1. Introduction
Salt affected soils are caused by excess accumulation of salt which are pronounced at the soil surface. Salt is often derived from geological formations featuring shale, marl, limestone, sylvite, gypsum, and halite, but variability of soil salinity is mostly due to parent material, soil type, and landscape position (Clay et al., 2001). Moreover, salts can be transported to the soil surface by capillary action from brackish water tables and can accumulate due to evaporation; they can also accumulate as a result of anthropogenic activities such as fertilization or oil production. Soil salinization is a universal problem and current estimations of the proportion of salt-affected soils in irrigated lands for several countries were 27 % in India, 20 % in Australia, 28 % in Pakistan, 50 % in Iraq and 30 % in Egypt (Stockle, 2013). The accumulation of soluble salts in the root zone greatly affect plant growth, resulting in lower crop yields and adversely affecting the soil fertility (Li et al. 2013).
Soil salinity is typically assessed by measuring the soil electrical conductivity in saturated paste extracts (ECe) or by using extracts with different soil-to-water ratios (Sonmez et al. , 2008). Developed in the mid-1950s, ECe is one of the most widely reported soil quality assessment parameters (Karlen et al., 2008), regular monitoring of which is essential for efficient soil and water management and sustainability of agricultural lands (Bilgili et al., 2011). Electrical conductivity can act as an indirect indicator of important soil physical properties (Rhoades et al., 1999) and provides important information about the impact that farm practices, such as irrigation and soil and crop management, have at both the field and regional scales. Therefore, reliable information on the nature and spatial extent of soil salinity is a prerequisite for restoring fertility and preventing further degradation. Thus, timely detection of the extent and magnitude of soil salinity is important for agriculture practices.
It is difficult to obtain up-to-date soil salinity information by using conventional techniques, to identify and monitor soil salinity because these techniques are time consuming and expensive and require high sampling densities and frequencies; hence efforts are being made to obtain more cost-effective methods for mapping soil salinity. During the last two decades, visible and near-infrared spectroscopy has been used as a rapid, cost-effective and relatively accurate method for analyzing conventional soil properties (Nocita et al., 2015). Previously, several studies indicated that pure sodium chloride is featureless in Vis-NIR regions because salt is not a strong or direct chromophore (Metternicht et al., 1997). However, the presence of salts in soils may result in subtle spectral responses when combined with—OH, which is common in soils. Therefore, soil salinity can be characterized by soil spectral reflectance or salinity spectral indices using partial least squares regression, artificial neural network, and stepwise multiple linear regression methods (Zhang et al., 2011) and can be detected using high-resolution spectroscopy (Jin et al.,2015). Accordingly, interest in using reflectance spectroscopy as a rapid and effective tool for mapping soil salinity has recently grown and several studies have estimated salt contents of air-dried soils with reasonable accuracy using reflectance spectroscopy (Yong-Ling et al., 2010).
Hyperspectral visible and near-infrared reflectance spectroscopy displays promise as a result of its performance, accuracy and cost effectiveness in the determination of most soil properties (Shepherd and Walsh 2002). Various statistical modeling techniques help to correlate a single reflectance spectrum of soil to a host of physical, chemical, mineralogical and microbiological attributes of that soil after proper calibration and validation of models. Principal component regression (PCR), partial least squares regression (PLSR), multivariate adaptive regression splines (MARS), artificial neural networks (ANN) are some of the commonly used diagnostics for calibration and validation of hyperspectral models (Bilgili et al. , 2010). The reliability of calibration of spectral data with chemical analysis data needs to be enhanced by factoring in variations on account of land use and choice of scale. The calibration process also needs to be made indubitable by using optimal sample size and sampling strategy. Once the calibration models between soil reflectance spectra and soil variables have been established, they can be used to predict unidentified parameters. Several regression methods based on visible near IR have been used to estimate soil salinity, and partial least-squares regression is the most common (Farifteh et al. 2007; Bilgili et al. 2011). The PLSR approach has inference capabilities that are useful for modelling a probable linear relationship between the measured reflectance spectra and salt content in soils (Farifteh et al. 2007). The MARS method is considered a nonparametric method that estimates complex nonlinear relationships among independent and dependent variables (Friedman 1991), and it has been effectively applied in different fields (Bilgili et al. , 2010; Felicísimo et al., 2012) and generally exhibits high performance results compared with other linear and non-parametric regression models, such as principal component regressions, classification and regression trees and artificial neural networks.
This study was conducted to evaluate multivariate regression models to predict electrical conductivity using Vis-NIR and MIR Spectra as a substitute to conventional soil analysis. The specific goals of this study were to find out sensitive regions of the spectrum for modelling electrical conductivity and compare the performance of multivariate regression models PLSR, RF, SVR, and MARS for predicting EC, both in the Vis-NIR and MIR spectral region.