1. INTRODUCTION
Soil salinization is one of the major land degradation processes in almost all parts of the world. About 836 million hectares (Mha) area of the world (3% of total geographical area) is affected with soil salinization, of which, 48% is saline and 52% is sodic (Singh, 2016). The extent of the salt-affected soils (SAS) in India is 6.73 Mha, of which 1.28 Mha (19%) is in the coastal region (CSSRI, 2018). Unlike the SAS formed by secondary salinization, the SAS of the coastal region possesses unique physical and chemical properties. The co-existence of the soil acidity (low soil pH) with high salinity level (high electrical conductivity, EC) is a unique feature of these soils. The ingression of saline water from sea and estuaries either naturally or through anthropogenic activities is the major reason for salinity in these soils. The excess amount of the salts affects the physical, chemical and biological properties of the soils and ultimately the plant growth (Mahajan et al. , 2016; Yuan et al. , 2007). The productivity of these soils is often very low. Quantitative assessment of the soil properties is very crucial to understand, maintain and improve the soil quality and enhance crop yield (Askari et al. , 2015; Xu et al. , 2018). But, the soil is a heterogeneous resource due to complex processes and mechanisms involved in soil formation (Morellos et al. , 2016). Real-time assessment of soil salinity at high temporal and spatial resolution and other properties are essential to manage these soils efficiently for crop production. Soil sampling and laboratory analysis to adequately assess the spatial and temporal variability of soil properties are time-consuming and expensive (Xuet al. , 2018) and often limited to smaller areas. Such a challenge has attracted researchers worldwide in recent years to find alternate ways to overcome it. Compared to the conventional laboratory analysis methods, hyperspectral remote sensing (HRS) has been proposed as one of the modern, valid and alternate techniques for monitoring the soil properties (Stenberg et al. , 2010). Moreover, multiple soil properties can be estimated using a single representative spectral scan of each sample (Vohland et al. , 2011).
Based on the spectral absorption and reflection features in visible (VIS) and near-infrared (NIR) region, remote sensing technology could be employed for the estimation of soil properties and salt content (Islamet al. , 2003). Researchers have investigated the use of VIS-NIR based spectroscopy for estimating different soil attributes in different soil types. Multivariate statistics is required to build the relationship between the complex features or patterns of soil spectral data and the soil properties (Araújo et al. , 2014; Stenberget al. , 2010). Among different multivariate statistical techniques, commonly used linear techniques are stepwise multiple linear regression (SMLR), principal component regression (PCR) and partial least square regression (PLSR), however, use of non-linear techniques like multivariate adaptive regression splines (MARS), random forest (RF), support vector machine regression (SVMR) are also becoming popular as relationships between spectral data and soil attributes are rarely linear (Araújo et al. , 2014; Xu et al. , 2018). Wanget al. (2018) found that the RF based model with 1.5 order derivative of absorbance was most effective, stable and accurate to quantify the soil salinity with coefficient of determination (R2) = 0.93, root means square error (RMSE) = 4.57 dS m-1 and ratio of performance to deviation (RPD) = 2.78. Cécillon et al. (2009) and Stenberg et al. (2010) investigated the applicability of the VIS-NIR remote sensing and observed absorption bands for NaCl, KCl and MgSO4 at 1930 nm, 1430 nm and 1480 nm, respectively. The use of PLSR and artificial neural network (ANN) to predict the salt concentration (NaCl, KCl, MgCl2 and MgSO4) has been successfully demonstrated by Farifteh et al. (2007). Nawaret al. (2014) recorded good prediction of soil EC based on the MARS model using soil spectral reflectance (R2 = 0.73, RMSE = 6.53 dS m-1, and RPD = 1.96) compared to PLSR. The use of VIS-NIR remote sensing for predicting the soil properties like soil organic carbon (SOC), soil pH, EC, total nitrogen (N), available N, total phosphorus (P), Mehlich 1 extractable P, total potassium (K), cation exchange capacity (CEC), moisture, soil texture, clay content, etc. has been studied widely (Araújo et al. , 2015; Cécillon et al. , 2009; Chang et al. , 2001; Christy, 2008; Morellos et al. , 2016; Schirrmann et al. , 2013; Vasqueset al. , 2008, 2009; Viscarra Rossel et al. , 2006; Wenjunet al. , 2014; Xu et al. , 2018). The non-linear multivariate techniques (SVMR and Back Propagation Neural Network (BPNN)) outperformed the linear techniques (PCR and PLSR) to predict the soil organic matter, total N, total P and total K in soil cores of paddy fields (Xu et al. , 2018). The use of the VIS-NIR spectra for characterizing soils gives a large number of predictor variables but using the full spectra at high-resolution compromises with the multi-collinearity and noise (Vohland et al. , 2011). Thus, the selection of a proper multivariate techniques for calibration and prediction of a variable is an important factor (Mouazen et al. , 2010; Nawar et al. , 2016; Xu et al. , 2018).
To the best of our knowledge, very few studies have been conducted on SAS and almost none on SAS having acidic soil reaction. The objectives of the study were to (1) study effect of the soil salinity on the VIS-NIR spectral reflectance pattern, (2) investigate potential of VIS-NIR spectroscopy to estimate the various properties of SAS and (3) compare the predictive ability of the linear and non-linear multivariate techniques for estimation of soil properties of SAS.