1 Estimation of Exchangeable Sodium Percentage with the aid of Sodium Adsorption Ratio in sedimentary soils PROFESSOR DR. SORUSH NIKNAMIAN PHD IN CELL AND MOLECULAR BIOLOGY, BOARD MEMBER OF WESTON A PRICE FOUNDATION, WASHINGTON DC, USA AND ACTIVE MEMBER OF FEDERAL HEALTH PROFESSIONALS AT US ARMY FORCES, UNITED STATES OF AMERICA Abstract Soil salinity and sodicity are two main factors limiting plant growth in irrigated agricultural land. Sodium adsorption ratio (SAR) and exchangeable sodium percentage (ESP) are two different criteria as an index of soil sodicity and salinity. Various approximate relationships between ESP and SAR have been reported for soils in different regions of the world. Since there is possibility that these relationships change substantially with clay content, mineralogy, salinity of equilibrium solution, and saturation percentage of soils, it seems essential doing specific studies for different regions. The purpose of this research was to i) find the relationship between ESP and SAR, and ii) estimate the ESP from SAR in alluvial soils of Sistan, the dry plain in east of Iran. Thus, 301 soil samples were collected from study area and analyzed. The best linear and logarithmic equations found between ESP and SAR using Datafit software were ESP = 8.89 × ln(SAR1:1) + 14.04 and ESP = 8.73 × ln(SAR1:5) + 14.59, that ESP variation was justified 78% and 76%, respectively. Then, the multi-layer perceptron neural network (MLP) and ANFIS system performance were investigated in order to estimate ESP. Results showed superior performance of MLP and ANFIS compared with the regression models. ESP estimation from SAR1:1 using ANFIS was more accurate than other models (coefficient of determination and root mean square error values were 0.99 and 0.014, respectively). These results indicate the superiority of the intelligent models in order to explain the relationship between ESP and SAR over linear and non-linear regression equations. Keywords: Soil salinity, Soil sodocity, Regression equations, SAR, ESP, MLP, ANFIS Introduction Knowing the relationships and correlations among different soil properties and expressing them quantitatively in the form of statistical models is one of the important aspects of soil investigation. These models, called pedo-transfer functions, comprising regression and artificial neural networks (Minasny et al., 2004). Important soil properties which are costly and time consuming in measurement, express as a function of easy obtaining attributes. First, PTFs was made of linear regression analysis, but gradually this analysis was replaced by nonlinear regression. Statistical analyses are based on accurate observed variables; while in the natural systems like soils, precision of observations is lower and the relationships arevague. Thus using the fitting methods of functions which are able to explain the ambiguous structure of system and provide real patterns is essential (Mohamadi and Taheri, 2005). In this regard, the artificial intelligence models which designed based on the nervous system of the human brain in the learning process is increasingly taken into consideration. Not requiring historical information about relationships amongst input and output and less sensitivity regarding error in data, are advantages of these models in comparison with regression transitional functions (Agyare, 2007). In other words, these models are able to predict variations of output variables with the lowest measured parameters and acceptable accuracy. Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 8 April 2019 doi:10.20944/preprints201904.0101.v1 © 2019 by the author(s). Distributed under a Creative Commons CC BY license.