Machine learning method for quick identification of water quality index (WQI) based on Sentinel-2 MSI data: Ebinur Lake case study

With the continuous development of water resources, water quality has become a popular issue of global concern (Pahlevan et al. 2019). The area of human activity continues to expand worldwide, and the impact on water quality is gradually becoming more severe (Mananze et al. 2018). The deterioration of water quality due to pollution is increasing due to urban cluster pollution emissions, industrial pollution emissions, and agricultural fertilizer pollution emissions (Shao et al. 2006; Lintern et al. 2018; Wang et al. 2019). Water pollution has had a considerable impact on human production and life and increased the health risks of ecosystems (Chen et al. 2020). Therefore, water quality and safety assessments are extremely important. Water quality has become the most important indicator in aquatic ecosystems. The quality of water directly affects the health of aquatic ecosystems and fully reflects the spatiotemporal heterogeneity and integrity of river water environment ecosystems (Staponites et al. 2019). Therefore, it is necessary to monitor the pollution of water quality in real time and record changes in water quality.

In the current situation, remote sensing can be used to accurately and quickly monitor river water quality, and there is an opportunity to regularly monitor aquatic systems using different sensor satellite products (Tyler et al. 2016). Remote sensing technology has become an important method in water resource surveys and monitoring, the remote monitoring of water environments, regional ecological remote sensing monitoring, urban environmental remote sensing monitoring and wetland protection. Reducing the interference of non-water factors and enhancing the expression of water body factors are key tasks in remote sensing to identify water body information (Li et al. 2016). Landsat image recognition technology for water bodies has been well established around the world. With the continuous deepening of machine learning methods, more and more water quality researches use this method to predict water quality (Gao et al. 2019; Hussein et al. 2019; Liang et al. 2020). This technology combines artificial neural network (ANN), support vector machine (SVM), and maximum likelihood (ML) classification methods to reflect water changes (Rokni et al. 2015). In multivariate data research, the partial least squares regression (PLSR) model can show better results in modeling and forecasting a large number of water quality indexes (Carrascal et al. 2010). And PLSR is widely used in the establishment of multivariate models, which can handle the collinearity between various variables and further reduce and weaken related data sets (Sidike et al. 2014; Wang et al. 2020). Random forest (RF) has a good application basis for water quality index prediction (Meyers et al. 2017), support vector machine (SVM) can achieve a better effect on nonlinear relationships when used alone (Nawar et al. 2016; Lucà et al. 2017). In PLSR-SVM, not only high-quality results can be obtained, but also spectral bands can be combined to improve the inversion effect (Zhang et al. 2021).

Additionally, this technology applies a cluster analysis method to establish a stable water index, extract surface water information (Wang et al. 2018b) and establish an automatic classification set (Fisher et al. 2016). The remote sensing of water quality mainly includes water color remote sensing and water quality remote sensing. Water color remote sensing mainly uses remote sensing to invert the optical characteristics and spectral characteristics of water bodies, such as the suspended matter and chlorophyll a levels. Conversely, if there are no significant optical and spectral properties for a given water body, the remote sensing inversion of other WQPs that have close relationships with direct WQPs can be achieved by water quality remote sensing; such parameters include total nitrogen and total phosphorus (Ma & Dai 2005). In research study, based on the change in chromaticity and the inversion of water body radiation and WQPs through spectroscopy, chromaticity is used to predict water quality. The spectral complexity of inland lakes has also gradually been considered (Bukata et al. 1983).

In 1991, Anatoly Gitelson (1991) studied the remote sensing of inland water quality through aerospace remote sensing and studied the potential associated with using chlorophyll a (Lintern et al. 2018), dissolved organic matter (DOM), and suspended solids (SM) concentrations. BABAN and Serwan (Baban 1993) used Landsat satellites to perform a regression analysis of TM data and surface WQPs in 1993 and to simulate and predict the water quality of inland lakes. Since 1990, Chinese scholars have studied the water quality of inland rivers through remote sensing (Yu et al. 2008). The research on chlorophyll and suspended solids has been continuously improved, and the inversion accuracy also been improved (Ma & Dai 2005). In the inversion of the water quality of inland rivers, by selecting different spectral band combinations and combining the water quality spectral characteristic parameters, the accuracy of establishing estimation spectral and water quality estimation models has been improved. With the continuous advancement of remote sensing technology, new methods of establishing the correlations with water quality through different wave bands of remote sensing satellites have been developed.

In recent years, Landsat 8 and Sentinel-2 have been widely used for remote sensing and the monitoring of water resources. Remote sensing methods have been used to extract surface water information from images, to analyze the sensitivity of different spectral bands to the surface water quality and to distinguish surface water quality levels. Rapid extraction requires technical support (Zheng et al. 2015). Stéfani (Stéfani et al. 2017) studied the effects of total suspended solids (TSS) and chlorophyll (Lintern et al. 2018) levels on the physiological response of oysters, highlighting the difference between the near-infrared (NIR) bands of Landsat imagery and Sentinel-2A imagery, which plays a significant role in band quantifying TSS. Fernanda (Fernanda et al. 2017) found that the NIR band of Sentinel-2A is more sensitive to Chl in Brazil's super-eutrophic reservoirs. Yang (Yang et al. 2018) found that Sentinel-2 multispectral satellite imagery can effectively reduce noise and be used to extract water more accurately than other products in a study of the automatic extraction of urban surface water. Landsat operational land imager (OLI) data are different from data corresponding to the four Vegetation red edge (VRE) bands of Sentinel-2 multispectral imagery (MSI), with better sensitivity to Chl. Additionally, the NIR bands of alternative algorithms and concentrations of Chl pigments have been effectively obtained (Moses et al. 2009). As a result, MSI data can be used to effectively monitor the growth and reproduction of various algae in water (Gower et al. 2008). Combining NIR (0.842 μm) and two shortwave infrared (SWIR) systems based on visible light (Halgamuge & Davis 2019) can enhance water inversion and the efficiency of vegetation detection. The WQI is a water index that covers all water resource parameters (Terrado et al. 2010). The WQI appears to lag behind the spatial and temporal changes in water quality samples when evaluating individual water samples, and problems related to the continuous large-scale comprehensive evaluation of water quality changes must be solved with the development of remote sensing technology (Wang et al. 2017). In this paper, WQPs are extracted to establish the corresponding WQI; then, the effects of the most prominent parameters of river basin water on river water quality are investigated.

The river water ecosystem in the Ebinur Lake Basin, Xinjiang, was studied. The Ebinur Lake Basin is located in an arid zone and adjacent to the largest arid zone in the temperate region of the Northern Hemisphere. The basin has typical regional characteristics, large climate changes and considerable evapotranspiration, and the artificial large-scale extraction of water for agricultural production has led to regional ecological and hydrological changes. Intensive non-point source pollution caused by the development of water and soil resources has affected the water quality of the basin to a certain extent. To this end, it is necessary to quickly identify and manage water quality through remote sensing images and provide technical support for environmental protection and safety.

In this article, the water quality in arid regions is studied through an analysis of the correlation between the water index and the WQI based on a spectral combination method, and a correlation analysis between the band combination index and the WQI is performed to identify the optimal WQI. The spectral band combination that is most suitable for water quality in arid areas is also determined. Furthermore, the effects of different parameters on the water quality in the entire basin is studied to provide important technical support for the future macrocontrol of water bodies.

The study was performed in the arid region of the hinterland of the Eurasian continent, the Ebinur Lake Watershed of northwestern Xinjiang, China (Figure 1). The Ebinur Lake Watershed is relatively complex. This watershed is located in the northwestern part of the Junggar Basin and is the lowest lake subbasin in the Junggar Basin. The northeastern part of the area is connected to the Gurbantunggut Desert, and the northwestern and western parts are the Alatau Mountains and low hills, respectively (Abuduwaili et al. 2008). Additionally, the Alashankou region, which is the largest outlet in China, is part of this basin. The valley is surrounded by mountains to the north, west and south. The plain is approximately 220 km from east to west and 60–120 km from north to south. The total area of the basin is 2,080 km². Ebinur Lake is located in the low-lying area of the river tail basin (Yu & Jiang 2003), and the source of recharge for the lake is the surrounding surface water and groundwater. The water of Ebinur Lake is mainly composed of water from several inflowing rivers, such as the Jing River, Boertala River and Kuitun River. The complexity of the geographical environment has affected the ecological environment of the Ebinur Lake Basin to varying degrees. The water quality of river water directly affects the growth of vegetation in the basin, the degree of salinization of the soil and the living conditions for human beings (Wang et al. 2017).

A multispectral instrument was launched onboard the Sentinel-2A satellite on June 23, 2015, from the European Space Agency. Sentinel-2A has 13 bands and three different spatial resolutions. The MSI ranges from VIS to NIR and SWIR, with spatial resolutions of 10, 20, and 60 m, respectively (Table 1). Notably, the MSI is the only terrestrial remote sensing satellite in the world with a ground resolution of up to 20 m. Four specialized bands (B5, B6, B7, and B8a) were designed to obtain the spectral characteristics of the vegetation in the near-infrared ‘red edge’ region (690–800 nm), and these bands are close to B4 at the red wavelength (Peterson et al. 2020).

We downloaded the Sentinel 2 image from the ESA Sentinel Scientific Data Hub (https://scihub.copernicus.eu/); this data is a radiometrically calibrated L1C product. We considered the data of field survey to choose the date of the image (October 2017, cloudless). To achieve the L2A product, we conducted atmospheric correction using the sen2cor toolbox in SNAP software (Main-Knorn et al. 2015). Meanwhile, geometric correction is carried out to ensure the verification accuracy to ±0.5 pixels. Using ENVI5.5 software, the 20-m and 60-m resolution bands were resampled to 10 m. To verify the water spectrum, the spectrum of the typical sample was extracted as Figure 2. In general, the water spectrum is obvious reflection in the B3 band, and the absorption continues to increase after the B4 band. The absorption effect is best in the B8a band, which is consistent with the changes in the water in the spectral range. By comparison, these spectra were accurate in this study.

River water quality data selection is mainly based on China's surface water environmental quality standard GB3838-2002 (GB3838-2002 2002); in this process, water samples were collected from the basin, and outdoor and indoor tests were conducted.

The sampling sites were located on the Boertala River in the west, the Jing River in the south and the Kuitun River in the east. Samples were collected from 38 sample points according to the trends of the three rivers, and samples from surface water bodies with a depth of 50 cm on the surface section of the river were extracted. One-liter water samples were collected, which were stored in polytetrafluoroethylene plastic bottles, and packed into benzene board insulation boxes with ice cubes and quickly transported back to the laboratory for the determination of the water quality indices. Water quality indicators were experimentally tested at each water sample point, and the measurement range of each indicator was strictly screened according to China's surface water environmental quality standard GB3838-2002 (GB3838-2002 2002), and passed based on the ‘Water and Wastewater Monitoring and Analysis Method’ of the State Environmental Protection Administration (State Environmental Protection Administration of China 2002). Twenty-two kinds of water quality indicators were experimentally monitored (Table 2).

The following 22 parameters were analyzed: ammonia nitrogen (NH₃-N), chemical oxygen demand (COD), biological oxygen demand (BOD₅), dissolved oxygen (DO), total nitrogen (TN), total phosphorus (TP), total hardness (TH),total suspended solid(SS), Chroma, turbidity, phosphate, hexavalent chromium, sulfide, Fe, Cu, Zn, volatile phenol, the chloride concentration, the cobalt concentration, the salt content (CON), total dissolved solids (TDS), and the pH. First, the data or named z-scores were standardized. Additionally, the concentration of the data through this step was analyzed, and the parameters were further extracted.

Then, a redundancy analysis (Kaplan & Avdan 2017) was performed on the 22 WQPs. The RDA involved a sorting method that combined regression analysis with principal component analysis (PCA) and an extension of multiresponse regression analysis. Conceptually, the RDA is a PCA of the fitted-value matrix of the multivariate multiple linear regression between the response variable matrix and the explanatory variables. The parameters in the RDA diagram are determined by the length and direction of the arrows, which represent the contribution of each variable to the two main components of the biplot.

There are many established water indices that can be applied to Landsat TM/ETM + data. This study selected four classic water indices for analysis. Based on the Landsat WQI, the improved WQI was developed to reflect the inversion of water quality in each band of Sentinel-2. The normalized difference water index (NDWI) was selected (McFeeters 1996; Yin et al. 2018; Alghamdi et al. 2020) by choosing the two optimal bands of water information from the remote sensing data. The optimal index was obtained by the spectral calculation of the green band and NIR of the VIS. A new water index was established; the automated water extraction index with no shadows (AWEI-nsh) can effectively remove the non-water pixel index information and is suitable for cases in which shadows are included in the image. Luo Chongliang (Luo 2015) extracted the Ebinur Lake boundary using the enhanced water index and normalized water index (NWI); these researchers found that the two types of indices were suitable for water extraction in the Ebinur Lake Basin and that the extraction accuracy was sufficient (Table 3).

As there are a variety of chemical, physical and biological WQPs, several researchers have proposed a WQI in the form of a simple expression for representing the general quality of surface waters (Zeinalzadeh & Rezaei 2017). This study constructed a WQI as a weighted multifactor environmental quality index that highlights the maximum values of major water quality indicators. The WQI can be used to extract and reflect WQPs and the water quality and composition (Wang et al. 2017).

The WQI is a quality index that directly and comprehensively reflects water quality (Ewaid et al. 2018). This index can quantify the degree of river water pollution objectively, scientifically and reasonably, and it can comprehensively and effectively reflect the effectiveness of comprehensive improvements in rivers. Compared with a single water quality assessment, WQI is a comprehensive water quality assessment method that integrates multiple water quality variables into an intuitive value (Zotou et al. 2020). This method can not only evaluate the overall situation of water quality (Şener et al. 2017), but also eliminate the differences caused by a single water body. Effective selection of the relevant water quality index can better apply WQI to water quality assessment at all levels (Pesce & Wunderlin 2000), and construct a WQI based on the actual water resources in the study area. This article contains many variables of water quality parameters, and these parameters are not all influential in terms of overall water quality. It is a relatively better way to invert the overall water quality by selecting water quality parameters with greater weight in the PCA analysis. This is a good, comprehensive water quality evaluation method (Şener et al. 2017).

In recent years, there have been more and more researches on the use of machine learning to model data, and modeling methods that can provide choices are also endless, but the three modeling methods (SVM, RF, PLSR) selected in this article are relatively mature machine learning methods, so this article chooses these three methods to model and predict the data, which can effectively improve the accuracy of data prediction.

An SVM is a supervised classifier based on small samples in machine learning that can quickly and accurately fit and predict samples; this method was proposed by Boser (Boser 2008). The SVM finds the difference between the maximal margins by points in different categories in the sample space; the results are mapped to the same sample space for classification and prediction of different categories (Zhao et al. 2018). This method has mainly been applied to classification and regression problems, where it is better applied. The regression model of the SVM is divided into a linear model and a nonlinear model (Kaya 2013). It is not invertible in the linear model, and the non-linear model can add the kernel function to the linear model (Nie et al. 2020). The method essentially finds a plane in the data such that all of the data in the set conform to the nearest plane distance.

The random forest (RF) is an algorithm based on classification and regression through a large number of decision trees as weak classifiers. The generated classifiers are diverse and aggregated (Halgamuge & Davis 2019). The RF selects the bagging method and randomly returns corresponding decision trees while continuing to select different training subsets, and then it votes on these decision tree classification results to determine the final classification result (Houborg & McCabe 2018). However, the RF and the traditional decision tree algorithms employ different training processes. The traditional decision tree selects the optimal attribute each time it selects the partition attribute, whereas the RF introduces random attribute selection when selecting the partition attribute of the node. The principle is similar to the tournament selection method in genetic algorithms (Peters et al. 2007). To achieve two random characteristics of the RF, classification features are randomly selected and the training subsets are randomly generated and are independent and identically distributed. With these two characteristics, various data sets can be quickly extracted without repeated accumulation, which is better for data classification.

PLSR is a new multivariate statistical analysis method (Geladi & Kowalski 1986). It is a regression-based modeling method based on a multivariate dependent variable corresponding to multivariate analysis. This linear model is better than the one-by-dependent variable for regression analysis based on the internal linear height correlation. However, PLSR can cope with the problem of a small sample size. This method combines the advantages of the canonical correlation analysis, PCA and multiple linear regression analysis to improve prediction and nonmodeling prediction types as well as data cognitive analysis. The optimal number of potential variables to be predicted in the future can be determined based on the number of factors with the smallest RMSE (Douglas et al. 2018).

In this study, 22 WQPs were chosen, normalized and introduced to a system with a standard value of 1 and an average of 0 for the analysis. To perform an accurate and economically feasible analysis of all the parameters, it is necessary to screen all the parameters and extract the indicators with the highest contributions.

For the 22 WQPs, the Z-score analysis is shown in Figure 3. For this box and whisker plot, the Y-axis represents the 22 WQPs, and the X-axis represents the WQPs based on normalized values. The green line in the middle of the box is the median of all data; for all the data, the median value is on the left side of 0. Additionally, the line is on the right side of 0 for DO and the pH. The left side of the blue border represents 25% of the data, and the right side represents 75% of the data. The overall quartiles of the data are small, and the middle part of the data is the most concentrated. The blue box enclosed by the black whiskers has the symbol ‘|’, which denotes the maximum and minimum values of the data. The red vertical lines on the left and plus signs on the right are the extensions of the whiskers to the most extreme data points that are not considered outliers.

The figure shows that most of the small and concentrated data have large abnormal values, indicating that the water quality at the sampling points in different rivers is very different and the values at some sampling points may exceed the standard; that is, all of the data sets have outliers. The parameters with more than four outliers include the TSS, turbidity, sulfur, Cu, Zn, volatile phenol, and cobalt concentration, and the largest outliers appear for these indicators. The water quality data are standardized, and it is clear that the median value of the data appears below the average in these cases.

In the first principal component of the horizontal axis (Figure 4), the positive correction coefficients of phosphate, Zn, TN, COD, BOD₅, the cobalt concentration, the salt content, hexavalent chromium, the chloride concentration, TDS, chroma, and NH₃-N are shown. Additionally, suspended solids, Fe, TP, turbidity, Cu, and sulfur volatile phenols are plotted. Moreover, on the vertical axis of the second principal component, the negative correction coefficients are DO and the pH. These points are scaled with respect to the maximum score value and the maximum coefficient length; therefore, only the relative locations can be determined from the plot.

Pearson's linear correlation is a statistic that reflects the linear correlation between two variables by the one-to-one correspondence of multiple parameters. After the redundancy analysis of the data, the nine parameters with the highest contributions were selected for linear correlation analysis. It can be seen from the correlation analysis in Table 5 that COD and TDS have the highest correlation with BOD₅ and TN, and the correlation coefficient is 0.891 at the 0.01 significance level. The COD in water will reflect the size of the electrolytes in the water, and BOD₅ has a certain restriction on the TN content in the water. Additionally, the TN content also limits changes in the COD in the water. The TDS is also influenced by the COD, BOD₅, DO, TN and salt content in the water. As shown above, each parameter in the water will be affected by other factors, and the water quality will change.

Sentinel-2 MSI imagery was used to retrieve the typical water index values in July 2016. Based on SPSS software, the correlation between the four classical water quality indices extracted from remote sensing images, and the nine WQPs were analyzed. The results are shown in Table 6.

According to the correlation analysis, the water quality parameters with high correlations were the COD, BOD₅, DO, TN, TSS, salt content, TDS, pH and NWI, suggesting that the NWI can effectively reflect most WQPs. The correlation between the NDWI and turbidity was as high as 0.128. Compared with other water indices, the band was not sensitive to water turbidity, and the turbidity correlation was generally low. The correlations between the NDWI and TDS and COD and TN were above 0.5. The EWI exhibited correlations with TN, the salt content and TDS of more than 0.5. The correlation between the AWEI-nsh and COD was 0.558.

In general, the correlation between the NWI and various water quality indicators in the classic water index was good, indicating that the WQI is sensitive to bands 1, 8, 11, and 12. Remote sensing image extraction is a sensitive reflection of the water indices.

By using 2D spectrum modeling, the multiband pixel values of the corresponding points on each image point are extracted, and three empirical indices, normalized index, differential index and ratio index, are established. Correlation analysis was performed on nine water quality parameters and 13 bands of Sentinel-2 sampled at each point. The analysis results are shown in Figure 5.

The color bar on the right side of the figure shows the correlation between the band combination and WQPs, where dark red is the maximum value of positive correlation and dark blue is the maximum value of negative correlation. As can be seen from the figure, the correlation coefficients of the two WQPs and band modeling effects, COD and TDS, are better than 0.7.

From the correlation among the water index and the spectral band index and the measured WQPs, the correlation coefficient between the water index and the WQP is below 0.7, and the highest is 0.679. In contrast, the spectral band is established. The index correlation is better. Although the index can only reflect the combination algorithm of two bands, it can also find the correlation of the optimal band and the band combination algorithm from 13 bands. Overall, the nine WQPs selected can be reflected in different bands of the image.

Table 7 shows the correlation results for the nine water quality parameters and the classic water index and the 2D modeled spectral index. The classic water index with the highest correlation among the nine water quality parameters was extracted, and the three water-based indicators with the highest correlation were extracted from the three 2D modeled spectral indices. Overall, compared with the classic water index, the 2D modeled spectral index has a better effect in water quality extraction. The overall effect is best when the TDS is inverted by the DI, and the correlation coefficient reaches 0.787, while the correlation coefficient of the NWI inversion of TDS is 0.679. Based on the two correlation coefficients of the nine WQPs, the correlation coefficients of the BOD₅, TN, TSS, salt content, and TDS modeled spectral indices are not much different from those for the water indices. This result suggests that the two methods have the same effect on the inversion of these five WQPs. In the inversion of DO, turbidity and the pH, the 2D modeled spectral index is much different than the water index. The inversion results of the classical water index for DO and the pH had negative correlation coefficients of −0.412 and −0.393, respectively. The inversion results of the 2D modeled spectral index showed positive correlation coefficients of 0.538 and 0.583. The two methods display a large difference in the inversion of turbidity, and the 2D modeled band index can better invert the degree of turbidity in water.

Overall, the 2D modeled band index can effectively invert the WQI, and the accuracy is higher than that of the water index. Thus, the 2D modeled band index is the optimal spectral index for water quality monitoring.

To estimate the overall water quality, a WQI index was introduced. WQI is a comprehensive evaluation of water quality indicators. It can analyze the overall effect of water quality through selected indicators, so that it can analyze the overall water quality situation more intuitively. We employed nine types of water quality selected from RDA in the WQI calculation, obtained WQI of 38 sampling points, and compared and analyzed WQI in the water index and modeled spectral index, respectively. As can be seen in Table 7, among the correlations between WQI and the four water indexes, the correlation between WQI and NDWI reached 0.341 at the level of 0.05, and the correlation with the other three water indexes was low.

Correlation analysis between WQI and three modeling indexes shows that RI of WQI and NDI, RI, and DI are 0.63, 0.69, and 0.62, respectively (Figure 6). The modeled water index has a better correlation with WQI than the classic water index.

The WQI was calculated from the nine selected WQPs, and the WQI values of 38 sampling points were obtained. For prediction, 24 models were selected, and 14 values were used for prediction verification. Machine learning (SVM and RF) and PLSR methods were chosen for prediction. The machine learning model was used as the main model for prediction, the partial least squares model was used as the auxiliary model, and the PLSR-SVM model was established for comparison to analyze the accuracy of each model.

From Table 8, by calibration modeling with 24 values, the modeling of the machine learning methods was performed. The results indicate that the RF modeling effect yields the best correlation (R²c = 0.814 and RMSEc = 109.411); however, the SVM modeling effect (R²c = 0.0.770 and RMSEc = 120.629) is preferred. Machine learning is a linear process, and the linear correlation of the PLSR method was low. The correlation coefficient and root mean square error of the PLSR-SVM are better than those of the previous three methods; notably, the correlation of R²c = 0.873 is better, and the degree of data dispersion is better at RMSEc = 102.334.

From the validation dataset, the modeling effect R²v displays the following order: PLSR-SVM > RF > PLSR > SVM; additionally, the highest R²v is 0.870. From RMSEv, SVM > RF > PLSR > PLSR-SVM, and the root mean square error (RMSEv) is small, indicating that the data are not discrete and the prediction effect is good. Additionally, RPD = Stdev/RMSE; a large RPD corresponds to a satisfactory prediction effect. The result indicates that the prediction effect of PLSR-SVM is better than that of the other methods, with RMSE = 62.927 and RPD = 2.755.

The WQI is modeled and predicted with the four modeling methods. Figure 7(a)–7(d) shows the modeling of 24 data points. In the modeling diagram, the red line around the red points is the fitted line. The surrounding red interval is the confidence interval. The smaller the interval is, the better the data modeling effect. It can be seen from the figure that the interval of the PLSR-SVM method in Figure 7(d) is smaller than that of the other three model regions. At R²c = 0.873, the model effect is best. In Figure 7(e)–7(h), the predictions involve 14 data points and the original data; the blue point is the prediction point, the blue line is the prediction fitting line, and the blue interval is the confidence interval of the prediction. In the interval, the confidence interval area of Figure 7(f) and 7(h) is the smallest, although the angle between the fitted line and the 1:1 line of Figure 7(h) is the smallest. Additionally, the predicted points are all distributed around the 1:1 line, and the predicted values yield R²v = 0.87 and RPD = 2.755, which reflect the good prediction effect of the PLSR-SVM method.

The reflectance of different spectral bands in remote sensing images is different from the reflectance of surface vegetation, soil and water. For such problems, differences are often analyzed (Pahlevan et al. 2019). Extracting the water pixels of all pixels from satellite images is the most important step in remote sensing detection, and it is also a popular issue in current research. Long-term water masks can be used to analyze time-varying water information and provide a reliable basis for surface water research (Fisher et al. 2016). Sentinel images can be used to extract water features under complex terrains such as mountains, snow, and cities (Kaplan & Avdan 2017). Water indices based on spectral bands can be used to quickly identify surface water and highlight certain features (Xu 2006). This article studies the extraction of the DN values corresponding to the Sentinel-2 data and sampling points and establishes a WQI and a two-dimensional spectral model based on this approach. First, the correlations between the water index and individual WQPs are studied. As shown in Table 6, the correlation between the NDWI, NWI, EWI and TDS is better than 0.5. The effect of AWEI-nsh on COD inversion reaches 0.558. Notably, the water spectral index can be used to invert the corresponding WQPs in the band calculation. Based on this method, we have established a two-dimensional spectral model of the Sentinel-2 MSI and three model indices; namely, the NDI, RI and DI. By comparing the correlation between the water index and the 2D spectrum model with that for a single WQI, the effect of the two-dimensional spectrum model is better than that of the water index. The two-dimensional spectrum model can filter the spectrum information and select the optimal band combination for analysis (Hong et al. 2018); such methods used in water quality monitoring will allow us to find more sensitive bands and band combinations for a single type of water quality. Through the correlation analysis of WQPs, water quality indices and spectral bands, the water quality of sampling points in the basin can be analyzed in depth.

In the Ebinur Basin, located in an arid area, many scholars have used Landsat OLI data to extract water quality information through different methods, identify water bodies in arid areas through different modeling methods, and analyze the correlation between the WQPs and water indices (Li et al. 2016; Wang et al. 2017). To extract and analyze the overall water quality of the river basin and facilitate an analysis of the overall water quality safety status, single types of WQPs are integrated into the comprehensive water index WQI. The WQI can be used to establish the relationship between multispectral bands and water quality reflectance, analyze the correlation between spectral bands and water quality, and establish a linear relationship with the sensitive bands of different indicators related to water quality (Feyisa et al. 2014). Based on the preliminary exploration of single types of WQPs and spectral modeling, we selected an improved 2D band model and WQI for further comprehensive water quality research. Figure 7 shows that the correlation between the three two-dimensional spectral modeling methods and the WQI is above 0.75, and the modeling effect is satisfactory. The WQI is based on a combination of different WQPs and converts a large number of WQPs into a single quantity (Sánchez et al. 2007; Feng et al. 2018). This method is currently widely used in water quality evaluation studies (Wu & Chen 2013). To conduct further prediction research on water quality, on this basis, the WQI was modeled and predicted by machine learning methods. This paper established four models: SVM, RF, PLSR and PLSR-SVM models. The PLSR-SVM provided a better model result and prediction of the basin WQI than did the other models. The PLSR-SVM and RF models have been used in other river basin water quality studies, and the prediction results have been shown to be better and more accurate than those of other models (Mananze et al. 2018). This finding confirms that the method used in this study is reasonable and provides a foundation for subsequent research.

In future research, it is necessary to consider the distribution of water samples and the selection of measured data in the entire basin.

This study explores the relationship between the water quality in the Ebinur Lake watershed and the multispectral bands of Sentinel-2 MSI data. The relevant WQPs are related to the water index and analyzed by spectral bands. The following results are obtained.

The Z-score and RDA are used to reduce 22 WQPs to nine while dividing them into different groups. COD, BOD₅, DO, TN, TSS, turbidity, the salt content, TDS and the pH are the nine selected WQPs with high contribution values. TDS, COD, and TN are the most influential WQPs.

The WQI is established through the selected nine WQPs, and modeling and prediction are performed through machine learning and linear correlation models. The PLSR-SVM model with a linear correlation and machine learning is the best model for modeling, with R²v = 0.87 and RPD = 2.755; the predictions with this approach are very accurate, and this approach can provide an effective method for water prediction.

This work was supported by the National Natural Science Foundation of China (41771470), the National Natural Science Foundation of China (41961059), and the National Natural Science Foundation of China (41561089). We are especially grateful to the anonymous reviewers and editors for reviewing the manuscript and offering instructive comments.

All relevant data are included in the paper or its Supplementary Information.