This study was conducted on Beiwucha Farm located in the Manasi River Basin, the north of Tianshan Mountains, Xinjiang Uygur autonomous region, in northwestern China (Figure 1). It was reclaimed in 2011 and cultivated cotton in 2012-2014. The research field covers 0.385 km² (550 × 70 m) with a 0.2% - 0.4% slope northwest to southeast; the area is 380 m above sea level and experience an arid climate with mean annual temperature, precipitation, evaporation and frost-free period of 7.6˚C, 148.5 mm, 1972 mm and 187 d over a 20-year

period, respectively.

The farmland for the present study was reclaimed in 2011 and cotton was planted in 2011-2014. Soil salinity is known as a most significant problem in this farm and large areas of salt-affected land was observed due to a very saline shallow water table (average mineralization of 35.0 g∙L⁻¹and water table of 0.5 to 2.5 m).

Initially, observation soil sites were arranged on a grid in the study plot, with an average interval of 10 m from northeast to southeast (X) and 10 m from southeast to northwest (Y) (Figure 1), and a total of 250 soil sites were selected. At each site, an electromagnetic induction (type EM38) was placed on the ground surface and measurements in both the vertical (EM_v) and horizontal (EM_h) modes of operation were made. Later 70 sites which covered a wide range of soil salinity were randomly chosen from the total 250 soil sites for soil sampling. We were focused on 0 - 80 cm arable layer for cotton root growth regions because of using drip irrigation under mulch in study area; soil samples were obtained by hand augering at 0 - 20, 20 - 40. 40 - 60 and 60 - 80 cm soil layers intervals to a depth of 0.8 m for the laboratory analyses of electrical conductivity (EC). In all, EM measurements (EM_h and EM_v) of 250 sites and soil samples of 70 in situ cores (280 soil samples) were collected. Soil sampling and EM gauging was carried out from March 25^th to April 5^th 2013 when soil moisture conditions were relatively uniform across the plot.

All soil samples were air-dried, crushed, and passed through a 2-mm sieve. The 1:5 soils: water suspensions were prepared by weighing 12 g of soil into a pop-top tube, adding 60 ml of deionized water, and shaking for 30 min on an end-over-end shaker. After being left for 30 min, the EC_1:5 of the supernatant were measured with a conductivity meter directly (Miller et al. 2006). In addition, 200 soil samples of 50 profiles that varied greatly in soil salinity were chosen and prepared for analysis of total dissolved salts (TDS).The TDS of soil samples was determined by summing up the dominant cations (Ca²⁺, Mg²⁺, Na⁺, K⁺) and anions (, , Cl⁻,) of 1:5 soil water extracts. A comparison between EC_1:5 and the TDS showed a strong positive linear relationship, and the equation of the fitted line relating the TDS (g∙kg⁻¹) to EC_1:5 (dS∙m⁻¹) values were given by:

The TDS values of the rest 80 soil samples (20 profiles) whose TDS values were not measured were estimated from EC_1:5 on the basis of this relationship (r² = 0.9517). The TDS value was used as a surrogate for the widely used EC_e to determine soil salinity in our study.

We used cross-validation to validate the accuracy of interpolation algorithms and examine the difference between the measured values and the predicted values using mean absolute errors (MAE) or relative MAE (MAE %) (Equation (2)), root-mean-square error (RMSE) or relative RMSE (RMSE %) (Equation (3)), ratio of prediction to deviation (RPD) (Equation (4)) and model efficiency factor (E) (Equation (5)). The RPD indicates strength of statistical correlation between measured and predicted values. MAE, RMSE and E values indicate degree of agreement between measured and predicted values. Detailed descriptions and definitions of these model performance parameters are given by [32] [33] .

where is the observed (measured) value of Z at locations, is the predicted value at the same locations, , is the average measured and predicted value, respectively. N is the number of values in the dataset.

Conventional statistical analyses were conducted using the software package SPSS 17.0 for Windows (SPSS Inc., Mat Lab, USA). Geostatistical analyses and mapping were performed by using Arc GIS 10.0 software package (Environmental Systems Research Institute, Redlands, CA).

Fifty sites out of the 70 in situ sampling sites were selected for the spatial simulation of soil salinity, and the rest 20 sites were used for validation of IDW, OK algorithms. In the calibration procedure, the TDS data was selected as the dependent variable and apparent soil electrical conductivity (EM_h and EM_v) made by the EM38 as the independent variable. The linear regression model was then fitted the data from the selected 50 sampling sites. This yielded of predictive equations across the whole soil profile (Table 1). Readers are referred [34] - [36] for similar approaches with variations.

Both EM_h and EM_v measurements showed significant linearity with the TDS at each soil layer. Multiple linear regression using EM_h and EM_v as independent variables gained higher regression coefficients than those fitted by linear regression (Table 1), the correlation coefficient of 0 - 20, 20 - 40, 40 - 60 and 60 - 80 cm were 0.9090, 0.9228, 0.896 and 0.9085 respectively. The multiple linear predictive equations were then used to reconstruct the salinity profiles at the 180 soil sites where EM measurements were made.

Summary statistics and normal test of soil salinity from multiple linear regression datasets was performed in different soil layers, and the results, present in Table 2, indicated that the average soil salinity value was comparatively high with regard to the generally accepted limit for most agricultural crops. The minimum and maximum of soil salinity were 9.56 g/kg in 20 - 40 cm soil layer and 40.06 g/kg in 0 - 20 cm soil layer, respectively, the means were greater than 24.0 g/kg in different soil layers. As it resulted also from the parameters of one- sample Kolmogorov-Smirnov (K-S) normality test (p < 0.05, two-tailed), skewness and kurtosis, the data of all variables did follow normal distributions.

The experimental semivariogram γ(h) measuring the spatial autocorrelation between data pairs as a function of the displacement between the pairs was calculated and the scatter plot of γ(h) vs. h (lag distance) was gained. Then different theoretical semivariance models were used to fit the calculated values, and the model with the best fitting value and the smallest nugget value was selected [37] . The fitted parameters of the semivariograms and cross-semivariograms at both orientations were summarized in Table 3. These semivariogram parameters include the nugget value C₀, partial sill (C), sill (C₀ + C), range value a and nugget-sill-ratio (). Validation information used to determine the goodness of fit includes determination coefficient (r²), residual sum of squares (RSS) and significance level of F test.

More often than not, positive nuggets were observed during surveys of most soil properties [38] . In this study, namely nugget-to-sill slightly positive nuggets of 5.58%, 0.97%, 0.62% and 0.34% were observed for soil salinity in different soil layers (Table 3), respectively. Cambardella et al. [39] defined nugget- sill-ratio of <25%, 25% - 75%, and >75% as categories of strong, moderate, and weak spatial dependence, respectively. In this study, indicating the spatial variance caused by structural factors (such as climate, matrix and topography) is much higher than that caused by stochastic factors (such as tillage, irrigation, fertilizer application, etc.) at whole study scale; the reclaimed farmland is also one of reasons for few human agricultural practices, perhaps. The goodness of fitting was determined by determination coefficient (R²), residual sum of squares (RSS) and significance level of F test. The results, also shown in Table 3, were quite satisfactory for all variables at different soil layers, because the statistics used, i.e. residual sum of squares was close to 0 and F test was significant at level of p < 0.05.

Ordinary kriging and IDW methods were applied to estimate the soil salinity of EM38 data at unsampled locations

using semivariogram models in Table 3. Considering the whole application, including trend analysis, semivariogram calculation and interpolation was applied to generate the spatial distribution map. Raster maps (2 m × 2 m) of soil salinity across the study area were presented in Figure 2.

A jack-knifing procedure of cross-validation was employed to statistically evaluate the performance of the spatial prediction of TDS data. The results showed that root mean square error (RMSE) of ordinary kriging for 0 - 20, 20 - 40, 40 - 60 and 60 - 80 cm prediction was 0.8852, 0.9 077, 0.8926 and 0.9143, respectively, while it was 0.9177, 0.9524, 0.9376 and 0.9408 for IDW, respectively, which indicates high prediction precision and reliability of IDW method to the interpolation plots of soil salinity (Figure 2).

Figure 2 shows the local and regional variation information of soil salinity at each of layers in 0 - 0.8 m soil solum. Soil salinity exhibited strip and block patterns, and similar spatial patterns can be observed from maps of all soil layers. As was plainly illustrated that, for both OK and IDW, the grids of TDS < 15.0 g/kg almost no position, the grids of 15.0 < TDS < 20.0 g/kg generally concentrated in the east-central positions, the grids of 20.0 < TDS < 25.0 g/kg mostly occurred in east-central and south-east portions of the study domain, the grids of 25.0 < TDS < 30.0 g/kg considered as the predominant soil salinization type, spread throughout the whole study plot, the grids of TDS > 30 g/kg were mainly situated in the northwest and southwest sporadically. It was also apparent that the northwest and southeast region of the study area were most salinized compared with other regions, especially at soil layers of 0 - 20 and 60 - 80 cm. This phenomenon can be explained by many reasons. For one thing, this highly salinized position is situated in low-lying area, more influences from the shallow water table and high groundwater mineralization to 60 - 80 cm layer. For another, intense evaporation and scarce precipitation in the study area have induced soluble salt to transport upward to 0 - 20 cm layer.

With the 20 observation soil sites (80 soil samples) randomly selected as external dataset to estimate the prediction yields of IDW and OK methods. Figure 3 shows good performance of TDS by both methods, comparing to coefficient of determination (R²), R² between predictions and observations ranges from 0.6074 to 0.5918 for IDW method and from 0.6219 to 0.5088 for OK method. This result reveals that IDW prediction value has better performance than OK in different soil layers, except 40 - 60 cm.

Based on the statistical parameters as discussed, both IDW and OK model are validated using the model efficiency factor E and RPD of predicted values. The E and RPD between reference measurements, i.e. accurate or good prediction if RPD and E values are higher than 2.5 and 0.80 respectively by Farifteh et al. (2007), suggest an accurate to good prediction. It is observed from Table 4 that IDW method (with E = 0.8873, 0.9075, 0.8483 and 0.901, RPD = 9.64, 8.01, 8.17 and 10.83 in 0 - 20, 20 - 40, 40 - 60 and 60 - 80 cm soil layer, respectively) predicts an accurate TDS slightly better than OK methods. In addition to MAE% and RMSE%, it is considered that IDW performs slightly better, differences between IDW and OK methods in predictions are significant for both variables based on the analysis of covariance (ANOVA) test (P < 0.05).