks (MLP), neural networks (Cascade-NN) and gene expression programming (GEP) were 0.95, 0.96 and 0.93, respectively, and the RMSE values were 3.74, 3.25 and 4.10 %, respectively. Comparison of the results of different input data scenarios indicated the low accuracy and difference between the results of the models in the first scenario, but adding the parameters of porosity and moisture at field capacity point to the input data in the second and third scenarios, increased the accuracy and difference between the results achieved by the models. Finally, it can be emphasized that the cascade-NN model was introduced as the superior option, using all the mentioned physical data.