Abstract

The research presented in this paper attempts to develop a more realistic model using multi-layer perceptron (MLP), a feed forward artificial neural network (ANN), instead of traditional models like multiple linear regression (MLR) for predicting some soil physico-chemical and hydrological properties. The study area (Guilan Province) which is located in northern Iran bordering to south side of Caspian Sea in a coastal zone has udic and thermic soil moisture and temperature regimes respectively. The estimated soil parameters were CEC, EC, ESP, MWD and final steady-state infiltration rate (IR). Although these parameters can be measured directly, their measurement is difficult and expensive, so pedotransfer functions (PTFs) provide an alternative by estimating these parameters from more readily available soil data. In order to predict the mentioned parameters, soil sampling was conducted at 500 points in the region. Measured soil variables included texture, O.C, porosity, EC, CEC, SAR, ESP, MWD, soluble cations and anions and IR. Then, ANN and MLR models were tested. The data set was divided into two subsets for calibration (80%) and testing (20%) of the models and their normality were tested by Minitab software and Kolmogrov-Smirnov method. In order to evaluate the models, root mean square error (RSME) was used. The comparison of RSME for two mentioned models showed that the ANN model gives better estimates rather than the MLR model. So that the levels of RMSE and R² derived by ANNs models for EC, CEC, ESP, MWD and IR were 0.24, 0.96; 1.25, 0.90;  0.18, 0.94;  0.04, 0.84 and;  1.55, 0.92 respectively while these parameters for MLR models were 1.98, 0.73;, 7.92, 0.60;  1.13, 0.66;  0.187, 0.51 and;  9.45, 0.57 respectively. The superiority of ANN models compared with MLR models was probably due to a nonlinear relationship between the dependent and independent variables. Furthermore, results indicated that training is very important in increasing model accuracy for one region.

Key words: Artificial neural networks, multiple linear regressions, soil properties