1. Introduction
Soil salinization is one of the most common worldwide land degradation processes, which seriously affects major soil degradation phenomenon such as soil dispersion, soil erosion, and engineering problems, and thus reduces agricultural production, and impacts environmental health [
1,
2,
3,
4]. According to Food and Agriculture Organization of the United Nations, the global surface area affected by soil salinization was more than 831 million ha by 2000, covering more than 6% of the Earth’s surface [
5]. Therefore, identification of the soil salinity and the salt-affected areas is essential for sustainable agricultural management, especially in arid and semi-arid areas [
6,
7].
Due to its high temporal sampling frequency and rapid nondestructive measurement, reflectance spectroscopy has been widely used to estimate soil salinity [
8,
9,
10,
11,
12,
13]. Hyperspectral data has become a promising source of data in many studies since it contains a large number of narrow contiguous spectral bands with high spectral resolution and thus enables the quantitative measurement of saline soils with different salinity levels [
14,
15,
16,
17]. During the last several decades, many studies have been focused on the effect of salinity on the spectral characteristics of saline soils. Taylor et al. [
18] found absorption valleys at 980 nm, 1170 nm, 1450 nm, and 1900 nm, and strong reflection peaks at 800 nm. Farifteh et al. [
19] carried out a laboratory experiment involving soils with six salt minerals and three soil textures in order to study the quantitative relationship between spectral response and salt concentration of saline soil samples, their results showed that the observed absorption features were broadened at wavelength above 1300 nm and the overall reflectance changed proportionally as the soil salinity increased. Wang et al. [
20] carried out a spectra measurement in of saline soil, the laboratory and found significant absorption features at 505 nm, 920 nm, 1415 nm, 1915 nm, and 2205 nm. Moreira et al. [
21] evaluated the variations in spectral reflectance of soil samples treated with different salt crystals using both principal component analysis and the continuum-removed method, after that they found the obvious absorption depths of major features at 1450 nm, 1950 nm, 1750 nm, and 2200 nm, their results also indicated the best negative correlations between reflectance and EC in the 1500–2400 nm range for CaCl
2 and MgCl
2.
In many studies, spectral indices were commonly used to process the hyperspectral data of saline soil after spectral measurements. Zovko et al. [
22] selected 246 soil samples for visible and near-infrared spectra measurements and defined a new geostatistical spectral index as covariate in ordinary cokriging for predicting electrical conductivity, spatial variation of soil salinity was then mapped in the Neretva River Valley, Croatia to propose an effective approach based on VNIR spectroscopy and geostatistics for mapping soil salinity. In order to predict the salt abundance in slightly saline soils in Burg Al-Arab, Egypt, Masoud [
23] carried out spectrometry measurements for 21 soil samples with different salinity levels to evaluate the simple wetness index (WI) and sophisticated spectral mixture analysis techniques including mixture tuned matched filtering (MTMF) and linear spectral unmixing (LSU).The prediction results of salinity levels showed significant high accuracy with R
2 of 0.88, 0.84 for MTMF, LSU and WI, respectively. After comparing three spectral indices including difference index (DI), normalized difference index (NDI) and ratio index (RI). Nawar et al. [
24] developed regression models with the EC in the El-Tina Plain, Egypt, their results showed RI can provide a reliable estimations of the soil EC with produced R
2, RMSE and RPD values of 0.8, 6.2, and 2.06.
Due to its superiority in dealing with high dimensional multi-collinearity, partial least square regression (PLSR) has been used as a very common multivariate statistical method in predicting the soil salinity. Weng et al. [
25] measured the reflectance spectra (350–2500 nm) from EO-1 Hyperion and tested PLSR as prediction model to evaluate the potential of predicting salt content in the Yellow River Delta of China, their results indicated that spectral bands at 1487–1527 nm, 1971–1991 nm, 2032–2092 nm, and 2163–2355 nm possessed large absolute values of regression coefficients. After collecting in situ field spectra of saline soils in the Yellow River Delta of China, Fan et al. [
26] established a PLSR model between soil salinity and field spectra extracted from Advanced Land Imager (ALI), the effectiveness of PLSR model was then demonstrated in retrieving soil salinity from new-generation sensors, their results indicated that the prediction accuracy was quite acceptable with R
2, RPD of 0.749, 3.584, respectively. In addition, many other multivariate statistical techniques such as artificial neural networks (ANN), principal component analysis (PCA), and multiple adaptive regression spline (MARS) were used for quantitative determination of soil salinity and illustrating the relationships between spectral characteristics and soil properties [
27,
28,
29]. Although the spectral characteristics of soil samples can reflect the difference of soil salt content to a certain degree, only soil samples after grinding over 2 mm sieve were used in most previous spectral measurements, which usually cannot represent the true surface state of saline soils. For many saline soils with complex surface conditions or high roughness, spectral measurement of 2mm soil samples often cannot give a good estimation of salt content.
For saline soil samples, especially for those soda saline-alkali soils with relatively high clay content, it is very common to shrink and crack for the soil surface during water evaporation. Volume scattering was generated in the crack regions and result in a significant mixed pixel spectral effect. However, studies on the correlation between salt content and the overall spectral reflectance of saline soils with desiccation cracks are still limited and qualitative [
30,
31]. Texture features always contain important information about the structural arrangement of the soil surface with desiccation cracks, which can also accurately reflect the hue information and the overall arrangement of crack patterns [
32,
33,
34,
35]. However, spectral measurements considering the texture features of soil surface cracks were rarely focused on. To address this gap, the quantitative relationship between the reflectance spectra and soil properties of soda saline-alkali soils were tried to be analyzed based on consideration of the grey level co-occurrence matrix (GLCM) texture feature. Optimal predictive models for four main soil properties of interest including Na
+, EC, pH and total salinity were also developed based on the original reflectance and the mixed reflectance considering the contrast texture feature (CON) for the block soil samples (soil blocks separated by crack regions generated during the desiccation cracking process) and the comparison soil samples (soil powders with 2 mm particle size) respectively. This can be used to improve the prediction accuracy of soil salinity with less work intensity of soil spectral measurement in field conditions.
4. Discussion
Many studies have shown that the cracking process of clayey soil is mainly affected by the mechanical properties such as the soil matrix suction, the tensile strength, and the surface energy, these factors generally depend on the mineral type and clay content [
45,
46,
47]. Du et al. [
48] and Zhang and Peng [
49] have shown that soil clay is an important factor affecting soil shrinkage and cracking process due to its obvious plasticity, swelling, and cohesiveness. Gray and Allbrook [
50] found that soil shrinkage and cracking ability is significantly positively correlated with soil clay content. Greene-Kelly [
51] showed that soil shrinkage and crack ability is stronger when the expansive clay minerals (such as smectite) dominate in the soil. After measuring the relationship between internal friction angle and cohesion and smectite content, Chen et al. [
52] illustrated that when the smectite content is above 5%, the internal friction angle and soil cohesion of soil particles begin to decrease while shrinkage cracking enhances. Zhang et al. [
53] carried out X-ray diffraction complete analysis for soil minerals, their measurement results showed that the mineral composition of the saline soils in Songnen Plain is dominated by quartz and feldspar, which accounts for 87.6% to 91.6% of mineral contents, and that only a small amount of kaolinite and almost no highly active smectite is contained. Illite/smectite mixed layer formation is the secondary mineral with an interlaying ratio above 0.5 and the activity index only from 0.33 to 0.48, indicating that the soda saline-alkali soil in Songnen Plain has relatively high compressibility, poor water permeability, low shear strength and limited expansion [
54]. From
Table 1, it can be seen that although relatively high clay contents (26.28% to 29.37%) were measured in this study, the Std and CV are only 1.54% and 5.49%, this indicates that the clay contents of soil samples are almost the same in this study. On the other hand, the soil sample preparation process was kept the same, and the cracking test was carried out under the same controlled conditions (temperature of 25 °C, humidity of 35% and pressure of 101 kPa), indicating that laboratory conditions also had no effect on the cracking process in this study.
Table 4 shows that the crack degree has a good correlation with the soil properties related to the salt content, the correlation coefficients between the soil properties and the CON are all above 0.68. This is because during the dehydration process of the soil samples, a bounded water film is forming between soil particles (especially Na+ with the large hydrolysis radius) due to the interaction of soil particles with exchangeable cations. The bounded water film reduces the soil cohesion and the tensile strength [
55], which also has a great influence on the internal friction angle and the shear strength of the soil [
56,
57,
58]. In addition, the well-known diffuse double layer (DDL) thickness can also be considered as an important role in soil surface cracking, which always becomes thinner as the salt contents of soil samples increase and directly enhances the shrinkage and desiccation cracking of soils during water evaporation [
59,
60,
61].
It can be seen from
Figure 5 that surface states of soil samples cannot affect the morphology of the reflectance curves but can largely change the amplitudes of the curves. The reflectance curves of the comparison soil samples are higher than those of the block soil samples. This is because after the desiccation cracking process, the area scattering on the surface of block soil samples strengthens with the complexity of surface morphology characterized by the CON texture feature, and the relative surface roughness of block soil samples increases, which reduces the energy reflected to the spectrometer, and thus enhances the spectral difference between different block soil samples. In addition, due to the desiccation cracking test was carried out under natural conditions, all block soil sampleswere air-dried with some residual water remainingin the block soil samples and, therefore, reduced the overall reflectivity compared with the comparison soil samplestotally driedinan oven.Moreover, the shadows appeared within crack regions and made the energy reflect back to the spectrometer close to zero since the incident sunlight was not perpendicular to the surface of the soil samples.
A gradual decrease in reflectance with increasing salt contents was evident, as shown in
Figure 5, especially for the wavelength above 1400 nm since NaHCO
3and Na
2CO
3 dominants the salt minerals in Soda saline-alkali soils. Howari et al. [
62] and Farifteh et al. [
63,
64] showed that due to the unique resonance caused by stretching and bending, the sharp decline in reflectance of NaHCO
3 and Na
2CO
3as the wavelength increases can be considered diagnostic since no spectra of other salt minerals were investigated. Similar finding was reported by Wang et al. [
65] that the Na
2CO
3-affected soil has the lowest reflectance shape comparing with saline soils affected by Na
2SO
4, NaCl and the soil without any salt, their results also indicate that a clear decreasing pattern of reflectance with increasing content of Na
2CO
3 can be distinguished especially for the wavelength above 1400 nm, which is quite in agreement with the results in this study. Although the residual water contents are very limited (ranging from 0.18% to 2.34% with variance only 0.03%, indicated in
Table 5), hygroscopic water can still generate spectral response difference among the block soil samples to a certain extent. This is because from the results studied by Huang et al. [
66], salt content can increase the water-holding capacity of soil and thus further reduce the spectral reflectance of soil samples with higher salinities. In addition, according to Tedeschi and Dell’ Aquila [
67] and Huang et al., the soil aggregate ability decreases with increasing salinity as a result of the deflocculation of clay colloids induced by the salt, which maybe another reason causing the salinity inversely proportional to the spectral reflectance of the soil samples.
Figure 5 also shows that the soil samples under different surface conditions have the same absorption bands and similar corresponding absorption characteristics. Due to the influence of water content in the atmosphere, the reflectance curves exhibit obvious absorption feature at 1900 nm. The reflectance curves also exhibit a weak absorption characteristics of NaCl at 1190 nm, 1150 nm and three weak absorption valleys of NaHCO
3 at 1470 nm, 1990 nm and 2170 nm, these weak absorption features are caused by the resonance by the stretching and bending between the ions of the soil salt minerals. The water evaporation during the drying process caused the salt ions to precipitate on the surface of soil samples, making the spectral response to the main soil properties of the block soil samples quite high.
The regression results given in
Table 2 and
Table 3 indicate that when considering the CON texture feature of crack images, the modeling accuracies of main soil properties can be greatly improved compared with those models developed based on the original reflectance. This is due to the fact that the desiccation cracking process of the soil samples can make the salt separate out on the surface of soil samples, which makes the spectral response of the block soil samples to soil properties stronger than that of the comparison soil samples. Since the reflectance of the crack region is close to zero and the cracking degree is proportional to the salinity of soil samples, a more spectral difference of the block soil samples can be found, and the estimation accuracy of soil properties is, therefore, improved.
In order to quantitatively evaluate the effectiveness of the regression models considering the CON texture feature, the measured soil properties and the estimated soil properties of all the validation soil samples were linearly fitted to a function of y = x shown in
Figure 7. In addition, the root-mean-square error (RMSE), the relative root-mean-square error (RMSE%), the mean absolute error (MAE), the relative mean absolute error (MAE%) and the predicted ratio of deviation (PRD) were calculated as the evaluation indexes, and the results are shown in
Table 6. According to the criteria given by Farifteh et al. [
68], both univariate and multivariate regression models developed from the block soil samples provide good accuracy for EC, salinity and Na
+ (0.82 < R
2 < 0.9 and RPD>2), the calculated RMSE, RMSE%, MAE, and MAE% also show that the regression models of EC, salinity and Na
+ have quite high reliability and stability. In addition, the univariate and multivariate prediction models developed from the comparison soil samples also show an acceptable prediction accuracy for EC, salinity and Na
+ (0.5 < R
2 < 0.65 and RPD > 1.5), but high RMSE, RMSE%, MAE The MAE% indicate that the reliability and stability are poor. However, although CON texture features considered, both univariate and multivariate regression models are inaccurate for pH, whether based on the block soil samples or the comparison soil samples.
For the soda saline-alkali soil with surface cracks under natural conditions, the regression models developed in this study can be well implemented to estimate the main soil properties such as EC, salinity and the cation of Na
+ with relatively high accuracy and less work intensity. In detail, only soil samples need to be photographed and then collected firstly in the field. Secondly, the spectral measurements can be carried out based on the obtained soil samples (the soil samples collected may be soil blocks or soil powders, corresponding to the block soil samples and the comparison soil samples mentioned in this study) under laboratory-controlled conditions, which usually takes less time than the spectral measurement under natural world. Thirdly, the CON texture features of soil samples with surface cracks can be calculated individually from the crack images obtained in the field. After that, the developed regression models in this study can, therefore, be used for soil properties prediction using the mixed spectral reflectance of the soil samples computed from Equation (6) in
Section 3.5. However, the method and the regression linear models developed in this study are still limited for only the salt-affected soil with relatively high clay content, which generates surface cracks commonly under natural conditions. Moreover, other anticipated problems, such as the height of the photograph, the CON scale, the humidity effects, and the atmosphere attenuation, also need to be considered in the field measurement.