Enhanced Water Quality Inversion in the Ningxia Yellow River Basin Using a Hybrid PCWA-ResCNN Model: Insights from Landsat-8 Data

Abstract

The real-time monitoring and evaluation of water quality provides a scientific basis for water resource management and promotes regional sustainable development. This study established a database using Landsat-8 satellite data and water quality data from the Ningxia Yellow River basin in China, spanning 2021 to 2023, and this paper proposes a custom residual convolutional neural network model with a hybrid attention mechanism, referred to as PCWA-ResCNN. The accuracy of the model in predicting turbidity, permanganate, ammonia nitrogen, and dissolved oxygen concentration was more than 95%. Compared to convolutional neural networks and long short-term memory models, this model performed better in predicting water quality parameters with significantly improved prediction performance. In terms of spatial distribution, the pollution degree in the middle reaches of the basin is relatively serious. However, the overall water quality is good, being mainly Class I and Class II water quality. The hybrid model established in this paper can better capture the complex nonlinear relationship between the observed values and the surface water reflectance, showing strong robustness. This model can be used for the water quality monitoring of complex inland rivers and lakes, and it can also provide effective support for relevant government departments to formulate scientific and reasonable water quality management policies.

1. Introduction

Water quality monitoring is an important basis for water quality evaluation and water pollution management. The quality of water has a significant impact on human health, ecosystems, economic development, and other areas. Therefore, improving water quality and strengthening water environmental protection have become issues that must be addressed urgently [1,2,3]. The Yellow River is one of the most important rivers in China. The river plays an important role in irrigating farmland, supplying water resources, and maintaining the ecological environment, and Ningxia is the only province whose entire territory is situated in the Yellow River basin. Therefore, large-scale, accurate, and rapid monitoring of water quality is of particular importance [4,5]. Although detailed water quality parameters can be obtained via on-site water quality measurements, timely and large-scale monitoring faces difficulties due to the complicated geographical environment of the Yellow River basin, which is easily restricted by weather and hydrological conditions [6]. Remote sensing has the advantages of a wide detection range, strong data comprehensiveness, and no restrictions imposed by conditions on the ground; therefore, the application of remote sensing technology can improve monitoring efficiency and accuracy, reduce monitoring costs, and provide scientific support for water resource management and environmental protection.

In recent years, the concentration estimation of water quality parameters combined with artificial intelligence technology has been extensively utilized in the field of water quality monitoring. Liu et al. [7] used the LSTM model to conduct a detailed analysis of dissolved oxygen, electrical conductivity, and turbidity in terms of the water quality of the Yangtze River. Cui et al. [8] established the point-center regression convolutional neural network (PSRCNN) of Sentinel-2 and Landsat-8 images to evaluate water transparency and achieved satisfactory results. He et al. [9] established the inversion model of water quality parameters of inland reservoirs by using a BP neural network, examined water quality parameters such as chemical oxygen demand (COD) and permanganate (CODMn), and evaluated the eutrophication of reservoirs. The W-WaveNet multi-site water pollution prediction method proposed by Yang et al. [10] combines adaptive graph convolution, a convolutional neural network, and a long short-term memory network, showing stable performance in predicting long series. Ni et al. [11] proposed a deep learning model, FTGCN, based on a graph convolutional network to realize the prediction of multivariate water quality data. Syariz et al. [12] developed a convolutional neural network WaterNet model for Chl-a concentration retrieval, which is superior to the artificial neural network (ANN) model. In addition, Sagan et al. [13,14] used spectral index, biooptical simulation, machine learning, and cloud computing techniques to monitor inland water quality, including dissolved oxygen (DO) and turbidity (TUB), and they analyzed the potential and limitations of these methods. Yang et al. [15,16,17,18] used convolutional neural networks and different remote sensing satellite data to invert Chl-a water quality parameters. Zhang et al. [19] predicted the water quality of ammonia nitrogen (NH₃-N) and permanganate (CODMn) in Dongping Lake of the Yellow River in China based on the convlstm model, with it showing satisfactory performance. Parra et al. [20] carried out research combining machine learning and RGB sensors to quantify and classify water turbidity (TUB), showing their application potential. Kuang et al. [21] used augmented learning to predict dissolved oxygen (DO) content in wireless sensor networks and demonstrated the advantages of machine learning in optimizing sensor network deployment and data acquisition strategies.

The authors of the above studies explored the significant development of the application of machine learning algorithms combined with remote sensing and sensor technology in water quality monitoring. However, both remote sensing technology and sensor technology have advantages and disadvantages: sensor technology provides high accuracy and real-time monitoring capabilities; however, it is constrained by limited coverage and high deployment and maintenance costs [22]. Remote sensing technology can realize large-scale water quality monitoring; however, its accuracy and real-time performance may be limited by weather conditions and the complexity of data processing [23]. The authors of future studies will need to continue to optimize models, improve data processing capabilities, and address current challenges in remote sensing monitoring to achieve comprehensive monitoring and effective management of water quality. Through the continuous optimization of technology and integration of different monitoring methods, water resources can be more effectively protected, improving environmental quality.

The main objectives of this paper are as follows: (1) to establish a new hybrid model by combining the newly fused attention mechanism, custom residual blocks, and convolutional neural networks; discuss the applicability of this model structure in the prediction of different water quality parameters; and compare and analyze with other neural network models to verify its effectiveness. (2) The same model was used to invert turbidity, permanganate, ammonia nitrogen, and dissolved oxygen, and then R², RMSE, MAPE, and RPD were used as evaluation bases to evaluate the effect of water quality parameter inversion by different models. (3) The spatial variation maps of four water quality parameters were drawn to comprehensively evaluate the water quality of the Yellow River basin in Ningxia, China.

2. Materials and Methods

The content of the following section is the core component of this paper. Firstly, the overall situation in the research area is introduced, and then the preprocessing of the downloaded remote sensing images and collected water quality data is described. Lastly, the design details of the PCWA-ResCNN, CNN, and LSTM models established in this study are empirically introduced.

2.1. Research Area

In the following study, the Yellow River basin of Ningxia, China (104°36′–106°55′ E, 37°15′–39°23′ N), was selected as the research area (Figure 1). Located in northwest China, the 317 km long river ends at the Mahuang Valley in Shizuishan City, a length that accounts for almost one-fourteenth of the entire Yellow River. The river flows through most parts of the Ningxia Hui Autonomous Region, including the cities of Yinchuan, Shizuishan, Wuzhong, and Zhongwei. The Yellow River in Ningxia comprises a long stretch, gentle river, wide riverbed, and abundant water resources. The main land types include cultivated land, forest land, and urban construction land, among which cultivated land and urban construction land are the main land types. The Yellow River basin in Ningxia has a temperate continental climate. Winters in the basin are cold and dry, water in the area may be frozen, and precipitation is rare and unevenly distributed. However, with the development of the economy and the increase in the population, the Yellow River basin in Ningxia region also faces various environmental problems. The effects of overexploitation and pollution of water resources, land degradation, and deterioration of the ecological environment have become increasingly prominent, posing challenges to local sustainable development [24,25]. Therefore, protecting the ecological environment of the Yellow River basin and promoting the rational utilization and protection of resources have become one of the leading issues in the development of this region.

2.2. Data and Processing

2.2.1. Water Quality Monitoring Data

All monitoring data for the Yellow River basin in Ningxia came from the surface water quality monitoring system, which was released to the public in 2014. The water quality measurement method is fully described in the Technical Specification for Automatic Monitoring of Surface Water (Standard No. HJ 915-2017) [26] issued by the Ministry of Ecology and Environment of the People’s Republic of China, which was implemented on 1 April 2018.

In the present study, a total of 9 monitoring sections were selected in the Yellow River basin of Ningxia (Table 1). The dissolved oxygen, permanganate, ammonia nitrogen, and turbidity data from 12 May 2021 to 15 September 2023 were selected, and the water quality index content was determined using an electrode method in each section. Considering the local climate throughout the year, lake ice in the winter affects the reflectivity of remote sensing bands, and as such, water quality data cannot be collected. Therefore, remote sensing images from April to November for each year of the study period were selected, which spans the warmer period of the year. The acquisition time of Landsat-8 OLI_TIRS satellite images was matched with the acquisition time of water quality monitoring section data, and the remote sensing image band where the water quality monitoring station was located was extracted and matched with the water quality data.

2.2.2. Remote Sensing Image Data

In the present study, Landsat-8 satellite data were selected as the data source. The aforementioned satellite was successfully launched by NASA on 11 February 2013, and it is one of the most commonly used satellites for water quality inversion and ground object recognition [27]. Landsat-8 carries two sensors covering a total of 11 bands (Table 2), providing global satellite data every 16 days [28,29]. The Landsat-8 Land Imager (OLI) dataset was obtained for the period of May 2021 to September 2023 via the Geospatial Data cloud platform. A total of 68 OLI images with less than 10% cloud cover were retrieved. These images provide high-resolution spatial data for water quality monitoring and facilitate accurate water quality inversion and dynamic change analysis.

2.2.3. Remote Sensing Image Preprocessing

Landsat-8 was used as the source of image data in the study presented herein. In the process of image preparation, clouds and shadows usually cause interference, thus affecting the observation and analysis of surface features [30]. In order to eliminate the effects of light conditions at different times and places, the digital meter values must be converted to radiation values. In addition, atmospheric absorption and scattering can cause energy loss and color changes in the image, further reducing the accuracy of the image. Therefore, radiometric calibration and atmospheric correction pre-processing steps are adopted to improve image quality. For atmospheric correction, the FLAASH method in ENVI 5.6 software was used in the present study to select an appropriate atmospheric model and aerosol model, and the 2-band (K-T) method was used for aerosol detection. This method estimates the aerosol optical thickness (AOT) and particle size distribution by analyzing the difference in radiation brightness between the short-wave infrared band (K band) and the thermal infrared band (T band), so as to realize atmospheric correction processing.

2.3. Methods

2.3.1. PCWA-ResCNN Model Construction

In the present paper, we propose a new PCWA (Position and Channel Wise Attention) mechanism. This mechanism combines the advantages of channel attention and location attention, as shown in Formulas (1)–(4). This mechanism can focus on the importance of both the channel dimension and spatial dimension of features in data prediction in order to capture key information more effectively and select and enhance important features in a targeted manner. In the PCWA attention mechanism, channel attention is used to add weighting to the channel dimensions of processing features to highlight the importance of different channels. Location attention, in contrast, reinforces the importance of different locations by adding weights to the spatial dimensions of features. Through this fusion, PCWA is able to more effectively capture key information in the data and select and enhance important features in a targeted manner, thereby improving the accuracy and efficiency of data predictions.

(1)$C_{att}\left(c\right)={\displaystyle \sum _{i=1}^H{\displaystyle \sum _{j=1}^{W}X}}\left(i,j,c\right)$

(2)$P_{att}\left(i,j\right)=\underset{c}{\max }X\left(i,j,c\right)$

(3)$A\left(i,j,c\right)=C_{att}\left(c\right)\cdot P_{att}\left(i,j\right)$

(4)$Z\left(i,j\right)=X\left(i,j,c\right)\cdot \left({\displaystyle \sum _{i=1}^H{\displaystyle \sum _{j=1}^{W}X}}\left(i,j,c\right)\cdot \underset{c}{\max }X\left(i,j,c\right)\right)$

In the formula, $H$ and $W$ represent the height and width of the feature graph, $c$ is the number of channels, $C_{att}\left(c\right)$ represents the attention weight of the $c$ channel, $P_{att}\left(i,j\right)$ represents the attention weight of the position $\left(i,j\right)$, and $A\left(i,j,c\right)$ and $Z\left(i,j,c\right)$ represent combined and applied attention, respectively.

In addition, a custom residual block is introduced to enhance the performance of traditional convolutional neural networks (CNNs). The residual block alleviates the problem of gradient disappearance in deep networks by introducing skip connections, and it improves the expressiveness and robustness of the model through feature reuse. Combined with the PCWA attention mechanism, the network performs well in feature extraction and selection, it is better able to capture complex features and patterns, and it has greater robustness and generalization ability.

By combining the PCWA attention mechanism and residual block with the CNN, a new network structure named PCWA-ResCNN is established (Figure 2). With these improvements, the PCWA-ResCNN model can not only extract important features between observations and surface reflectance but also effectively select and reuse these features, thus significantly improving the overall prediction performance. The improvement of this structure induces the network to perform better when handling complex tasks, and it provides strong support for research and application in related fields.

The loss function $L_{loss}$ of this model is

(5)$L_{loss}={\displaystyle \frac{1}{2}}{\displaystyle \sum _{i=1}^R{(t_i-y_i)}^2}$

In the formula, $t_i$ is the target output, $y_i$ is the predicted network value, and $R$ is the number of responses.

2.3.2. CNN Model Construction

A convolutional neural network (CNN) can effectively process high-dimensional data such as images, and it has the advantage of local acceptance domain and weight sharing, which greatly reduces training costs and makes it able to mine data fluctuation features more efficiently. Therefore, the CNN structure used in the present study is 7-(LY1-LY2-LY3)-2-1, which consists of 7 input bands, 3 hidden layers, 2 fully connected layers, and 1 output layer. Convolution layers of 16, 32, and 64 filters (the convolution kernel size of each layer is 2 × 1) are used to capture the feature information of different levels. Simultaneously, the RMSprop optimizer is adopted, whose advantages include its self-adaptive learning rate, robustness to sparse gradients, fast convergence, etc., and it can solve the problem in that the learning rate in SGD cannot be adjusted adaptively.

2.3.3. LSTM Model Construction

A long short-term memory neural network (LSTM) can learn feature representation directly from the original input sequence through end-to-end learning, avoiding the complexity of manual feature extraction using traditional methods. This feature makes LSTM perform well in time series data processing. The LSTM model structure in the present study includes a sequence input layer, an LSTM layer, a ReLU activation layer, a dropout layer, and a full-connection layer. The sequence input layer processes data for seven features. The LSTM layer comprises a total of 256 units, which only output the final state of the sequence. The ReLU activation layer and two fully connected layers (128 and 64 neurons, respectively) further process the data, adding nonlinearity. The output layer consists of a single neuron and is suitable for regression tasks.

2.4. Correlation Analysis

In the following study, after preprocessing the Landsat-8 remote sensing images, the band reflectance data of the monitoring points were extracted. In order to explore the relationship between the remote sensing data and water quality parameters, Pearson correlation analysis was performed, and the results are shown in Figure 3. The results show that there was a strong correlation between TUB and B5. CODMn showed a strong correlation with both B4 and B5; however, the correlation with B5 was more prominent. NH₃-N is strongly correlated with B1 to B4, especially with B2. DO is strongly correlated with B3, B6, and B7, and the correlation with B7 is the greatest.

Based on the above findings, we selected all parameters, B1–B7, of the pre-processed Landsat-8 remote sensing images as input features for the construction and training of three models, PCWA-ResCNN, CNN, and LSTM, all of which adopted the structural design of seven inputs and single outputs. From the collected data, a total of 170 sample points were extracted, of which 136 (80%) were used for training the model and the remaining 34 (20%) were used for validating the model.

2.5. Statistical Evaluation of Indicators

In the present study, the coefficient of determination (R²), root mean square error (RMSE), mean absolute percentage error (MAPE), and residual prediction deviation ratio (RPD) were used to evaluate the various models. The calculation Formulas (6)–(9) of each indicator are shown below:

(6)$R^2=1-{\displaystyle \frac{{\displaystyle \sum _{i=1}^n{\left({\widehat{y}}_i-y_i\right)}^2}}{{\displaystyle \sum _{i=1}^n{\left({\overline{y}}_i-y_i\right)}^2}}}$

(7)$RMSE=\sqrt{{\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n{\left({\widehat{y}}_i-y_i\right)}^2}}$

(8)$MAPE={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n\left|\frac{{\widehat{y}}_i-y_i}{y_i}\right|}$

(9)$RPD={\displaystyle \frac{SD}{RMSE}}$

where $\widehat{y}$ is the predicted value of the sample, $\overline{y}$ is the mean value of the measured sample, $y$ is the measured sample value, $n$ is the number of samples, expressed by $i$ = 1, 2, ∙∙∙, n, and $SD$ is the standard deviation. Models with high R² and RPD and low RMSE and MAPE are considered to be suitable for quantitative inversion.

3. Results

In the following section, the study results are presented, including the correlation analysis, summary, and graphical visualization of the prediction results of water quality parameters. Lastly, four types of water quality parameters are inverted to generate spatial distribution feature maps, and through a comparison with the results of previous studies, the changes in water quality in the basin in recent years are evaluated.

3.1. Prediction Results and Analysis of Water Quality Parameters

As can be seen from Table 3, the PCWA-ResCNN model has the best fitting characteristics, and the R² of TUB, CODMn, NH₃-N, and DO parameters are 97.92%, 95.29%, 95.01%, and 96.12%, respectively, showing satisfactory prediction ability. Compared with the traditional CNN model, the prediction accuracy of PCWA-ResCNN improved by 4.16%, 8%, 8.58%, and 7.09%, respectively. The LSTM model was not effective in predicting TUB, CODMn, and NH₃-N, with R² of 84.99%, 77.9%, and 68.42%, respectively. However, the prediction effect of the DO parameter was poor, and the R² value was only 33.87%. The experimental results show that the LSTM model is not suitable for predicting DO parameters in the Yellow River basin of Ningxia. The change in water quality in the Yellow River basin involves multi-scale and nonlinear factors. The LSTM model mainly focuses on the long-term and short-term dependence of time series data, and the processing ability of these multi-scale and nonlinear factors is limited. The mixed model of deep learning can be used to consider the multi-dimensional characteristics and complexity of the data to improve the prediction effect.

In terms of RMSE, MAPE, and RPD, the PCWA-ResCNN model has excellent performance on all parameters. This fact is reflected in its smaller RMSE and MAPE, in addition to its higher RPD. These experimental results further validate the superiority of the PCWA-ResCNN model in water quality parameter prediction.

The scatter diagram of observed and predicted values of water quality parameters by the PCWA-ResCNN, CNN, and LSTM models is shown in Figure 4. These scatter plots clearly show the relationship between the predicted and observed values of each model. By observing these scatter plots, it can be seen that the PCWA-ResCNN model shows excellent performance in predicting water quality parameters, regardless of the R², RMSE, MAPE, or RPD indicators, which is far superior to the traditional CNN model and LSTM model. The PCWA-ResCNN model performs well in processing complex data structures and feature extraction. The R² values of the four water quality parameters of TUB, CODMn, NH₃-N, and DO all exceed 0.95, showing strong robustness and an excellent feature extraction ability. The model can capture local and global information effectively, and it has a satisfactory spatial feature learning ability. Therefore, it is recommended to use the PCWA-ResCNN model to predict and analyze water quality parameters in practice in order to obtain more reliable and accurate results.

3.2. Spatiotemporal Analysis of Water Quality

According to the “Surface water Environmental Quality Standards” (GB3838-2002) [31] that establish index evaluation standards, as shown in Table 4, and according to the surface water environmental functions and protection objectives, Class I water is suitable for source water and nature reserves, Class II water is suitable for primary drinking water sources, class III water is suitable for secondary drinking water sources, and class IV water is suitable for industrial water. Class V water is suitable for agricultural water use. Turbidity is not included in water quality assessment because there is no corresponding standard limit value, but it can be used as a reference index to judge whether water quality is affected by sediment and salinity. High NH₃-N levels indicate water pollution, which is mainly caused by agricultural fertilizers and domestic sewage. CODMn indicates organic content; high DO levels indicate the health of the water body and support the survival of aquatic life. In Figure 5, the spatial variation characteristics of water quality inversion for TUB, CODMn, NH₃-N, and DO are introduced.

As shown in Figure 5, in the water quality inversion diagram of the Yellow River basin in Ningxia, it was found that the TUB content in the middle reaches of the Yellow River basin in Ningxia has increased significantly, which is primarily due to the increase in sediment content caused by soil erosion. As such, soil and water control in this region still need to be strengthened. In the entire basin, the CODMn concentration in the upper and lower reaches meets the Class II water quality standards, and that in the middle reaches meets the Class III and Class IV water quality standards, indicating that the impact of CODMn pollution is relatively serious. The average content of NH₃-N reached the Class I water quality standard, and, in some areas, it reached the Class II water quality standard. In general, the influence of NH₃-N on the Yellow River basin in Ningxia was minor. The average concentration of DO content in the entire basin meets the water quality standards of Classes I and II. According to the inversion maps of the four water quality parameters, the water quality in the upper and lower reaches of the Yellow River basin in Ningxia is satisfactory; in comparison, the water quality in the middle reaches is poor, but the overall water quality is satisfactory, mainly constituting Class I and Class II water quality. These findings are consistent with the actual information on surface water environmental quality released by the local environment office.

From 2004 to 2018, the pollution of the Yellow River basin in Ningxia was relatively serious, the main pollution indicators were NH₃-N and CODMn, and the rainy season and dry season were dominant, but since 2010, the water quality in the Yellow River basin has gradually improved [32]. At present, the average erosion intensity of the land in the Yellow River basin still reaches 178 tons per square kilometer per year, and the turbidity content is relatively high. However, over the past 10 years, a series of measures have been taken, such as afforestation and grassland restoration, to increase vegetation coverage, reduce soil erosion by a considerable degree, and reduce water pollution. Compared with ten years ago, CODMn has decreased by about 40% and NH₃-N content has decreased by about 50% in the Yellow River basin [33]. The results of the spatial change analysis show that the water quality in the Yellow River basin has improved significantly in recent years; however, more attention needs to be paid to the water quality in the middle reaches, monitoring and control need to be strengthened, and comprehensive and long-term management measures need to be taken to achieve continuous improvement of water quality and sustainable development of the water environment in the Yellow River basin.

4. Discussion

4.1. Advantages

In recent years, the CNN model has been significantly improved and proven to be reliable. In the present study, the same model was used to predict four water quality parameters simultaneously, and the PCWA-ResCNN model has a high accuracy in predicting the concentration of water quality parameters, namely, TUB, CODMn, NH₃-N, TUB, CODMN, and NH₃-N. The R² values of DO were 0.97, 0.95, 0.95, and 0.96, respectively (Figure 4), indicating that this technology has the potential of high-precision prediction in solving water environmental problems and showing strong generalization ability in the inversion of different water quality parameters, which can be applied in the water quality inversion of major and small rivers, lakes, and reservoirs. Due to the accuracy and reliability of deep learning methods, an increasing number of scholars are making use of SVR, RF, and neural network models to predict water quality parameters. For example, Maier et al. [34,35] used CNN models of different dimensions to simulate and train Chl-a concentration, and they achieved high-precision results. However, only a few water quality parameters were selected; thus, further optimization of the model and its application in the prediction of other water quality parameters still requires further research and exploration. Yang et al. [15] also conducted a similar study. Compared with their study, we increased the input dimension and introduced a custom attention mechanism and residual block to achieve a higher accuracy. Compared with other models, our model has high robustness and can maintain stable performance even when faced with data changes, noise interference, and system changes.

The authors of some studies have reported the problem of gradient disappearance [36,37]. In order to address this challenge, we can introduce more complex attention mechanisms, which can not only effectively alleviate the problem of gradient disappearance but also enhance the model’s learning ability for important features and context information. By ensuring the steady flow of gradients between the layers of the network, it is also expected to significantly improve the accuracy and generalization ability of spatio-temporal prediction models in practical applications. Therefore, we focused on designing a more robust and efficient neural network architecture to overcome the challenges brought by the gradient disappearance problem and promote the further application and development of deep learning in complex data analysis and prediction tasks. The prediction results of the PCWA-ResCNN model proposed herein (Figure 2) also demonstrate the potential of this model, which has been widely applied in the inversion of water quality in the Yellow River basin.

4.2. Limitations

Although the PCWA-ResCNN model performs well in retrieving water quality parameters of the Yellow River in Ningxia, there are still some limitations to the present study. First of all, for the study area referenced herein, the sampling points were far from one another and the Yellow River flow rate was fast. Even if the prediction accuracy was high, it would be easy for the important characteristics of water quality between the sampling points to be lost, which would affect the overall evaluation of water quality. Due to the wide water areas such as lakes and reservoirs, the water flow is usually slow or static; therefore, the PCWA-ResCNN model has great application potential in these waters. Secondly, the currently available satellite data are limited, and remote sensing data with a high spatial resolution, high temporal resolution, and wide spectral range are relatively scarce [38,39]. The acquisition of remote sensing data usually requires a certain time interval, and the required high temporal resolution data may not be obtained in time, which affects the real-time and effective inversion of water quality parameters. The impact of climate and environmental factors (such as rainfall, temperature change, and pollution source change) is difficult to fully capture and quantify in the model, which may lead to differences in performance under different climatic conditions [40].

In addition, the PCWA-ResCNN model has a high computational complexity and high resource requirements, which increases the difficulty of deployment and application, especially in resource-constrained environments [41]. The “black box” nature of deep learning models limits their interpretability, posing challenges to understanding the mechanisms of water quality change and developing management strategies [42]. Lastly, the adaptability of the model in different regions and time periods needs to be further verified. Differences in geographical and hydrological characteristics in different regions may lead to inconsistent model performance.

5. Conclusions

In the study presented herein, remote sensing data sets were established using Landsat-8 data and water quality monitoring data of the Yellow River basin in Ningxia, a new deep learning PCWA-ResCNN model was developed, and multiple indicators were used to test the architecture to verify its performance in water quality inversion. The same data set was compared with the traditional CNN and LSTM models. The results of water quality inversion show that the PCWA-ResCNN model has the ability to perform high-precision inversion in all water quality parameters and has more significant advantages than traditional CNN and LSTM models. In addition, the spatial variation map of water quality inversion shows that the water quality of the Ningxia Yellow River basin is mainly Class I and Class II, which accords with the actual situation of the study area. Therefore, the PCWA-ResCNN model structure can be used as a low-cost and effective method to estimate the concentration of conventional pollutants and water quality parameters in the Yellow River basin of Ningxia and can also be applied to the inversion of water quality parameters in any water quality sampling area. It is suggested that future research should seek to enhance simulations of climatic and environmental factors by expanding data sets, optimizing calculations, interpreting models, and ensuring adaptability across regions and time periods. This will enhance the application breadth and reliability of the model to better meet the needs of water quality monitoring in different regions and different time periods.

Author Contributions

Conceptualization, Q.L.; methodology, Q.L.; software, Q.L.; validation, Z.G., J.L., and X.L.; formal analysis, Z.G.; investigation, Q.L. and B.B.; resources, Q.L.; data curation, Q.L.; writing—original draft preparation, Q.L.; writing—review and editing, Z.G., J.L., B.B., and X.L.; visualization, Q.L. All authors have read and agreed to the published version of the manuscript.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The original contributions presented in the study are included in the article, further inquiries can be directed to the corresponding author.

Acknowledgments

The remote sensing data sets referenced herein were provided free of charge by the geospatial data cloud platform.

Conflicts of Interest

The authors declare no conflicts of interest.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1. Map of sampling sites in the Ningxia Yellow River basin, China.

Figure 2. PCWA-ResCNN model structure.

Figure 3. Pearson correlation between water quality parameters and each band.

Figure 4. Scatter diagram of prediction using the water quality inversion model.

Figure 5. Spatial variation characteristics of water quality parameters.

Table 1

Monitoring section information of the Yellow River basin in Ningxia.

Sampling Point No.	Monitoring Section	Longitude	Latitude
1	Mahuang Ditch	106°44′53″	39°22′35″
2	Sand Lake	106°20′52″	38°49′20″
3	Pingluo Yellow River Bridge	106°39′58″	38°48′46″
4	Yingu Highway Bridge	106°24′40″	38°21′33″
5	Yazidang Reservoir	106°32′12″	38°09′07″
6	Yesheng Highway Bridge	106°12′58″	38°08′14″
7	Jinsha Bay	105°55′34″	37°49′51″
8	Xiangshan Lake	105°12′01″	37°29′25″
9	Zhongwei Xiaheyan	105°07′27″	37°29′11″

Table 2

Landsat 8 band introduction.

Sensor	Wave Band	Wavelength Range/μm	Spatial Resolution/m
OLI	B1-Coastal	0.43–0.45	30
OLI	B2-Blue	0.45–0.51	30
OLI	B3-Green	0.53–0.59	30
OLI	B4-Red	0.64–0.67	30
OLI	B5-NIR	0.85–0.88	30
OLI	B6-SWIR1	1.57–1.65	30
OLI	B7-SWIR2	2.11–2.29	30
OLI	B8-PAN	0.50–0.68	15
OLI	B9-Cirrus	1.36–1.38	30
TIRS	B10-TIRS1	10.60–11.19	100
TIRS	B11-TIRS2	11.50–12.51	100

Table 3

Precision analysis statistics of PCWA-ResCNN, CNN, and LSTM models.

Water Quality Parameter	Model	${\mathbf{R}}^2\uparrow$	$\mathbf{RMSE}\downarrow$	$\mathbf{MAPE}\downarrow$	$\mathbf{RPD}\uparrow$
TUB (NTU)	PCWA-ResCNN	0.9792	80.4109 NTU	2.5953	7.0347
	CNN	0.9376	139.2386 NTU	3.1927	4.0626
	LSTM	0.8499	215.8538 NTU	3.4165	2.6206
CODMn (mg/L)	PCWA-ResCNN	0.9529	0.3005 mg/L	0.0876	4.6780
	CNN	0.8729	0.4938 mg/L	0.1549	2.8474
	LSTM	0.7790	0.6512 mg/L	0.1941	2.1591
NH3-N (mg/L)	PCWA-ResCNN	0.9501	0.0090 mg/L	0.2291	4.5441
	CNN	0.8643	0.0149 mg/L	0.3120	2.7550
	LSTM	0.6842	0.0227 mg/L	0.5212	1.8063
DO (mg/L)	PCWA-ResCNN	0.9612	0.1997 mg/L	0.0188	5.1517
	CNN	0.8903	0.3356 mg/L	0.0369	3.0651
	LSTM	0.3387	0.8241 mg/L	0.0897	1.2482

Table 4

Environmental quality standard limits for surface water.

Evaluation Factor	Evaluation Level
	I	II	III	IV	V
CODMn (mg/L)≤	2	4	6	10	15
NH3-N (mg/L)≤	0.15	0.5	1	1.5	2
DO (mg/L)≥	7.5	6	5	3	2

References

1. Galgani, L.; Loiselle, S.A. Plastic pollution impacts on marine carbon biogeochemistry. Environ. Pollut.; 2021; 268, 115598. [DOI: https://dx.doi.org/10.1016/j.envpol.2020.115598] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33158618]

2. Olaka, L.A.; Ogutu, J.O.; Said, M.Y.; Oludhe, C. Projected climatic and hydrologic changes to Lake Victoria Basin Rivers under three RCP emission scenarios for 2015–2100 and impacts on the water sector. Water; 2019; 11, 1449. [DOI: https://dx.doi.org/10.3390/w11071449]

3. Yang, W.; Yao, J.; He, Y.; Huang, Y.; Liu, H.; Zhi, Y.; Qian, S.; Yan, X.; Jian, S.; Li, W. Nitrogen removal enhanced by benthic bioturbation coupled with biofilm formation: A new strategy to alleviate freshwater eutrophication. J. Environ. Manag.; 2021; 292, 112814. [DOI: https://dx.doi.org/10.1016/j.jenvman.2021.112814] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34030016]

4. Moss, B. Cogs in the endless machine: Lakes, climate change and nutrient cycles: A review. Sci. Total Environ.; 2012; 434, pp. 130-142. [DOI: https://dx.doi.org/10.1016/j.scitotenv.2011.07.069] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/21962562]

5. Yin, D.; Wang, L.; Zhu, Z.; Clark, S.S.; Cao, Y.; Besek, J.; Dai, N. Water quality related to Conservation Reserve Program (CRP) and cropland areas: Evidence from multi-temporal remote sensing. Int. J. Appl. Earth Obs. Geoinf.; 2021; 96, 102272. [DOI: https://dx.doi.org/10.1016/j.jag.2020.102272]

6. Schaeffer, B.A.; Schaeffer, K.G.; Keith, D.; Lunetta, R.S.; Conmy, R.; Gould, R.W. Barriers to adopting satellite remote sensing for water quality management. Int. J. Remote Sens.; 2013; 34, pp. 7534-7544. [DOI: https://dx.doi.org/10.1080/01431161.2013.823524]

7. Liu, P.; Wang, J.; Sangaiah, A.K.; Xie, Y.; Yin, X. Analysis and prediction of water quality using LSTM deep neural networks in IoT environment. Sustainability; 2019; 11, 2058. [DOI: https://dx.doi.org/10.3390/su11072058]

8. Cui, Y.; Yan, Z.; Wang, J.; Hao, S.; Liu, Y. Deep learning–based remote sensing estimation of water transparency in shallow lakes by combining Landsat 8 and Sentinel 2 images. Environ. Sci. Pollut. Res.; 2022; 29, pp. 4401-4413. [DOI: https://dx.doi.org/10.1007/s11356-021-16004-9]

9. He, Y.; Gong, Z.; Zheng, Y.; Zhang, Y. Inland reservoir water quality inversion and eutrophication evaluation using BP neural network and remote sensing imagery: A case study of Dashahe reservoir. Water; 2021; 13, 2844. [DOI: https://dx.doi.org/10.3390/w13202844]

10. Yang, S.; Zhong, S.; Chen, K. W-WaveNet: A multi-site water quality prediction model incorporating adaptive graph convolution and CNN-LSTM. PLoS ONE; 2024; 19, e0276155. [DOI: https://dx.doi.org/10.1371/journal.pone.0276155]

11. Ni, Q.; Cao, X.; Tan, C.; Peng, W.; Kang, X. An improved graph convolutional network with feature and temporal attention for multivariate water quality prediction. Environ. Sci. Pollut. Res.; 2023; 30, pp. 11516-11529. [DOI: https://dx.doi.org/10.1007/s11356-022-22719-0] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36094707]

12. Syariz, M.A.; Lin, C.H.; Nguyen, M.V.; Jaelani, L.M.; Blanco, A.C. WaterNet: A convolutional neural network for chlorophyll-a concentration retrieval. Remote Sens.; 2020; 12, 1966. [DOI: https://dx.doi.org/10.3390/rs12121966]

13. Sagan, V.; Peterson, K.T.; Maimaitijiang, M.; Sidike, P.; Sloan, J.; Greeling, B.A.; Maalouf, S.; Adams, C. Monitoring inland water quality using remote sensing: Potential and limitations of spectral indices, bio-optical simulations, machine learning, and cloud computing. Earth-Sci. Rev.; 2020; 205, 103187. [DOI: https://dx.doi.org/10.1016/j.earscirev.2020.103187]

14. Zhao, Y.; Yu, T.; Hu, B.; Zhang, Z.; Liu, Y.; Liu, X.; Liu, H.; Liu, J.; Wang, X.; Song, S. Retrieval of water quality parameters based on near-surface remote sensing and machine learning algorithm. Remote Sens.; 2022; 14, 5305. [DOI: https://dx.doi.org/10.3390/rs14215305]

15. Yang, H.; Du, Y.; Zhao, H.; Chen, F. Water quality Chl-a inversion based on spatio-temporal fusion and convolutional neural network. Remote Sens.; 2022; 14, 1267. [DOI: https://dx.doi.org/10.3390/rs14051267]

16. Harvey, E.T.; Kratzer, S.; Philipson, P. Satellite-based water quality monitoring for improved spatial and temporal retrieval of chlorophyll-a in coastal waters. Remote Sens. Environ.; 2015; 158, pp. 417-430. [DOI: https://dx.doi.org/10.1016/j.rse.2014.11.017]

17. Chu, H.J.; He, Y.C. Remote sensing water quality inversion using sparse representation: Chlorophyll-a retrieval from Sentinel-2 MSI data. Remote Sens. Appl. Soc. Environ.; 2023; 31, 101006. [DOI: https://dx.doi.org/10.1016/j.rsase.2023.101006]

18. Xue, Y.; Zhu, L.; Zou, B.; Wen, Y.-M.; Long, Y.-H.; Zhou, S.-L. Research on inversion mechanism of chlorophyll—A concentration in water bodies using a Convolutional Neural Network model. Water; 2021; 13, 664. [DOI: https://dx.doi.org/10.3390/w13050664]

19. Zhang, H.; Xue, B.; Wang, G.; Zhang, X.; Zhang, Q. Deep learning-based water quality retrieval in an impounded lake using landsat 8 imagery: An application in Dongping lake. Remote Sens.; 2022; 14, 4505. [DOI: https://dx.doi.org/10.3390/rs14184505]

20. Parra, L.; Ahmad, A.; Sendra, S.; Lloret, J.; Lorenz, P. Combination of Machine Learning and RGB Sensors to Quantify and Classify Water Turbidity. Chemosensors; 2024; 12, 34. [DOI: https://dx.doi.org/10.3390/chemosensors12030034]

21. Kuang, L.; Shi, P.; Hua, C.; Chen, B.; Zhu, H. An enhanced extreme learning machine for dissolved oxygen prediction in wireless sensor networks. IEEE Access; 2020; 8, pp. 198730-198739. [DOI: https://dx.doi.org/10.1109/ACCESS.2020.3033455]

22. Ren, J.; Zhu, Q.; Wang, C. Edge Computing for Water Quality Monitoring Systems. Mob. Inf. Syst.; 2022; 2022, 5056606. [DOI: https://dx.doi.org/10.1155/2022/5056606]

23. Yigit Avdan, Z.; Kaplan, G.; Goncu, S.; Avdan, U. Monitoring the water quality of small water bodies using high-resolution remote sensing data. ISPRS Int. J. Geo-Inf.; 2019; 8, 553. [DOI: https://dx.doi.org/10.3390/ijgi8120553]

24. Brierley, G.J.; Han, M.; Li, X.; Li, Z.; Huang, H.Q. Geo-eco-hydrology of the Upper Yellow River. Wiley Interdiscip. Rev. Water; 2022; 9, e1587. [DOI: https://dx.doi.org/10.1002/wat2.1587]

25. Wu, N.; Liu, S.M.; Zhang, G.L.; Zhang, H.M. Anthropogenic impacts on nutrient variability in the lower Yellow River. Sci. Total Environ.; 2021; 755, 142488. [DOI: https://dx.doi.org/10.1016/j.scitotenv.2020.142488]

26. HJ 915-2017 Technical Specifications for Automatic Monitoring of Surface Water; People’s Republic of China: Beijing, China, 2018.

27. Kuhn, C.; de Matos Valerio, A.; Ward, N.; Loken, L.; Sawakuchi, H.O.; Kampel, M.; Richey, J.; Stadler, P.; Crawford, J.; Striegl, R. et al. Performance of Landsat-8 and Sentinel-2 surface reflectance products for river remote sensing retrievals of chlorophyll-a and turbidity. Remote Sens. Environ.; 2019; 224, pp. 104-118. [DOI: https://dx.doi.org/10.1016/j.rse.2019.01.023]

28. Wu, H.; Guo, Q.; Zang, J.; Qiao, Y.; Zhu, L.; He, Y. Study on Water Quality Parameter Inversion based on Landsat 8 and Measured Data. Remote Sens. Technol. Appl.; 2021; 36, pp. 898-907.

29. Holden, C.E.; Woodcock, C.E. An analysis of Landsat 7 and Landsat 8 underflight data and the implications for time series investigations. Remote Sens. Environ.; 2016; 185, pp. 16-36. [DOI: https://dx.doi.org/10.1016/j.rse.2016.02.052]

30. Guo, J.; Yang, J.; Yue, H.; Hou, C.; Li, K. Landsat-8 OLI multispectral image dehazing based on optimized atmospheric scattering model. IEEE Trans. Geosci. Remote Sens.; 2020; 59, pp. 10255-10265. [DOI: https://dx.doi.org/10.1109/TGRS.2020.3036972]

31. GB 3838-2002 Environmental quality standards for surface water; People’s Republic of China: Beijing, China, 2002.

32. Liu, Y.L.; Zheng, Y.A. Water quality assessment and spatial-temporal variation analysis in Yellow River Basin. Huanjing Kexue; 2022; 43, pp. 1332-1345.

33. Quan, J.; Xu, Y.; Ma, T.; Wilson, J.P.; Zhao, N.; Ni, Y. Improving surface water quality of the Yellow River Basin due to anthropogenic changes. Sci. Total Environ.; 2022; 836, 155607. [DOI: https://dx.doi.org/10.1016/j.scitotenv.2022.155607] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35500708]

34. Maier, P.M.; Keller, S.; Hinz, S. Deep learning with WASI simulation data for estimating chlorophyll a concentration of inland water bodies. Remote Sens.; 2021; 13, 718. [DOI: https://dx.doi.org/10.3390/rs13040718]

35. Aptoula, E.; Ariman, S. Chlorophyll-a retrieval from sentinel-2 images using convolutional neural network regression. IEEE Geosci. Remote Sens. Lett.; 2021; 19, 6002605. [DOI: https://dx.doi.org/10.1109/LGRS.2021.3070437]

36. Miao, Q.; Pan, B.; Wang, H.; Hsu, K.; Sorooshian, S. Improving monsoon precipitation prediction using combined convolutional and long short term memory neural network. Water; 2019; 11, 977. [DOI: https://dx.doi.org/10.3390/w11050977]

37. Tran, H.; Leonarduzzi, E.; De la Fuente, L.; Hull, R.B.; Bansal, V.; Chennault, C.; Gentine, P.; Melchior, P.; Condon, L.E.; Maxwell, R.M. Development of a deep learning emulator for a distributed groundwater–surface water model: ParFlow-ML. Water; 2021; 13, 3393. [DOI: https://dx.doi.org/10.3390/w13233393]

38. Xiao, Z.-H.; Wang, D.-M.; Wen, J.; Famg, H.-B.; Hu, Y.-X.; Xu, N. Overview of the ground application system of satellite-aviation-ground remote sensing data at home and abroad. Chin. J. Geomech.; 2015; 21, pp. 117-128.

39. Liu, P. A survey of remote-sensing big data. Front. Environ. Sci.; 2015; 3, 45. [DOI: https://dx.doi.org/10.3389/fenvs.2015.00045]

40. Farahani, S.D.; Farahani, A.D. Machine learning models for predicting the performance of solar-geothermal desalination in different meteorological conditions. Ain Shams Eng. J.; 2024; 15, 102591. [DOI: https://dx.doi.org/10.1016/j.asej.2023.102591]

41. Kamath, V.; Renuka, A. Deep learning based object detection for resource constrained devices: Systematic review, future trends and challenges ahead. Neurocomputing; 2023; 531, pp. 34-60. [DOI: https://dx.doi.org/10.1016/j.neucom.2023.02.006]

42. Oviedo, F.; Ferres, J.L.; Buonassisi, T.; Butler, K.T. Interpretable and explainable machine learning for materials science and chemistry. Acc. Mater. Res.; 2022; 3, pp. 597-607. [DOI: https://dx.doi.org/10.1021/accountsmr.1c00244]