Quantitatively estimating main soil water-soluble salt ions content based on Visible-near infrared wavelength selected using GC, SR and VIP

Study area

Hetao Irrigation District (HID), with Yin Mountains at its north, the Yellow River at its south, Ulanbuh Desert at its west and Baotou at its east, lies in Bayannur League, Inner Mongolia, China. It consists of irrigation areas of Ulan Buh, Jiefangzha, Yongji, Yichang and Urat, and it is China’s largest irrigation district with a total size of 5740 km² (Yu et al., 2010). In addition, HID is an important production base of cereal and oil plants in China with major crops of wheat, corn and sunflower. Shahaoqu Irrigation Area (SIA), a typical region of saline soil in HID, was chosen as the study area. SIA (107°05′∼107°10′E, 40°52′∼41°00′N) is located in the central east of Jiefangzha Irrigation Area. SIA belongs to typical continental climate, having hot summers, chilly winters, rare precipitation and strong evaporation. Its mean annual temperature, precipitation, potential evaporation is about 7.1 °C, 155 mm and 2,000 mm, respectively. Physiographically, the mean elevation and slope of SIA are about 1,030 m and 1/10,000, respectively. According to the World Reference Base for Soil Resources (WRB), the local soil texture is mainly silty clay loam with varying degrees of saline soil. Over the years, due to its gentle terrain slope, poor groundwater runoff, intense land surface evaporation and irrational farming activities, about 60% of the land within the district has been affected by various degree of salinization, which seriously restricted the agricultural development (Wu et al., 2008; Gao et al., 2015).

Sample collection and chemical analysis

The Hetao irrigation district administration gave field permit approval to us (NO. 2017YFC0403302). To ensure the representativeness of soil samples, the samples were randomly gathered from a total of 120 sampling units on a grid of 16 m ×16 m (because the spatial resolution of GF-1 satellite imagery is 16 m) in the study area during October 12∼22, 2017 (Fig. 1). In each unit, approximately 0.5 kg of topsoil (0–5 cm) was collected at four randomly selected sampling sites and then mixed thoroughly to obtain a representative sample. Overall, a total of 120 soil samples were acquired, and each sample was stored in a plastic bag, labeled and sealed. A portable global position system (GPS) was used to determine the coordinates of sampling points. Subsequently, the soil samples were transported to the lab to receive a series of such treatments as sufficient natural air-drying for two weeks and rubbing through a 2 mm sieve to exclude small stones and other impurities. Each sample was divided into two subsamples to be used for spectra collection and physiochemical analysis.

Each 50 g of soil sample was put into a respective flask, and 250 ml of distilled water (the ratio of water to soil is 5:1) were added into each flask. The water-soluble ion contents were measured in the filtrate obtained from full soaking, oscillation and filtration (Aboukila & Norton, 2017). Ca²⁺ and Mg²⁺ were measured using EDTA titration, Na⁺ and K⁺ flame photometry, CO₃²⁻ and HCO₃⁻ double indicator-neutralization titration, Cl⁻ silver nitrate titration, and SO₄²⁻ EDTA indirect complexometry (Bao, 2000). The content of CO₃²⁻ was too low (approximately 0) in some soil samples because CO₃²⁻ is liable to integrate with Ca²⁺ and Mg²⁺ as sediment in a weak alkaline solution (Table 1). Coefficient of variation (CV) reflects the degree of discreteness, and a positive correlation exists in two variables. The high CV helps to build a robust model (Dai et al., 2015). The grading of CV showed a wide range of variation among different ions, among which the ion contents of K⁺, Na⁺ and SO₄²⁻ are over 100%, showing a strong variability, and those of CO₃²⁻, Cl⁻, Ca²⁺, Mg²⁺ and HCO₃⁻ are between 10% and 100%, having a moderate variability.

Laboratory spectral measurements and pretreatments

The soil samples were put into black vessels with a diameter of 10 cm and depth of 2 cm for spectral data collection and the surfaces were smoothed with a straightedge in the laboratory. The spectral data of the soil samples were measured using ASD (Analytical Spectral Devices, Inc., Boulder, CO, USA) FieldSpec^®3 spectrometer with spectral range from 350–2,500 nm. This instrument is equipped with two sensors whose spectral resolutions are 1.4 nm and 2 nm, for the region of 350–1,000 nm and 1,000–2,500 nm, respectively. The spectral data was measured in a dark room with the light sources which have halogen lamps of 50 W, 50 cm from the sample soil surfaces, and 30° incident angle to reduce the effects of external factors to the minimum. The field angle of fiber-optics probe is 5°, and it is 15 cm from the sample soil surface. The light source and spectrometer had been fully preheated, and the spectrometer had been corrected with a standardized white panel (99% reflectance) prior to each measurement to reduce measurement error. Each sample soil was measured in four directions (3 turns, each is 90°), the spectrum was collected five times in each direction, and altogether there were 20 curves of the spectrum (Hong et al., 2018b). These curves were used as the raw spectral reflectance (R_raw) after having the arithmetic mean in ViewSpecPro software version 6.0. The gaps of the spectral curves near 1,000 nm and 1,800 nm were corrected using the Splice Correction function (Xiao, Li & Feng, 2016a).

The fluctuation would affect the accuracy of subsequent modeling because of such disturbance as the external environment, instrument noise and random error in spectral data collection. In general, a series of effective pretreatment, including smoothing, resampling and transformation etc., can eliminate the external noise to some degree, and then enhance the spectral characteristics (Ding et al., 2018). Therefore, it is necessary to pretreat R_raw in the following steps. (i) The marginal wavelength (350–399 nm and 2,401–2,500 nm) of higher noise in each soil sample was removed, then remaining spectrum data was smoothed with filter method (window size is 5 and polynomial order is 2) using Savitzky-Golay (SG) (Savitzky & Golay, 1964) via Origin Pro software version 2017SR2. (ii) The spectral data between 400 and 2,400 nm was resampled with a 10 nm of sample interval to keep the spectral features and remove redundant information (Xu et al., 2016). A new spectral curve consisting of 200 wave bands was obtained. (iii) The precise R_raw−SNV was obtained by using the standard normal variable (SNV) to eliminate the effects of soil particle size, surface scattering and baseline shift on the spectrum data (Xiao, Li & Feng, 2016b; Barnes, Dhanoa & Lister, 1989). The spectral curves of R_raw and R_raw−SNV are shown in Figs. 2A and 2B. Notably, comparison indicated that the spectral curve in Fig. 2B was much smoother than that in Fig. 2A, which made for the subsequent modeling.

Gray correlation (GC)

The GC, as one grey system theory, seeks the primary and secondary relations and analyzes the different effects of all the factors in a system (Deng, 1982; Li et al., 2016). Its calculation process is as follows: the reference sequence is $X_0=\left\{x_0\left(t\right),t=1,2,\dots ,n\right\}$, the comparative sequence is $X_i=\left\{x_i\left(t\right),t=1,2,\dots ,n\right\}$, and the formula of the gray correlation degree (GCD) between X₀ and X_i is (1)${\displaystyle \mathrm{GCD}=\frac{1}{n}\sum _{t=1}^n\gamma \left(x_0\left(t\right),x_i\left(t\right)\right)}$ where $\gamma \left(x_0\left(t\right),x_i\left(t\right)\right)=\frac{{min}_i{min}_t\left|x_0\left(t\right)-x_i\left(t\right)\right|+\rho {max}_i{max}_t\left|x_0\left(t\right)-x_i\left(t\right)\right|}{\left|x_0\left(t\right)-x_i\left(t\right)\right|+\rho {max}_i{max}_t\left|x_0\left(t\right)-x_i\left(t\right)\right|}$

ρ is the distinguishing coefficient within $\left[0,1\right]$ . ρ was set as 0.1 in this paper.

The inconsistent dimension between the spectral data and the contents of different ions has some effects on the data analysis. Therefore, normalizing the spectral data preprocessing method can reduce these disadvantageous effects (Liu, Yang & Wu, 2015; Wang et al., 2018b). In this paper, the larger the GCD of a certain band is, the closer relation the band and the ion content has, and vice versa.

Variable importance in projection (VIP)

The VIP is a variable selection method based on PLSR (Oussama et al., 2012). The explanatory power of the independent variables to the dependent variables is achieved by calculating the VIP score. The independent variables are sequenced according to the explanatory power (Qi et al., 2017). The VIP score for the j-th variable is given as: (2)${\displaystyle {\mathrm{V\; IP}}_j=\sqrt{\frac{p\ast \sum _{f=1}^F{\mathrm{SSY}}_f\ast {\mathrm{W}}_{jf}^2}{{\mathrm{SSY}}_{total}\ast F}}}$

Where p is the number of independent variables; f is the total number of components; SSY_f is the sum of squares of explained variance for the f-th component and p the number of independent variables. SSY_total is the total sum of squares explained of the dependent variable. ${\mathrm{W}}_{jf}^2$ gives the importance of the j-th variable in each f-th component. The higher value VIP_j has, the stronger explanatory power the independent variable has over the dependent variable. The VIP scores of independent variables have been recognized as a useful measure to identify important wavelengths when the score is more than 1 (Wold, Sjöström & Eriksson, 2001; Maimaitiyiming et al., 2017).

Model construction and validation

Two-thirds of the samples were used for modeling (n = 80) and one third for validation (n = 40) using Kennard-Stone (K-S) to calculate the Euclidean distance among different samples to ensure the statistical characteristics of modeling and the validation datasets resembled that of the whole sample set (Kennard & Stone, 1969).

The PLSR and SVR models were applied to the quantitative inversion of different water-soluble salt ion contents in the saline soil in this paper. The PLSR model is a new stoichiometric statistical model. Compared with the traditional multivariate least squares regression (MLSR), PLSR can overcome the multicollinearity among the variables, reduce the dimension, synthesize and filter the information, extract the aggregate variables with the strongest explanatory power in the system, and exclude the noise with no explanatory power (Wold, Sjöström & Eriksson, 2001). The optimal fitting model was built using the number of optimal principal components through full cross validation. SVR model is a new machine learning method based on the principle of structural risk minimization provided by the statistical learning theory. This model is characterized by its ability of solving such problems as limited sample size, nonlinear data processing and spatial pattern recognition of high-dimension data (Vapnik, 1995). During the modeling in this study, the type of SVR and kernel were set as epsilon-SVR and linear function, respectively; the penalty parameter C and nuclear parameter g were acquired by a grid-searching technique and a leave-one-out cross validation procedure. The optimal values of C and g were selected when the minimum RMSE_CV (root mean squared error of cross validation) was produced (Xiao, Li & Feng, 2016b). The two models were constructed and validated using the Unscrambler software version X10.4 (CAMO AS Oslo, Oslo, Norway).

Precision indices of determination coefficient of calibration (R_c²), determination coefficient of prediction (R_p²), root mean squared error (RMSE) and ratio of performance to deviation (RPD) were used to evaluate the performance of these models. RPD classification was adopted to facilitate the interpretation of predictive results: a model is considered as excellent when RPD ≥ 2.5, as very good when 2.0 ≤ RPD < 2.5, as good when 1.8 ≤ RPD <2.0, and as satisfactory when 1.4 ≤ RPD <1.8 and can only distinguish between high and low values when 1.0 ≤ RPD <1.4 (Viscarra Rossel, Taylor & McBratney, 2007). Generally, the most robust model would be the one with the largest R_c², R_p² (approach to 1) and RPD value and the lowest RMSE value.

Correlation between water-soluble salt ions content and spectral reflectance

The correlation coefficients (Pearson correlation) between each soil salt ion content and R_raw−SNV in the range of 400–2,400 nm were tested with the significance level of P < 0.01 (|r| = 0.234 or above). The curves of correlation coefficients of soil salt ions were plotted in Fig. 3 and the numbers of bands passing the significance test were counted in Table 2.

The curve patterns of SO₄²⁻, Cl⁻, Ca²⁺, Mg²⁺, K⁺ and Na⁺ were similar (Fig. 3). From 400 nm to about 550 nm, the correlation coefficients rose sharply from negative to positive, moved with a gentle depression until 1,400 nm, plummeted and surged up to 1,560 nm (among the curves, the change of Ca²⁺ was the sharpest), and maintained a relative stable state to 1850 nm. And then from 1,850 to 2,400 nm, dramatic oscillating variations alternated between rise and fall. In the intervals of 400–1,400 nm and 1,850–2,400 nm the curve pattern of CO₃²⁻ was similar to that of other ions such as SO₄²⁻. But between 1,400 nm and about 1,850 nm, the curve took on a unique pattern: sustained oscillating rise. The coefficient curve of HCO${}_3^{-}$ displayed a smaller variation, smoothly fluctuating between −0.2 and 0.2. The complex variation of the coefficient curves of different ions revealed rich spectral information.

Selection of characteristic wavelength

Characteristic wavelength selection based on SR method

Feature band intervals were selected by stepwise regression method in SPSS software version 23.0 (IBM, Chicago, IL, USA), and the significance levels of variables acceptance and rejection were set at 0.10 and 0.15 (Zhang et al., 2018). The parameter indexes of feature band intervals selection were shown in Table 4 by stepwise regression method at maximum adjusted R².

Table 4:

Parameter indexes of feature band intervals selection by stepwise regression method.

Water-soluble salt ions	Sensitive band numbers	Band intervals/nm	Adjusted R2	Standard error	Sig.
Ca2+	7	1,040∼1,050, 1,090∼1,100, 1,900∼1,910, 1,920∼1,930, 2,200∼2,210, 2,310∼2,320, 2,370∼2,380	0.942	0.529	<0.001
Cl−	8	730∼740, 910∼920, 1,890∼1,900, 1,970∼1,980, 1,990∼2,000, 2,180∼2,190, 2,200∼2,210, 2,290∼2,300	0.975	1.063	<0.001
CO32−	4	1,280∼1,290, 1,360∼1,370, 1,380∼1,390, 1,420∼1,430	0.836	0.012	<0.001
HCO3−	3	2,200∼2,210, 2,260∼2,270, 2,290∼2,300	0.934	0.085	<0.001
K+	6	740∼750, 810∼820, 1,160∼1,170, 1,890∼1,900, 2,210∼2,220, 2,390∼2,400	0.817	0.706	<0.001
Mg2+	6	1,130∼1,140, 1,930∼1,950, 1,990∼2,000, 2,100∼2,110, 2,170∼2,180	0.973	0.152	<0.001
Na+	6	740∼750, 820∼830, 1,860∼1,870, 2,210∼2,220, 2,260∼2,270, 2,390∼2,400	0.942	1.812	<0.001
SO42−	6	610∼620, 1,140∼1,150, 1,960∼1,970, 2,210∼2,220, 2,290∼2,300, 2,390∼2,400	0.947	3.255	<0.001

DOI: 10.7717/peerj.6310/table-4

Great difference existed among the optimal SR models of different ions, and the numbers of band intervals accepted by the model range from 3 to 8 (Table 4). The SR model fitted well with the adjusted R² greater than 0.8 when the number of selected independent variables was considered. Meanwhile, SR model of each ion was statistically significant (p < 0.001). Therefore, the band intervals selected by the SR models were used as the independent variables of PLSR and SVR models.

Table 5:

Max VIP scores and band intervals of soil water-soluble salt ions content with standard normal variable reflectance.

Water-soluble salt ions	Sensitive band numbers	Maximum VIP scores	Maximum VIP scores intervals/nm
Ca2+	69	1.97	1,440∼1,450
Cl−	85	1.42	560∼570
CO32−	67	2.01	1,440∼1,450
HCO3−	79	2.37	1,410∼1,420
K+	69	1.73	1,880∼1,890
Mg2+	69	1.49	1,870∼1,880
Na+	83	1.55	1,880∼1,890
SO42−	74	1.74	1,880∼1,890

DOI: 10.7717/peerj.6310/table-5

Construction and analysis of PLSR model

The sensitive bands were obtained using different band selection methods of GC, SR and VIP to build PLSR model. The results of PLSR model were shown in Table 6.

The models of the six ions Ca²⁺, Cl⁻, CO₃²⁻, Mg²⁺, Na⁺ and SO₄²⁻ performed well using VIP method (R_c² is close to 1). The models based on the bands of Ca²⁺, Cl⁻, Mg²⁺, Na⁺ and SO₄²⁻ selected using the SR method displayed good fitting effect, and those of Ca²⁺, Mg²⁺ and Na⁺ using the GC method exhibited good fitting effect.

In terms of verification accuracy, VIP method had excellent prediction of Ca²⁺, Na⁺, SO₄²⁻, SR method had excellent prediction of Ca²⁺, Mg²⁺, Na⁺, SO₄²⁻ (the RPD of Ca²⁺ was up to 3.95), and GC method did not show strong prediction power over any ions. On the contrary, all the three models demonstrated poor forecasting power over HCO₃⁻. The RPDs of SR-HCO₃⁻ and VIP-HCO₃⁻ were 0.64 and 0.93 respectively. Therefore, VIP method had the best modeling effect and SR method had the best forecasting effect, and GC method had poor modeling and forecasting effects on the salt ions inversion in the PLSR model.

Construction and analysis of SVR model

The sensitive bands were obtained by using different band selection methods of GC, SR and VIP to build SVR model. The results of SVR model were shown in Table 7.

The modeling accuracy of SVR model was similar to that of PLSR model. But the verification accuracy of ions was different between the two models. VIP method had the excellent prediction of Ca²⁺, Cl⁻, Mg²⁺, Na⁺, SR method had the excellent prediction of Ca²⁺, Mg ²⁺, Na⁺, SO₄²⁻, and GC method did not show strong prediction power over any ions. The prediction results of Ca²⁺ were the best: the RPD of VIP and SR models were 3.93 and 3.97, respectively. Overall, in the SVR model, VIP method exhibited the best performance for modeling and predicting the salt ions content, SR method was the second, and GC method was relatively poorer.

Comparison among the results of different salt ions content in estimating

The optimal band selection method varied in some degree from the optimal modeling method (Tables 6 and 7). The comparison was made between the measured value and the estimated value of all the ions concerned under the optimal model (Fig. 6). The sequence of the forecasting power of the ions was Ca²⁺ > Na⁺ > Cl⁻ > Mg²⁺ > SO₄²⁻ > CO₃²⁻ > K⁺ > HCO₃⁻, and it was the same as that of the modeling power.

Obviously, the verification result showed that most data points of the five ions, Ca²⁺, Na⁺, Cl⁻, Mg ²⁺ and SO₄²⁻, were concentrated near line 1:1. The optimal models of these five ions had very strong predicative power with the RPD above 2.5 (Tables 6 and 7). Compared with the previous researches, model prediction effects of K⁺ and Na⁺ (Qu et al., 2009); Ca²⁺, Na⁺ and Mg²⁺ (Viscarra Rossel & Webster, 2012); HCO₃⁻, Ca²⁺, Cl⁻, Mg²⁺ and SO₄²⁻(Dai et al., 2015); HCO₃⁻, Ca²⁺ and SO₄²⁻ (Peng et al., 2016a); K⁺, Na⁺, Ca²⁺ and SO₄²⁻ (Wang et al., 2018a) were satisfactory. Although the results of this study are not exactly the same as these previous researches, it still shows the rationality own to some extent. In addition, this result shows that band selection has realized the goal of removing the irrelevant information, and plays a major role in improving the inversion accuracy of salt ions.

In Fig. 6, the data points of CO₃²⁻ and K⁺ were relatively dispersed in the verification result. The CO₃²⁻ had a relatively good predictive power (RPD = 1.80) and the K⁺ had a normal predictive power (RPD = 1.43). Notably, HCO₃⁻ had no predicative power (RPD = 0.96) because the slope was under the 1:1 line and the data points were most discrete (Fig. 6D). The predicting effect of HCO₃⁻ was different from that of Peng et al. (2016a) and Dai et al. (2015), but similar to that of Wang et al. (2018a). The cause of this result needs to be further studied. Overall, it is vital to make some efforts to improve the robustness and accuracy of these ion models. Xiao, Li & Feng (2016b) failed to predict Na⁺, Mg ²⁺ and Ca²⁺, but applied the SVR model to forecasting SAR after the SNV transformation and the performance was satisfactory (RPD = 2.13). Analogously, first derivative reflectance (FDR) index was calculated to effectively predict SAR by Xiao, Li & Feng (2016a). In addition, Viscarra Rossel & Webster (2012) forecasted the content of Na⁺ after logarithmic pretreatment with VIS-NIR spectral technique (RPD = 2.10). Thus, salt ion indexes construction and variable transformation processing are helpful approaches to improve the correlation with the spectra so as to establish satisfactory models.

A little difference existed in the applicability between PLSR and SVR models on inversing the content of ions. Both methods could produce satisfactory results in conformity with that of Peng et al. (2016a). In addition, the optimal inversion models and prediction models for each ion were different: SR-PLSR model and SR-SVR model for Ca²⁺, VIP-SVR model and SR-PLSR model for CO₃²⁻, SR-PLSR model and VIP-PLSR model for K⁺, VIP-PLSR model and GC-PLSR model for HCO₃⁻, respectively. Among them, the performance of the optimal inversion model of Ca²⁺ resembled that of the prediction model. The results suggested that the ion models with poorer performance frequently demonstrated uncertainty in the inversion process (Peng et al., 2016a). Generally, as the major water-soluble ion components in the two highly soluble salts of sodium and kali, Na⁺ and K⁺ exhibit great difference in the spectral characterization degree (Dai et al., 2015). Therefore, the spectral characters of water-soluble salt ions are not necessarily determined by the number of dissociative ions, so more pertinent experiments and analysis should be conducted to explore the response mechanism.

Correlation analysis and inversion performance

The raw spectral reflectance curve of each soil sample presented distinct shapes (Fig. 2A). One of the prime reasons for this phenomenon is that the absorption features in these soil samples were related to soil salt crystal contents and types, as well as various chemical bonds (e.g., C-H, O-H, N-H). The results were in accordance with those in previous studies (Viscarra Rossel et al., 2006; Viscarra Rossel & Webster, 2012; Dai et al., 2015; Peng et al., 2016a; Wang et al., 2018a), which demonstrated that soil VIS-NIR spectra could be used to determine part of soil salt ions contents in some degree.

Traditionally, correlation analysis helps reveal the relationships between soil salt ions content and VIS-NIR spectra, and it indicates modeling effects to some degree (Weng, Gong & Zhu, 2008). In the current research, the number of the significant bands of different ions could be sequenced from the largest to the smallest as follows: Cl⁻ (96%) > Ca²⁺ (95%) > Mg²⁺ (93%) > Na⁺ (90.5%) > K⁺ (89%) = SO₄²⁻ (89%) > CO₃²⁻ (73%) > HCO₃⁻ (0.5%), the correlation coefficients of different ions ranged from the largest to the smallest as: Cl⁻ (−0.882) > Ca²⁺ (−0.877) > Mg²⁺ (−0.848) > Na⁺ (−0.752) > SO₄²⁻ (0.749) > K⁺ (0.630) > CO₃²⁻ (0.552) > HCO₃⁻ (0.235) (Table 2). Thereby, five ions (Cl⁻, Ca²⁺, Mg²⁺, Na⁺ and SO₄²⁻) had more significant relationship with reflectance spectra. Although there were some differences between forecasting power ranking and correlation ranking, the optimal models of these five ions had the excellent predictive results (Fig. 6). Nevertheless, the other three ions (K⁺, CO₃²⁻ and HCO₃⁻) had weak correlations and unsatisfactory predictive power. In particular, HCO₃⁻ had only one significant band and the worst prediction effects. But in most cases, the sensitive band numbers of HCO₃⁻ were not the least in comparing the results of the three wavelength selection methods (Tables 3–5). Thus, we conjecture that the different calculation mechanisms cause a certain inconsistency between modeling performance and sensitivity. In addition, the optimal method of finding out their responding spectrum varies from one ion to another in the soil. In future study, it is practically significant to adopt various methods to select the optimal bands in the inversion of soil ions.

Effects of wavelength selection on estimation models

The massive complex spectra often contain a large amount of redundant information irrelevant to the ions contents. The selection of feature spectra is hence a critical step to create a robust model. From Tables 3–5, we could see the great difference exist in the number of wavelength selected with the three methods: VIP method had the largest number of wavelengths (34.5%∼42.5%), SR method had the smallest number of wavelengths (1.5%∼4%) and number of wavelengths (7%∼55%) varied greatly by GC method.

Our experiment with three wavelength selection methods also indicated that different methods yielded different results. Among the three methods, the VIP method produced the best results, followed by SR method, while the GC method performed least ideally. We argue that the GC method is not necessarily an inappropriate method as some results are still acceptable. However, GC method could distinguish the primary relationships among the factors in the system by calculating and comparing GCD (Deng, 1982; Liu, Yang & Wu, 2015). In the field of spectral analysis, the application of GC method could better identify sensitive spectral indices, select sensitive bands and optimize inversion model (Li et al., 2016). On the other hand, Wang et al. (2018b) used GC method to extract the feature bands of soil organic matter content to construct the model with stronger generalization capability. Therefore, the soil compositions have a strong impact on the performance of spectral model. This conclusion is consistent with previous research results (Viscarra Rossel et al., 2006; Viscarra Rossel & Webster, 2012; Xiao, Li & Feng, 2016b). The VIP values were calculated with VIP method, in the process of PLSR analysis to further evaluate the significance of each wavelength for model prediction (Wold, Sjöström & Eriksson, 2001; Maimaitiyiming et al., 2017; Qi et al., 2017). VIP method often produces the best results in the modeling set because it can distinguish between useful information and inevitable noises in the set. Oussama et al. (2012) adopted this method to reduce almost 75% of the total data set for a simplified model of high accuracy. Additionally, as a simplified regression linear model, SR method not only preserves significant bands but also solves multicollinearity problems effectively (Xiao, Li & Feng, 2016a; Xiao, Li & Feng, 2016b). It has great optimization effect on model complexity by adjusting the significance level of selected and excluded variables (Zhang et al., 2018). Compared with the selection results with VIP method, SR method could be used to extract fewer bands to establish ions (except for K⁺, CO₃²⁻ and HCO₃⁻) forecasting models with RPD above 1.80. Therefore, it is meaningful to make further simplification of the model while ensuring its accuracy.

Research limitations

This study clearly demonstrated that VIS-NIR spectral analysis technique is an effective method to detect salt ions content of salinity soil in the irrigated district. In terms of extracting feature wavelengths to estimate ions content, our work provides a comprehensive comparison and evaluation approaches. Such endeavor is critically and practically important to further enhance the model performance of the soil salt ions. The application of machine learning algorithms with strong applicability to solve nonlinear relationship between variables, such as Ant Colony Optimization-interval Partial Least Square (ACO-iPLS), Recursive Feature Elimination based on Support Vector Machine (RF-SVM), and Random Forest (RF) has been proved to be a useful approach to obtain the effective information of soil organic matter (Ding et al., 2018). To further improve the prediction accuracy, the more machine learning algorithms should be applied to the analysis of sensitive spectral regions and the construction of stable models in future study. In addition, the application of multi-source remote sensing platforms such as Landsat, GaoFen-5, Hyperion and unmanned aerial vehicle (UAV) in soil salt ions estimation has not been investigated. Therefore, further research should focus on the possible combination of multiple approaches and remote sensing data at different scales to estimate soil salt ions content.