Soil salinization is increasing because of climate change and agricultural expansion, particularly in arid and semiarid regions. The global extent of primary salt-affected soils has been reported as , whereas secondary salinization affects , and nearly 20% of all irrigated land is salt affected.1 Soil salinization plays an important role in both soil degeneration and environmental deterioration, resulting primarily from global climate change and human activities.23.–4 Reducing soil degeneration and salinization may help to constrain the influence of climate change and human activities and provide an incentive for developing agriculture and improving the ecological environment in semiarid and arid areas.5 Because soil salinization usually causes a series of large-scale ecological and environmental problems, satellite-based earth observation has become an indispensable tool for estimating soil salinization. This method provides satellite-based remote sensing data generally at the pixel scale and spectral characteristics that delineate soil salinization. However, monitoring and mapping salt-affected soils are known to be difficult because dynamic processes are involved.67.–8 The greatest challenge is that soil salinization is a subsurface water–salt movement process. Therefore, to determine soil conditions, some approaches to soil salinity estimation make use of indirect indices, such as vegetation coverage, groundwater level, and groundwater salt content.9 In recent years, many studies have explored various approaches to soil salinity estimation by means of remote sensing techniques.6,9,10,11 Hyperspectral remote sensing data have been used to estimate soil salinity, with the main advantage of this method being its ability to resolve the reflectance responses between ground objects and image features with fine spectral detail, usually in hundreds of narrow, contiguous spectral wave bands.11,12 However, the soil reflectance responses of satellite-based remote sensing images are affected by soil conditions, such as water content and vegetation coverage.9,12 To avoid mixed pixels for cases in which soil conditions are not homogeneous, a spectral mixture analysis is necessary to resolve fine-scale land surface features.13 Quantitative soil salinity estimation using satellite-based remote sensing with higher accuracy remains a challenge. Existing research focuses mostly on providing a qualitative estimation of the surface soil salinity in which salinization has an influence on the surface soil at a depth of .14 This research reached different conclusions using sensitive bands of remote sensing images on soil salinity monitoring.14,15 As the soil salinization is related to the subsurface water–salt movement, the topsoil salinity monitoring cannot meet the practical needs of soil management. Recently, to gain subsoil information, ground remote sensing techniques, such as those that use ground-penetrating radar or the EM38-MK2 conductivity meter, provided soil information at deeper depths than those using satellite or aerial remote sensing data.16,17 These have been employed for soil condition delineation purposes,6,16,17 but these ground remote sensing techniques can provide subsoil information only in a small area. The study of the relationship between soil salt components and soil reflectance spectra as well as the possibility of using hyperspectral data for profile soil salinity estimation has been minimal. In this study, we investigated the possibility of using hyperspectral data for topsoil salinity estimation and for determining the relationship between soil salt components and soil reflectance spectra. The results showed that the bands sensitive to various levels of chemical components of soil salt were shown to differ, and they were controlled by the dominant component of the soil salt. It may be possible to use the hyperspectral data to estimate topsoil salt components with the aid of a significant correlation between topsoil spectra and soil salt ions.
Materials and Methods
The study area is a typical arid area of the Wigan–Kuqa River Delta Oasis in the northern region of the Tarim River Basin, Xinjiang Province, China, located at 41°25′ to 41°50′ N, 83°00′ to 83°45′ E (Fig. 1). The terrain is high to low from north to south with an average surface slope of 0.8%, the altitude is about 950 m, and the climate is warm temperate continental arid. The long-term average evaporation is 2124 mm, the average annual precipitation is 52 mm, and the ratio of evaporation to precipitation is . Large-scale rock halite and gypsum constitute the main deposits. The underground water level and salt content are high, and the chloride and sulfate contents in the soil are higher. The soil salinization depth is about 1.0 m, though soil salinity decreases gradually with increasing depth. As soils become more saline, vegetation becomes less and less likely to survive in the soil. Bushes grow in mild saline soils, while bare ground appears in the areas with high soil salinity levels. In addition, the soil is composed of a large amount of fine particles with poor permeability, providing good conditions for soil salinization. The groundwater level is , and the groundwater salt content is larger. The soil textures are mainly sandy loam and clay. The soil saline pH value is , chlorine is prevalent, and NaCl is the dominant soil salt component. Interannually, soil salt moves dramatically at the vertical level, whereas it moves less dramatically at the horizontal level.18 The groundwater level with a depth of 1.0 m is shallow, and the crop growth is dependent on irrigation.
The HJ-1A hyperspectral remote sensing satellite was launched by China in 2008. We used one scene of the HJ-1A hyperspectral image (HSI) covering the study area on October 31, 2010, which was cloud free with a resolution of 100 m. It contained 115 spectral bands, and the spectral range was 459 to 956 nm, from blue to near infrared (NIR). The average spectral resolution is 4.32 nm. The signal-to-noise ratio is 50 to 100. The relative radiometric calibration accuracy is , and the absolute radiometric calibration accuracy is . The HJ-1A HSIs are reliable, which has been proven in many ecological and environmental studies.1920.–21 The images were products with radiation calibration and preliminary georeferencing and were downloaded from the website Ref. 22. They were further processed (including georeferencing and atmospheric correction) using ENVI4.8 software. The georeferencing used more than 20 ground control points, which were selected from a -scale topographic map. The root mean-squared error (RMSE) of the image rectification was 0.2 pixels. A quick atmospheric correction (QUAC) method was used to implement the atmospheric correction. QUAC is an atmospheric correction method for multispectral and hyperspectral imagery that works with the visible and near-infrared through shortwave infrared wavelength range, and it determines the atmospheric compensation parameters directly from the information contained within the scene using the observed pixel spectra.
Bands 1 to 21 and 113 to 115 had some noise, so bands 22 to 112 in the wavelength coverage of 508 to 926 nm were selected to analyze the relationship between total soil salt content (SSC), which distinguishes the soil salinization condition, and the reflectance spectra. Bands 1 to 21 and 113 to 115 were removed.
In addition, the SSC and chemical ion content, including , , , , , , and , of the soil samples at depths of 0 to 10 cm were used in this study. They were measured using the methods introduced in Sec. 2.3.
To reduce the effect of soil condition on soil spectra, we conducted a field survey of the area. Soil sample locations with similar soil conditions were selected, and the locations of the 25 soil sample points were determined using a handheld global positioning system. We collected the soil samples at a time that was nearly consistent with the time of the satellite transit in the study area. Soil samples for the 25 points were collected between October 12 and 19, 2010. In accord with the characteristics of the salinized soil,18 soil samples were collected from depths of 0 to 10 cm. The soil features are very similar for the sampling points at a grid size of . We collected samples from five points that show an -type distribution in a grid of and used the average SSC value of the five samples to represent each soil layer sampling point SSC value. A total of 125 soil samples were collected.
The soil moisture content was measured by weighing the samples before and after the soil samples were air-dried. The soil samples were sieved at 0.5 mm to remove large debris and vegetation. The SSC and soil pH were measured following the National Standard of Soil Physiochemical Analysis.23 The stoving residue method was used to measure the SSC of the soil samples. According to the saline soil classification standard of Xinjiang Province, China, these soil samples include nonsalinized soil; slightly, moderately, and severely salinized soil; and solonchak. To study the effect of the soil salt chemical components, including , , , , , , and , on the soil spectral reflectance, the chemical ion content was measured. The contents were measured by different methods (Table 1).24
Evaluation methods of soil sample chemical ions.
|Chemical ions||HCO3−||Cl−||SO42−||Ca2+ and Mg2+||Na+ and K+|
|Methods||Double indicator titration||Silver nitrate titration||Ethylene diamine tetraacetic acid (EDTA) indirect titration method||EDTA complex-metric titration||Flame photometry|
To determine the spectral features of soil salinity, soil spectra were measured in the laboratory. The soil spectra measurement was conducted using an analytical spectral device (ASD) FieldSpec3 Portable Spectroradiometer (American Analytical Spectral Devices Inc.) at wavelengths of 350 to 2500 nm with sampling intervals of 1.4 nm from 350 to 1050 nm and 2 nm from 1000 to 2500 nm. The spectral resolution of the measurements can reach 1 nm after resampling. Before each measurement, the reference panel was initialized for the ASD. The detector was operated at a height of 10 cm directly above the soil sample center. The spectral reflectance of each soil sample was measured 14 times. Flat plastic dishes with diameters of 20 cm were filled with soil samples, and each soil sample surface was flat. The soil samples were illuminated with two tungsten quartz halogen filament lamps. One lamp was placed on each side of the sample with the light beams angled at 45 deg with respect to the vertical direction. The distances between the two lamps and from the lamp to the dish edge were 20 cm. The reflected light was collected with a 25 deg field of view. Since the shadow of the detector was out of the field of view, it could not be detected.
The spectral reflectance curves of the soil samples measured in the laboratory were constructed to analyze the basic spectral features of the saline soil. Soil reflectance spectra curves were processed by splice correction and the Savitzky–Golay method to smooth the spectra curve features and to remove subtle noise.25 After processing, the average of the 14 reflectance measurements was calculated from the soil sample spectra to represent the salinity of each soil sample.
The selection of sensitive bands and spectral indices is a key to the soil salinity estimation using data from remote sensing. We used a correlation analysis and partial least squares regression (PLSR) model to select sensitive bands and spectral indices and to construct predictive models for soil salinity. Therefore, the correlation coefficient (), the coefficient of determination (), and the RMSE were used to assess the feasibility of these methods. SPSS Statistics 17.0, IBM MATLAB7.0, and R language were used to conduct these calculations. The soil sample spectral reflectance was extracted from the remote sensing images, and the correlation coefficient between the soil reflectance spectra and the SSC was then calculated as
We analyzed the soil salt component using the -type clustering analysis method. Clustering analysis is a statistic method for grouping objects of random kinds into respective categories. It was used to find the most appropriate kinds through mathematical statistics and some information available to be collected when no priori hypotheses were used. It is a multivariate statistics method for studying classification. Clustering analysis can be classified into two major types according to research purpose, -type clustering and -type clustering. The -type one was also called index clustering, which was used to sort various kinds of indices. The most important thing of -type clustering is to define similarity, that is, how to quantify similarity. The first step of clustering is to define the measurable similarity of the two indices, that is, similarity coefficient. We selected the absolute value of the Pearson correlation coefficient as the similarity coefficient of -type clustering analysis. The two variables tend to be more similar when the absolute value of the Pearson correlation coefficient is larger.
Furthermore, soil moisture and vegetation, which are related to soil salinization, can affect soil reflectance. To remove the effects of vegetation and soil moisture on the soil spectra, a spectral vegetation index (VI) was introduced to estimate soil salinity.9 Several types of VI are adapted to various conditions, but the difference vegetation index (DVI) is better suited to low vegetation density areas, where vegetation shows a moderate- or low-level coverage. This is applied to salinization areas with lower vegetation coverage; therefore, we selected DVI as the spectral index for estimating soil salinity.
To estimate the salinity using the HJ-1A HSI, spectral transforms of the first-order differential of reflectance, the first-order differential of the reciprocal of the logarithm of reflectance, and the reciprocal of the logarithm of reflectance were carried out. The first-order differential of the transformed spectra was calculated26,27
The soil chemical ion content at depths of 0 to 10 cm was measured in October 2010. The results indicated that the main chemical component of the soil was NaCl. was the most prevalent ion, followed by (Table 2). To study the relationship between soil spectra and soil salt components, the soil samples were divided into two groups according to the measured results. There were 21 soil sample points in the first group (Fig. 1), and there were four soil sample points that were on the margins of the study area in the second group. The first group contained more (Table 2), and the ratio of to was 0.044. The second group contained more and (Table 2). The ratio of to was 0.460, and the ratio of to was 0.085. The soil salt component in the first group was purer, with NaCl as the dominant chemical component in the group. The chemical components in the soil salt were more complex in the second group. In addition, the soil sample points in the first group were on the inside of the study area, and the soil sample points in the second group were on the margins of the study area. This indicates that the soil salt components varied from the inside to the outside of the study area.
The chemical ions content in soil layer in October 2010.
|Chemical ions (g/Kg)||HCO3−||Cl−||Ca2+||Mg2+||SO42−||Na+||K+|
|All soil samples||Mean||0.177||12.276||0.405||0.821||0.578||12.448||0.001|
|The coefficient of variation||1.508||1.512||1.457||2.865||0.394||1.131||2.000|
|The first group||Mean||0.064||12.932||0.421||0.182||0.579||13.09||0.0004|
|The coefficient of variation||0.469||1.153||1.504||0.670||0.408||1.134||1.000|
|The second group||Mean||0.770||8.828||0.321||4.178||0.574||9.078||0.004|
|The coefficient of variation||0.083||1.122||0.984||1.226||0.375||1.098||0.750|
Statistics of the salt content in the study area in October 2010.
|Soil layer salt content (g/kg)||Minimal value||Maximal value||Mean||Standard deviation||The coefficient of variation|
|The first group||0.500||131.500||37.618||36.753||0.977|
|The second group||8.900||62.000||28.750||25.195||0.876|
The spatial variation in soil salinity within the study area was analyzed using statistical methods with the field data. For the SSC in October 2010, the results of the descriptive statistical analyses (Table 3) showed that the salt content of all the soil samples from the 0- to 10-cm soil layer ranged from 0.500‰ to 131.500‰ with a mean of 36.199‰. The higher mean SSC value of the 0- to 10-cm soil layer indicates that the soils in the study area were highly affected by salt and that the salt was concentrated in the topsoil.
In terms of the spatial distribution of soil samples, more than 90% of the samples were affected by salt. The soil samples were distributed in croplands, grassland, and sandy land or at the edges of water bodies. Highly saline soils were primarily distributed in soil and weeds affected by salt; therefore, the collected soil samples represent salinity variations in the study area. The pH value of each soil sample was , which indicated an alkaline soil.
Soil Spectral Features
Satellite-based remote sensing spectra are affected by many factors resulting in mixed spectra. Therefore, it is necessary to study the saline soil spectra features to extract the saline soil information using remote sensing images.
The laboratory-measured spectral reflectance curves shown in Fig. 2 were similar to those in the study by Ayetiguli et al.,14 with a slight difference. Compared with that study,14 the absorption bands that occur in the shortwave-infrared (SWIR) bands were weaker, and the spectral reflectance around the wavelength 2150 nm in SWIR bands was larger than in the visible (VIS) and NIR bands. This is especially true for the spectral reflectance around the wavelength 500 nm in the VIS band, which had a larger difference. The distribution of the spectral reflectance values was also more scattered. Therefore, we found that differences in soil salinity could also be distinguished using the SWIR bands of the remote sensing images in the study area, except for those of the VIS and NIR. Based on the analysis above, blue, green, red, NIR, and SWIR bands were tested to estimate the SSC with the measured spectra from the laboratory as discussed in Sec. 3.3.1. In addition, the laboratory-measured spectral reflectance curves of the soil samples in the first group had similar features to those of the soil samples in the second group (Fig. 3).
The soil spectral reflectance in HJ-1A images (Fig. 4) had similar features to those of the aforementioned laboratory measured spectral data. The soil spectral reflectance increased as the wavelength increased in the wavelength region of 540 to 926 nm. The NIR had the largest spectral reflectance. In addition, we determined that not all the soil reflectance values increased as the SSC increased. Some soil sample points with high salt content had lower reflectance values and vice versa. Therefore, the reflectance spectra of saline soil are also affected by other soil factors, such as soil type, organic matter content, soil moisture, and vegetation. To estimate soil salinity using remote sensing technology, it is necessary to remove these effects.28
Selection of soil salinity sensitive spectral indices with measured spectra in the laboratory
To verify the reliability of the HJ-1A data, measured hyperspectral data using the ASD instrument in the laboratory were explored first. Since the average spectral resolution of ASD is similar to the HJ-1A images, the relationships between the SSC in the 0- to 10-cm soil layer and reflectance spectra29 were analyzed. First, seven transforms of the reflectance spectra were carried out, including the first-order differential of reflectance, the first-order differential of the root mean square of reflectance, the first-order differential of the logarithm of reflectance, the first-order differential of the reciprocal of the logarithm of reflectance, the first-order differential of the reciprocal of reflectance, the reciprocal of the logarithm of reflectance, and the reciprocal of reflectance. Second, we analyzed the correlations between the SSC and the soil reflectance (Fig. 5). Finally, we analyzed the correlations between the SSC and the seven transformed spectra (Figs. 6Fig. 7–8). When we directly analyzed correlations between the reflectance spectra and the SSC, the results showed that the reflectance spectra (from VIS to SWIR) had gradually reduced correlations with SSC. The greatest correlation coefficient (0.62) was found between the blue band (397 nm) of the VIS and SSC, whereas the lowest value (0.36) occurred between the SWIR band (1650 to 1700 nm) and SSC. For the NIR bands, the best correlation coefficient was 0.382 that occurred at wavelength 762 nm. However, the correlation coefficient between the SSC and the transformed reflectance spectra increased to some extent in some bands, such as the blue band (355 to 397 nm), NIR band (758 to 770 nm), and SWIR band (1400 to 1450 nm and 2200 to 2350 nm). Through the spectral transforms, there is a remarkably improved correlation coefficient (0.3) between two bands (NIR and SWIR) and the SSC. Meanwhile, the calculated results showed that the function of the first-order differential of reflectance and the first-order differential of the reciprocal of the logarithm of reflectance was almost the same, and they achieved better correlations with the SSC. In the VIS bands, the best was 0.73 at a wavelength of 390 nm. In the NIR bands, the best was that occurred at a wavelength of 758 nm.
Selection of soil salinity sensitive bands and spectral indices using the HJ-1A images
We selected sensitive bands with measured hyperspectral data using the ASD instrument in the laboratory as described in Sec. 3.3.1. To estimate soil salinity using satellite remote sensing data, the sensitive band selection and relationship analysis between the spectral reflectance of the remote sensing images and the SSC are very important. The correlation coefficient between the spectral reflectance and SSC was calculated using the method discussed in Sec. 2.3.
For all the soil samples, the spectral reflectance of each band in the HJ-1A image had a positive and strong significant correlation with the salt content at 0 to 10 cm of depth. The largest correlation coefficient of 0.800 appeared at the wavelength 510.975 nm in the VIS band (blue band) (Fig. 9). For the soil samples in the first group, the spectral reflectance of each band in the HJ-1A image also had a positive and stronger significant correlation with the salt content at 0 to 10 cm. The largest correlation coefficient of 0.857 also appeared at the wavelength 510.975 nm of the VIS band (blue band) (Fig. 10). Therefore, the strongest relationship between soil reflectance derived from the HJ-1A HSIs and SSC occurred in the blue bands of VIS, which was consistent with the relationship between soil salinity and reflectance in the laboratory. The relationship between soil salinity and the reflectance of the HJ-1A images was reliable.
In addition, the relationship between the spectral reflectance at the wavelength 510.975 nm and soil moisture of all soil samples at 0 to 10 cm was analyzed. The correlation coefficient was , but the significance level ( value) was larger than 0.05, so soil moisture in the 0- to 10-cm soil layer did not have a significant effect on soil spectra. Soil spectra were affected mostly by the SSC between the soil salt and the soil moisture.
Based on the analysis in Sec. 3.3.1, three transforms of the reflectance spectra were carried out, including the first-order differential of reflectance, the first-order differential of the reciprocal of the logarithm of reflectance, and the reciprocal of the logarithm of reflectance. The correlation between the SSC and the three transformed reflectance spectra was analyzed.
For all soil samples, the correlation between the SSC of the 0- to 10-cm soil layer and the transformed spectra using the reciprocal of the logarithm of reflectance was the largest, even larger than that between the SSC and the untransformed reflectance spectra (Table 4). For the soil samples in the first group, the correlation result between the SSC of 0- to 10-cm soil layer and the reflectance spectra was the largest, larger than that between the SSC and the transformed spectra using the first-order differential of the reciprocal of the logarithm of reflectance (Table 5). The correlation result between the SSC and the transformed spectra by the method of the first-order differential of the reciprocal of the logarithm of reflectance is shown in Fig. 11. Additionally, we found that SSC had stronger correlations with the reflectance spectra derived from the HJ-1A (Table 4) compared to those correlations with reflectance spectra measured by the ASD in the laboratory. The differences were caused by the differences in soil salt distributions. The salt is distributed in the upper layer of the topsoil in the field, but it is distributed homogeneously in the soil samples in the laboratory.
Largest correlation coefficient between SSC of all soil samples and soil reflectance and transformed spectra.
|Soil spectral reflectance and transformed spectra||Reflectance||The reciprocal of the logarithm of reflectance|
0.01 confidence level.
Largest correlation coefficient between SSC of the first group and soil reflectance and transformed spectra.
|Soil spectral reflectance and transformed spectra||Reflectance||The first-order differential of the reciprocal of the logarithm of reflectance|
0.01 confidence level.
We distinguished the soil salt component using the -type clustering method. For soil samples in the first group, according to the similarity coefficient of soil salt chemical ions obtained through clustering analysis (Table 6, Fig. 12), and had the largest similarity, which can be classified into the same category, and and had a larger similarity and can be classified into the same category. For all the soil samples, according to the similarity coefficient of soil salt ions obtained through clustering analysis (Table 7, Fig. 13), and had the largest similarity and can be classified into the same cluster, and and had a weaker similarity and can be classified into the same cluster to some extent. Based on the above analysis, we knew that NaCl and can be distinguished in the first group, and for all the soil samples, NaCl can be distinguished; can be considered as one category to some extent.
Clustering analysis results of the soil chemical ions contents of the first group.
|Clustering||Clusting||Similarity coefficient (%)|
Clustering analysis result of the soil chemical ions contents of all soil samples.
|Clustering||Clusting||Similarity coefficient (%)|
We then examined the relationship between the spectral reflectance and the soil salt chemical ion content of the 0- to 10-cm soil layer. For all soil samples (Fig. 14 and Table 8), had the largest correlation coefficient, followed by , , , , , and . The wavelengths with the largest correlation coefficients occurring for , , , , and were 546.745, 510.975, 605.435, 605.435, and 584.520 nm, respectively. The wavelength (510.975 nm) with the largest correlation coefficient occurred for both and ; therefore, the soil salt chemical component should be NaCl. In addition, , , and had negative correlation coefficients. For the first group (Fig. 15 and Table 9), had the largest correlation coefficient, followed by , , , , , and . The wavelengths with the largest correlation coefficients occurring for , , , , and were 543.815, 543.815, 584.520, 609.065, and 540.920 nm, respectively. The wavelength (510.975 nm) with the largest correlation coefficients occurred in both and ; therefore, the soil salt chemical component should be NaCl. The wavelength (543.815 nm) with the largest correlation coefficients occurred in both and ; therefore, the soil salt chemical component should be . In addition, and had negative correlation coefficients. The main chemical ion of the soil had the strongest relationship with soil reflectance, and this result occurred at the wavelength of 510.975 nm, which is same as the salt content in the 0- to 10-cm soil layer. Moreover, the correlation coefficient between and contents and the spectral reflectance was larger than that of the SSC correlation coefficient. Compared with the result of clustering analysis of soil salt components, as for the first group, this result is consistent with the result of clustering analysis; for all samples, this result is consistent with the result of clustering analysis to a large extent. In addition, the soil salt component of the soil samples in the first group was purer, and more soil salt components can be distinguished using hyperspectral data, so soil salt components can be distinguished using hyperspectral data.
Highest correlation coefficient between spectral reflectance in various bands of the HJ-1A in October 2010 and the soil chemical ion content of all soil samples.
|Soil salt ions||HCO3−||Cl−||Ca2+||Mg2+||SO42−||Na+||K+|
0.01 significance level.
Highest correlation coefficient between spectral reflectance for various bands of the HJ-1A in October 2010 and the soil chemical ions content of the first group.
|Soil salt ions||HCO3−||Cl−||Ca2+||Mg2+||SO42−||Na+||K+|
0.01 significance level.
Vegetation cover proved useful for soil salinity estimation,14 for the HJ-1A HIS at the wavelength 834.265 nm of the NIR band, the spectral reflectance had the largest correlation coefficient with SSC, whereas that at 711.495 nm in the red band was the lowest. Therefore, we constructed a DVI using the HJ-1A HIS to estimate soil salinity
Based on the above analysis, the soil spectral reflectance at wavelength 510.975 nm had the strongest relationship with the salt content of the 0- to 10-cm soil layer. The DVI was constructed using the HJ-1A HSI; therefore, the spectral reflectance at wavelength 510.975 nm () and the DVI were selected to estimate the salinity of the 0- to 10-cm soil layer.
Soil salinity of the 0- to 10-cm top soil layer has been estimated using satellite-based remote sensing images.14 This estimation is generally based on the soil spectral features, and the selection of the spectral indices is usually determined with regional differences to some extent. According to the analyses in Secs. 3.2 and 3.3.1, the VIS and infrared (IR) bands can be used to estimate soil salinity, and DVI can also be used for this purpose. For the remote sensing images with different spectral and spatial characteristics, it is necessary to choose the proper methods and spectral indices to estimate soil salinity. In terms of the features of the HJ-1A HSIs, we built models using statistical methods and various spectral indices to estimate soil salinity at the 0 to 10 cm depth.
The soil salinity has been approximately described by a linear function of soil reflectance.28 PLSR models that reduce the dimension and multicollinearity of the data used widely in many fields,14,28,30,31 so we applied the PLSR model to estimate soil salinity at the depth of 0 to 10 cm in this study. The spectral indices were selected in Sec. 3.3.2 as independent variables in the model. The salt content of the 0 to 10 cm () was the dependent variable. We assessed the constructed models using , RMSE, and the value. RMSE was frequently used to measure the difference between values predicted by the model and the values actually observed from the ground sample points. A smaller RMSE value means that the model prediction has a better accuracy. The predictive accuracies that were illustrated by , RMSE, and the value were tested using a leave-one-out cross-validation method.
To explore the effects of the soil salt component levels on the predictive accuracies, all soil samples and the soil samples of the first group were used to build predictive models.
We first built the predictive models using the soil samples of the first group. First, the predictive models were established using the PLSR method. The was estimated by the spectral index and the DVI in model 4.
The accuracy of the fit of these models is listed in Table 10. Of the salinity estimation models for the 0- to 10-cm soil layer, model 4 had a better fitting accuracy . The predictive accuracies that were illustrated by , RMSE, and the value were tested using a leave-one-out cross-validation method (Table 10, Figs. 16 and 17). The predictive accuracy for was better using model 4 (Table 10, Fig. 16). The predictive accuracy for was poorest using model 5 (Table 10, Fig. 17).
Compared with the soil salinity estimation models that were constructed by spectral indices , the predictive accuracy of the model 4 was higher, the value was higher, and the RMSE values were reduced (Table 10), so the spectral index DVI was useful in increasing the predictive accuracy for soil salinity in the study area. According to the relationship between soil spectral reflectance and the chemical ion content, as shown in Table 10, the content of the 0- to 10-cm soil layer had the largest correlation coefficient with spectral reflectance. Therefore, we used the spectral reflectance at a wavelength of 510.975 nm () to estimate the content () in the 0- to 10-cm soil layer using the PLSR model in model 6.
Constructed models 4–6 using spectral indices based on HJ-1A images.
|Spectral indices||Models||R2||P value||RMSE|
As indicated by the results shown using model 6 in Table 10, the prediction accuracy of the model that was constructed using the same spectral index was significantly greater than that for the salt content of the 0- to 10-cm soil layer . The predictive accuracies that were illustrated by , RMSE, and the value were tested using a leave-one-out cross-validation method (Table 10, Fig. 18). Therefore, estimating the dominant salt of the soil using reflectance spectra will lead to greater prediction accuracy (Fig. 18).
We then built the predictive models using all the soil samples and the same methods and spectral indices that were mentioned above. The was estimated by the spectral index and the DVI in model 7.
As a comparison, the soil salinity at a depth of 0 to 10 cm was estimated by the spectral index in model 8.
The accuracy of fit of these models is listed in Table 11. Of the salinity estimation models for the 0- to 10-cm soil layer, model 7 had a better fitting accuracy . The predictive accuracies that were illustrated by , RMSE, and the value were tested using a leave-one-out cross-validation method (Table 11, Figs. 19 and 20). The predictive accuracy for was better using model 7 (Table 11, Fig. 19). The predictive accuracy for was poorest using model 8 (Table 11, Fig. 20). This indicates that the spectral index DVI was also useful in increasing the predictive accuracy of soil salinity using all the soil samples in the study area.
Constructed models 7–9 using spectral indices based on HJ-1A images.
|Spectral indices||Models||R2||P value||RMSE|
In addition, we used the spectral reflectance at wavelength 510.975 nm () to estimate the content in the 0- to 10-cm soil layer using the PLSR model and all the soil samples in model 9.
As indicated by the results shown for model 9 in Table 11, the prediction accuracy of the model was significantly greater than that for the salt content of the 0- to 10-cm soil layer . The predictive accuracies that were illustrated by , RMSE, and the value were tested using a leave-one-out cross-validation method (Table 11, Fig. 21). Therefore, this showed that estimating the dominant salt of the soil using reflectance spectra would lead to greater prediction accuracy.
Comparing models 4 and 7, model 4 had higher predictive accuracy. As the soil salt components varied from the inside to the outside of the study area, model 4 estimated soil salinity inside the study area, whereas model 7 estimated soil salinity for the total study area. The estimated SSC at 0 to 10 cm depth () in the study area using model 7 is shown in Fig. 22.
We estimated effectively the topsoil salinity using HSIs. We found NaCl was the dominant chemical component of the soil salt in the study area. Due to the limitation of wavelength of the HJ-1A images, we explored only the effects of SSC and soil salt components in soil reflectance spectra, and we predicted soil salinity using the data band 508 to 926 nm. In the future, it is urgently needed to study the effects of regional differences and the potential of other bands for estimating soil salinity. These needs are discussed below in detail.
First, the bands from blue to NIR perform equally in predicting soil salinity. Based on the data analyses, there is no significant difference in the correlation coefficients between SSC at 0 to 10 cm and the spectral reflectance from the blue band to NIR band (508 to 926 nm) (Figs. 9 and 10). We also obtained similar results in the laboratory (Fig. 5). From the SWIR bands to blue bands, the correlation coefficients increased gradually from 0.36 to 0.62, though there is a limited increase in correlation coefficients (Fig. 5). Additionally, the spectral reflectance of each band is positively correlated with the SSC, i.e., the brighter the soil, the higher the degree of salinization. Using spectral reflectance, VIS bands are superior to the IR bands. For the SWIR to VIS bands, spectra transformation, such as using the first-order differential of the reflectance spectra, increased the correlation in some bands (Fig. 6). This could be attributed to the reflectance spectra difference caused by the soil salinity that was detected by the narrow and contiguous spectral wave bands of the HSIs. The correlation between the single band reflectance was reduced,26,27 so the proper spectra transformation enhanced prediction accuracy. Therefore, we concluded the spectra transformation could improve the correlation between the soil salinity and the reflectance spectra. This further proved that the VIS, NIR, and SWIR bands could be used to estimate soil salinity in the study area.
Second, the past finding showed that vegetation cover, soil moisture, soil components, and other factors can affect spectral reflectance values from optical remote sensing images.28 These factors increased the uncertainty of the prediction results. However, it was very difficult to remove these effects on the soil reflectance spectra directly. These factors had direct or indirect relationships with soil salinization, and they can be considered as indices for soil salinity estimation. This would include vegetation cover for the soil salinity estimation in this study and one by Ayetiguli et al.14 Due to regional differences in soil salt elements, vegetation cover, and soil moisture, the power of these factors and sensitive bands on the soil salinity estimation may change accordingly. Abb pointed out that vegetation indices were poor predictors of soil salinity.32 These different conclusions are partially explained by the difference in vegetation types and the degree of soil salinization. Therefore, it is worth our effort to study the effect of these factors on soil reflectance spectra and the degree of soil salinization.
Third, we proposed that the spectral reflectance was useful in estimating the dominant salt component. This is because the saline soil reflectance spectra are controlled by the dominant salt component. Based on the analyses in Sec. 3.3.2, the sensitivities of different bands of remote sensing images were highly impacted by the dominant component of the soil salt (Figs. 14 and 15). Moreover, the composition of soil salt is extremely complex. The correlation was stronger between the dominant component of salt and the reflectance spectra than that between the SSC and the reflectance spectra. In particular, we need to mention the limitation of the wavelength coverage of the HJ-1A images (508 to 926 nm), which is referred to in Sec. 3.3.2. We observed a weaker correlation between the transformed spectra and the 0- to 10-cm soil layer salt content (Table 5). For the measured reflectance spectra in the laboratory and using the first-order differential transformation of the reflectance spectra, the largest correlation coefficient value of 0.735 occurred at the wavelength of 390 nm, which was outside the wavelength coverage of the HJ-1A images.
Fourth, we examined the relationship between the spectral reflectance and the soil salt chemical ion content of the 0- to 10-cm soil layer in Sec. 3.3.2, which showed that the wavelength with the largest correlation coefficient occurring between each soil chemical ion and soil reflectance was different, and the sensitive bands for the various soil salt chemical components differed. The sensitive bands for the soil salinity estimation are controlled by dominant salt chemical components, varying with the various soil salt chemical components or chemical component levels in the soil. We were able to determine easily the most sensitive band for higher accuracy soil salt salinity estimation based on the relationship between the main soil salt chemical ion and the soil reflectance. In addition, there exists the possibility of using the hyperspectral data to distinguish the topsoil salt components.
Finally, we discussed how different bands contributed to the soil salinity estimation. Previously, Ayetiguli et al.14 pointed out that the NIR bands had the highest contribution to soil salinity estimation, followed by the blue band, red band, and green band, using Quickbird images and the PLSR method. Said et al.15 concluded that the NIR band and SWIR-1 band provided the highest contribution to the estimation of soil salinity, followed by the red, SWIR-2, green band, and the blue band, which made the lowest contribution, using Landsat 7 ETM+ images and the PLSR method. In this study, we found that the blue band had the highest correlation with soil salinity in the study area. The correlation between SSC and reflectance spectra reduced gradually from the VIS to SWIR bands. We found that the soil spectra reflectance was significantly impacted by the SSC, and the sensitive bands will change with the difference of soil salt component. Since many factors, such as soil salt chemical components and their content, regional differences, soil moisture, the band range, soil parent material, and vegetation cover, act on the results, we cannot draw the conclusion that the difference in the sensitive bands from the remote sensing images on SSC was caused by the difference in the soil salt component. We can only claim that the sensitive bands from the remote sensing images on SSC were different based on the difference in the soil salt component in our analysis. More research must be conducted related to these areas in the future.
It is challenging to conduct timely observations regarding the water–salt movement of the subsurface using remote sensing data. In this study, we have explored the distinguish of the soil salt components using HSIs and estimated the soil salinity by assigning indices derived from the HJ-1A images into established models. Our results indicated that the spectra reflectance measured in the laboratory showed good similarity to the reflectance spectra derived from the HJ-1A HSIs. The reflectance spectra in the blue bands had the largest correlation coefficient with the salinity in the 0- to 10-cm soil layer in the study area. The correlation between SSC and reflectance spectra reduced gradually from the VIS to SWIR bands. Results from models demonstrated improved prediction accuracies by proper transformation of the reflectance [blue (355 to 397 nm), infrared (758 to 770 nm), and SWIR (1400 to 1450 nm and 2200 to 2350 nm)]. The vegetation cover index improves our model performance in the study area. The PLSR models estimated the soil salinity in this study well. We briefly summarize the performance of our approach in the following. The soil salt components had a variation from the inside to the outside in the study area. The salinity of the 0- to 10-cm depth soil layer can be estimated with these statistical results for parts of the study area (, , and ) and for the total study area (, , and ).
In particular, we analyzed the soil salt components using the clustering analysis method, compared with the clustering analysis result, and we distinguished the soil salt components using hyperspectral data. It was proved that hyperspectral data can be used to distinguish soil salt components. We also found that remote sensing data are very sensitive to different soil salt chemical components or chemical component levels in the soil.
The choice of the image band was greatly influenced by the dominant component of the soil salt. Using information for the dominant salt will lead to greater prediction accuracy. Our methods achieve a relatively higher accuracy in predicting the dominant salt element () compared with the soil salinity estimation accuracy for the 0- to 10-cm depth soil layers using the same method and spectral indices. This provided a new perspective on soil salt estimation using remote sensing, which is very useful in the field of soil management and studies of the soil salinization mechanism.
However, the above conclusions were obtained based on data collected in an arid area in China. Caution must be exercised when interpreting these results for other areas as regional differences are important factors that should be included.
This research was supported by the National Natural Science Foundation of China, Nos. 41561081 and 41001198, the State Key Program of National Natural Science of China, No. 41331175, the Suzhou Science and Technology Program of Applied Basic Research, No. SYG201319, the Hubei Provincial Natural Science Foundation of China, No. 2014CFB725/ZRY2014000982, and by the key Laboratory of Oasis Ecology of Xinjiang University, China, No. XJDX0201-2010-13.
H. Y. Li et al., “Mapping the three-dimensional variation of soil salinity in a rice-paddy soil,” Geoderma 195–196, 31–41 (2013).GEDMAB0016-7061http://dx.doi.org/10.1016/j.geoderma.2012.11.005Google Scholar
T. G. Ammarl et al., “Soil salinity changes in the Jordan Valley potentially threaten sustainable irrigated agriculture,” Pedosphere 23, 376–384 (2013).PDOSEAhttp://dx.doi.org/10.1016/S1002-0160(13)60029-6Google Scholar
K. M. Anil, “Climate change impact, mitigation and adaptation strategies for agricultural and water resources, in Ganga Plain (India),” Mitigation Adapt. Strategies Global Change 18, 673–689 (2013).http://dx.doi.org/10.1007/s11027-012-9381-7Google Scholar
A. Yakirevich et al., “Modeling the impact of solute recycling on groundwater salinization under irrigated lands: a study of the Alto Piura aquifer, Peru,” J. Hydrol. 482, 25–39 (2013).JHYDA70022-1694http://dx.doi.org/10.1016/j.jhydrol.2012.12.029Google Scholar
Y. Guo et al., “Integrating remote sensing and proximal sensors for the detection of soil moisture and salinity variability in coastal areas,” J. Integr. Agric. 12, 723–731 (2013).http://dx.doi.org/10.1016/S2095-3119(13)60290-7Google Scholar
J. S. M. Gabriel, “Spatial variation of soil salinity in the mexicali valley, Mexico: application of a practical method for agricultural monitoring,” Environ. Manage. 50, 478–489 (2012).EMNGDC1432-1009http://dx.doi.org/10.1007/s00267-012-9889-3Google Scholar
R. J. Yao et al., “Accuracy and uncertainty assessment on geostatistical simulation of soil Salinity in a coastal farmland using auxiliary variable,” Environ. Monit. Assess. 185, 5151–5164 (2013).http://dx.doi.org/10.1007/s10661-012-2932-2Google Scholar
N. M. Khan et al., “Assessment of hydrosaline land degradation by using a simple approach of remote sensing indicators,” Agric. Water Manage. 77, 96–109 (2005).AWMADF0378-3774http://dx.doi.org/10.1016/j.agwat.2004.09.038Google Scholar
N. M. Khan and Y. Sato, “Monitoring hydro-salinity status and its impact in irrigated semi-arid areas using IRS-1B LISS-II data,” Asian J. Geoinform 3, 63–73 (2001).Google Scholar
Z. E. Mashimbye et al., “Model-based integrated methods for quantitative estimation of soil salinity from hyperspectral remote sensing data: a case study of selected South African soils,” Pedosphere 22, 640–649 (2012).PDOSEAhttp://dx.doi.org/10.1016/S1002-0160(12)60049-6Google Scholar
Y. L. Weng et al., “A spectral index for estimating soil salinity in the Yellow River delta region of China using EO-1 hyperion data,” Pedosphere 20, 378–388 (2010).PDOSEAhttp://dx.doi.org/10.1016/S1002-0160(10)60027-6Google Scholar
F. Wang et al., “Detecting soil salinity with arid fraction integrated index and salinity index in feature space using Landsat TM imagery,” J. Arid Land 5, 340–353 (2013).http://dx.doi.org/10.1007/s40333-013-0183-xGoogle Scholar
S. Ayetiguli et al., “Estimating soil salinity in Pingluo county of China using QuickBird data and soil reflectance spectra,” Int. J. Appl. Earth Obs. Geoinf. 26, 156–175 (2014).http://dx.doi.org/10.1016/j.jag.2013.06.002Google Scholar
N. Said et al., “Modeling and mapping of soil salinity with reflectance spectroscopy and landsat data using two quantitative methods (PLSR and MARS),” Remote Sens. 6, 10813–10834 (2014).http://dx.doi.org/10.3390/rs61110813Google Scholar
D. Moghadas et al., “Joint full-wave form analysis of off-ground zero-offset ground penetrating radar and electromagnetic induction synthetic data for estimating soil electrical properties,” Geophys. J. Int. 182, 1267–1278 (2010).http://dx.doi.org/10.1111/j.1365-246X.2010.04706.xGoogle Scholar
D. M. Andrea et al., “Early-Time GPR signal attributes to estimate soil dielectric permittivity: a theoretical study,” IEEE Trans. Geosci. Remote Sens. 3, 1643–1654 (2013).IGRSD20196-2892http://dx.doi.org/10.1109/TGRS.2012.2206817Google Scholar
L. C. Liu, “A preliminary study of soil salinization in the northern Tarim Basin,” Chin. J. Soil Sci. 5, 196–200 (1986).Google Scholar
N. Majid et al., “Evaluation of atmospheric correction models and Landsat surface reflectance product in an urban coastal environment,” Int. J. Remote Sens. 35, 6271–6291 (2014).IJSEDK0143-1161http://dx.doi.org/10.1080/01431161.2014.951742Google Scholar
L. G. Zhou et al., “Estimation of higher chlorophylla concentrations using field spectral measurement and HJ-1A hyperspectral satellite data in Dianshan Lake, China,” ISPRS J. Photogramm. Remote Sens. 88, 41–47 (2014).IRSEE90924-2716http://dx.doi.org/10.1016/j.isprsjprs.2013.11.016Google Scholar
H. J. Tian et al., “Estimation of hloroph -yll-aconcentration in coastal waters with HJ-1A HSI data using a three-band bio-optical model and validation,” Int. J. Remote Sens. 35, 5984–6003 (2014).IJSEDK0143-1161http://dx.doi.org/10.1080/01431161.2014.934403Google Scholar
Y. K. Li et al., Soil Agricultural Chemical Routine Analysis Methods, Science Press, China (1983).Google Scholar
S. D. Bao, Agro-Chemicals Analysis of Soil, pp. 178–200, China Agriculture Press, Beijing, China (2000).Google Scholar
A. Savitzky and M. J. E. Golay, “Smoothing and differentiation of data by simplified least squares procedures,” Anal. Chem. 36, 1627–1639 (1964).ANCHAM0003-2700http://dx.doi.org/10.1021/ac60214a047Google Scholar
E. A. Cloutis, “Hyperspectral geological remote sensing: evaluation of analytical techniques,” Int. J. Remote Sens. 17, 2215–2242 (1996).IJSEDK0143-1161http://dx.doi.org/10.1080/01431169608948770Google Scholar
Z. Huang et al., “Estimating foliage niteogen concentration from HYMAP data using continuum removal analysis,” Remote Sens. Environ. 93, 18–29 (2004).http://dx.doi.org/10.1016/j.rse.2004.06.008Google Scholar
J. Farifteh et al., “Quantitative analysis of salt-affected soil reflectance spectra: a comparison of two adaptive methods (PLSR and ANN),” Remote Sens. Environ. 110, 59–78 (2007).http://dx.doi.org/10.1016/j.rse.2007.02.005Google Scholar
L. Lei et al., “Study on the soil salinization monitoring based on measured hyperspectral and HSI data,” Spectrosc. Spectral Anal. 34, 1948–1983 (2014).Google Scholar
N. B. Julio et al., “Use of near infrared hyperspectral tools for the screening of extractable polyphenols in red grape skins,” Food Chem. 172, 559–564 (2015).FOCHDJ0308-8146http://dx.doi.org/10.1016/j.foodchem.2014.09.112Google Scholar
V. Toffanin et al., “Effectiveness of mid-infrared spectroscopy for prediction of the contents of calcium and phosphorus, and titratable acidity of milk and their relationship with milk quality and coagulation properties,” Int. Dairy J. 41, 68–73 (2015).IDAJE60958-6946http://dx.doi.org/10.1016/j.idairyj.2014.10.002Google Scholar
Hongnan Jiang was working as a lecturer in Xinjiang University of China. His main research interest focused on the application of remote sensing on ecology and environment problems.
Hong Shu is a professor. His research has been focused on spatiotemporal geostatistics, spatiotemporal data mining, and data assimilation.
Lei Lei is a research assistant at Xinjiang University. His research has been focused on remote sensing application.