Exploring the potential of spatial artificial neural network in estimating topsoil salinity changes of in arid lands

Soil salinity is one of the most important environmental issues, especially in arid and semi-arid regions, due to the influence of various parameters such as climate variables. Nowadays, the use of computational intelligence systems has expanded as a new strategy for soil studies based on satellite imagery. The purpose of this study is comparison of performance and efficiency of two multivariate regression methods as linear methods, and artificial neural network as a nonlinear method, to model and estimate salinity on topsoil in Jarghouyeh_e_Sofla plain. For this purpose, 61 soil samples were collected from 0 to 10 cm depth in study area and electrical conductivity values were extracted in laboratory. Two types of data were used: electrical conductivity of sampling points as independent variables and satellite data including salinity indices and Landsat Operational Land Imager sensor bands of Landsat8 as associated variables. The combination of input parameters was carried out in regression and neural network by using backward regression and principal component analysis, respectively. Therefore, data were divided into two series: train series (60 to 70%) and test series (20 to 30%). The results of assessment based on correlation coefficient and root mean square error in the neural network and regression was equal to 0.65, 27.74 and 29.9 and 31.85, respectively. It showed that the neural network has the highest precision in forecasting soil salinity.