Assessing WQI Using Spatial Land-Use Context Derived from Google Earth Imagery and Advanced Convolutional Neural Networks in South Korea

Abstract

Assessing water quality indices (WQIs) derived from physicochemical measurements accurately and efficiently is essential for effective water resource management. However, conventional monitoring approaches based on single-point measurements and limited spatial coverage face constraints in representing large-scale river environments. To address these limitations, this study integrates high-resolution Google Earth RGB imagery with national water quality monitoring data from South Korea to construct an image-based dataset for WQI estimation. Water quality monitoring records from 1762 sampling sites collected between January 2000 and September 2020 were used to calculate WQI values. The index was computed using seven parameters—temperature, pH, dissolved oxygen, total solids, biochemical oxygen demand, nitrate, and phosphate—following the standard weighting procedure. Corresponding Google Earth RGB imagery acquired within ±1 day of field measurements over the same 2000–2020 period was compiled, resulting in 34108 image–sample pairs. Based on this integrated dataset, a ResNeXt-based convolutional neural network enhanced with convolutional block attention modules was implemented and applied to estimate WQI values from spatial land-use context and river morphology captured in RGB imagery. The proposed model demonstrated superior predictive performance compared to baseline neural network models, achieving a coefficient of determination (R²) of 0.94 and an index of agreement (IOA) of 0.96. Grad-CAM analysis indicates that the model primarily utilizes spatial land-use patterns, riparian context, and river morphology rather than direct visual signals from the water surface itself. These findings suggest that RGB imagery contains spatial information related to observed WQI values. Accordingly, the framework provides a spatially continuous perspective on river conditions that may support large-scale monitoring efforts.

1. Introduction

Climate change has intensified environmental variability worldwide, increasing the importance of effective environmental management and monitoring across multiple ecosystems [1,2,3]. Among these systems, rivers are particularly sensitive to climate-driven changes, as they integrate hydrological, ecological, and anthropogenic influences. As a result, maintaining good water quality is essential for sustaining healthy river systems, and accurate assessment of water quality conditions and variability is critical for monitoring river conditions and identifying potential sources of pollution [4]. The water quality index (WQI), proposed by Horton [5] and Brown [6], is a numerical expression that assesses water quality by combining various parameters related to multiple pollutants into a single score. The National Sanitation Foundation WQI (NSFWQI) is an assessment tool developed using the Delphi technique with 142 experts to determine the weights for nine water quality indicators. However, calculating the WQI requires measuring various water quality parameters (WQPs), and currently, the concentrations of these parameters at the sampling points are primarily determined through single-point monitoring [7,8]. Although this method can accurately reflect various water quality indicators, it is time-consuming and costly and poses challenges in surveying extensive river areas [4,9]. Accurate estimation of WQI values and their spatial variability has become increasingly important for effective environmental management and sustainable water resource planning. In response to the limitations of traditional monitoring systems, recent research has increasingly focused on integrating remote sensing and artificial intelligence (AI) to enhance spatial coverage, temporal frequency, and analytical capabilities in water quality assessment. Remote sensing technology, in particular, has seen significant advances, with numerous studies utilizing satellite imagery in combination with in-situ measurements to estimate key water quality parameters (WQPs). For example, artificial neural networks (ANNs) have been applied to Landsat TM imagery for predicting chlorophyll-a, turbidity, and phosphorus concentrations [10], while random forest (RF), support vector machine (SVM), and support vector regression (SVR) have demonstrated robust performance when applied to GOCI imagery for monitoring chlorophyll-a and suspended particulate matter (SPM) [11]. Similarly, Landsat-8 data have been used to develop both back-propagation neural networks and convolutional neural networks (CNNs) for estimating a range of WQPs, including total suspended solids (TSS), chemical oxygen demand (COD), biochemical oxygen demand (BOD), and dissolved oxygen (DO), with strong predictive performance [9,12].

Efforts to generalize these approaches across broader spatial and temporal contexts have led to the application of extreme learning machines (ELMs), SVR, and linear regression (LR) models using multi-sensor satellite data—from Landsat-8 OLI, Sentinel-3 OLCI, and Sentinel-2 MSI—integrated with field data collected between 2013 and 2019 [13]. In addition, nutrient-related parameters, such as total nitrogen (TN), ammoniacal nitrogen (AN), and total phosphorus (TP), have been estimated using diverse machine learning (ML) models, including ANN, random forest regression (RFR), regression trees (RTs), and gradient boosting machines (GBMs) [14]. Deep learning-based models such as CNN, fully connected networks (FCNs), multilayer perceptrons (MLPs), recurrent neural networks (RNNs), and long short-term memory (LSTM) networks have also been employed to model physicochemical variables like electrical conductivity (EC) and DO with improved accuracy [15]. Meanwhile, UAV imagery has emerged as a high-resolution remote sensing alternative, successfully combined with models such as CatBoost, XGBoost, AdaBoost, RF, K-nearest neighbors (KNN), deep neural networks (DNNs), and LR for localized WQP analysis [4]. Furthermore, comprehensive water quality indices (WQIs) have been estimated using hybrid models such as the M5 model tree and multivariate adaptive regression spline (MARS), as demonstrated in large-scale applications, including the Hudson River [16].

Despite the remarkable progress achieved through spectrally rich remote sensing approaches, their practical implementation still faces several challenges in terms of cost, complexity, and operational constraints. While previous studies have achieved high accuracy by leveraging spectral bands from multispectral or hyperspectral satellite sensors, these methods often require access to complex and high-cost datasets and depend heavily on atmospheric correction and calibration procedures.

In contrast, this study employed freely available high-resolution Red–Green–Blue (RGB) imagery from Google Earth, which improves accessibility, interpretability, and operational feasibility in water quality monitoring. Although RGB imagery does not directly measure physicochemical water properties, it captures visible surface characteristics and surrounding landscape features that may be indirectly related to WQI variability. We, therefore, hypothesize that spatial land-use patterns and river morphology observable in RGB imagery contain sufficient information to approximate WQI values, even without direct spectral water quality signals. With the advancement of computer vision techniques, especially convolutional neural networks, RGB-based models have demonstrated strong capabilities in recognizing spatial features without relying on spectrally rich data. Building on this, the present study applies a convolutional neural network to RGB imagery in order to estimate WQI values across diverse regions in South Korea, with the objective of exploring widely accessible RGB imagery as an alternative source of spatial land-use information for WQI estimation.

2. Study Area and Data Processing

2.1. Study Area

The Republic of Korea features rivers that predominantly originate from the high mountains in the east and flow westward; over 70% of its total area comprises mountainous terrain. As of 2023, South Korea had a population of approximately 51.78 million. The rapid economic growth in the 1980s has made a major contribution to the development and population concentration in major cities, such as Seoul and Busan.

The major rivers in South Korea are the Han River, which is around 494 km in length and originates in the Taebaek Mountains, flowing through Seoul and discharging into the West Sea. It has a watershed area of approximately 26,219 km². The Nakdong River, with a length of 510 km, also begins in the Taebaek Mountains and has a watershed area of about 23,817 km². The Geum River starts in the Jirisan Mountains, flows through Chungcheong Province, and then into the West Sea, spanning about 401 km with a watershed area of around 10,032 km². Lastly, the Yeongsan River passes through the Jeolla region and drains into the Yellow Sea, covering around 130 km with a watershed area of approximately 3467 km². Figure 1 illustrates the geographical locations in this study; the 1762 red dots in the figure represent data collection points.

2.2. WQP Collection

This study used datasets from “The Open AI Dataset Project (AI-Hub, S. Korea, Dong-gu, Daegu, Republic of Korea).” All data information can be accessed through “www.aihub.or.kr (accessed on 12 June 2023).”

From January 2000 to September 2020, water quality data were collected from a network of 1762 monitoring points. The dataset included measurements of water temperature, pH, dissolved oxygen (DO), biochemical oxygen demand (BOD), chemical oxygen demand (COD), suspended solids (SS), total nitrogen (TN), ammonia nitrogen (NH3-N), nitrate nitrogen (NO3-N), total phosphorus (TP), phosphate phosphorus (PO4-P), electrical conductivity (EC), and chlorophyll-a. Although additional parameters were available in the raw monitoring records, several variables exhibited substantial missing values or limited spatial coverage across monitoring sites. Therefore, only parameters with sufficient temporal continuity and nationwide consistency were retained for statistical summarization and subsequent analysis.

The locations and timestamps of the measurements were also included.

Table 1. Statistical characteristics of WQPs measured in South Korea.

Parameter	Unit	Mean	Std.	Max.	Min.
pH	-	7.86	0.53	9.2	6.5
Chemical oxygen demand	mg/L	3.12	1.17	6.05	0.1
Biochemical oxygen demand	mg/L	1.49	0.66	3.5	1
Suspended solids	mg/L	3.81	3.51	10.25	0.1
Dissolved oxygen	mg/L	10.81	2.58	18.25	3.45
Total nitrogen	mg/L	2.04	0.82	4.22	0.02
Total phosphorus	mg/L	0.02	0.02	0.06	0.001
Temperature	°C	13.64	8	34	−0.1
EC	μS/cm	166.26	84.17	390	2.2
Ammonia nitrogen	mg/L	0.06	0.05	0.2	0.001
Nitrate nitrogen	mg/L	1.55	0.72	3.51	0.001
Chlorophyll a	mg/m3	5.11	4.9	14.6	0.02
Phosphate phosphorus	mg/L	0.008	0.007	0.02	0.0002

lists the statistical characteristics of the parameters for WQPs.

2.3. WQI Calculation

The WQI is a numerical indicator used to evaluate water quality by integrating various WQPs into a single index [17,18]. Although numerous methods have been proposed to facilitate WQI calculation [19,20,21], the NSFWQI method, developed by the National Sanitation Foundation, remains widely used [8,22,23]. The NSFWQI calculates a dimensionless WQI score ranging from 0 to 100 based on nine WQPs: temperature, pH, DO, turbidity, total solids, BOD, nitrate, phosphate, and fecal coliforms. This method uses the Delphi technique, as consented by 142 experts [24]. In this study, the WQI was calculated based on the NSFWQI framework. However, because turbidity and fecal coliform data were not available in the monitoring dataset, the index was computed using seven of the nine original parameters. The resulting index, therefore, represents a modified version of the NSFWQI rather than the original nine-parameter formulation and is hereafter referred to as the WQI to distinguish it from the standard NSFWQI. Although 19 water quality parameters were included in the overall monitoring dataset (Section 2.2), only these seven parameters were incorporated into the WQI computation. The WQI was calculated using the following formula:

$$WQI={\displaystyle \sum _{i=1}^n}S{I}_i$$

(1)

$$S{I}_i=R{W}_i\times Q_i$$

(2)

$$Q_i={\displaystyle \frac{C_i}{S_i}}\times 100$$

(3)

$$RW={\displaystyle \frac{AW}{{\sum }_{i=1}^{n}A{W}_i}}$$

(4)

The quality grade (Qi) of each parameter is obtained by dividing its measured concentration (Ci) by the corresponding standard value (Si), as expressed in Equation (3). The relative weight (RW) is calculated according to Equation (4) by normalizing the assigned weight (AW) of each parameter with respect to the total sum of assigned weights. To maintain methodological consistency, the original assigned weights were re-normalized according to Equation (4) so that the sum of relative weights equaled one, following approaches adopted in previous studies when incomplete NSFWQI parameters were reported [25]. The adjusted relative weights were derived from the original weight factors reported by Brown et al. [6] and are summarized in

Table 2. Weight factors of the National Sanitation Foundation Water Quality Index.

Parameters	Allocated Weight (AW)	Relative Weight (RW)
Turbidity	0.08	-
Biochemical oxygen demand	0.11	0.14
Dissolved oxygen	0.17	0.24
Fecal coliform	0.16	-
Nitrate	0.10	0.13
pH	0.11	0.14
Temperature	0.10	0.13
Total solids	0.07	0.09
Total phosphates	0.10	0.13

.

The exclusion of turbidity and fecal coliform has specific interpretive implications. Turbidity is closely associated with suspended sediment dynamics and is widely used as an indicator of erosion-related water quality degradation [26]. Fecal coliform serves as a standard indicator of microbiological contamination originating from domestic sewage and agricultural runoff [27]. Consequently, the modified WQI used in this study has reduced sensitivity to sediment-related disturbance and pathogenic contamination compared to the original NSFWQI, and this limitation should be considered when interpreting the results.

As outlined by Brown et al. (1970) [6], the resulting WQI scores can be categorized into five classes: Excellent (100–91), Good (90–71), Medium (70–51), Bad (50–26), and Very Bad (25–0). This classification provides a standardized reference for interpreting WQI values.

2.4. Image Data Collection

With rapid advancements in remote sensing technology, very high-resolution imagery such as WorldView-4 (WV-4, 0.3 m resolution, operated by the USA) and Google Earth Images (1 m resolution, operated by Google) has become increasingly available [28]. The Google Earth Pro program was used in this study to collect historical satellite images of water quality measurement points across South Korea. A total of 34,108 images from 1762 points, corresponding to the measurement times and locations from January 2000 to September 2020, were acquired. The image collection focused on the exact latitude and longitude of each sampling location, and only images captured within ±1 day of the field measurements were selected. The images were then collected from within radii of 500, 1000, and 2000 m of each measurement point for analysis in this study. These Google Earth images are freely available and offer access to historical satellite imagery.

2.5. Image Data Preprocessing

To ensure the quality of the RGB satellite imagery, a preprocessing procedure combining automated filtering and manual inspection was implemented. Raw imagery may contain visual artifacts such as blurriness, occlusion, or brightness distortion, which can adversely affect model performance. Therefore, such anomalies were systematically identified and removed in advance.

Blurriness was evaluated based on image sharpness using the variance of the Laplacian operator applied to the grayscale version of each image. The sharpness measure was computed as follows:

$${\mathrm{Var}}_{\mathrm{L}}=\mathrm{Var}\left({\nabla }^{2}I\right)$$

(5)

where $I$ is the grayscale image, and ${\nabla }^{2}I$ denotes the result of the Laplacian operation. An image was considered blurry if ${\mathrm{Var}}_{\mathrm{L}}$ fell below a predefined threshold, which was set to 100 in this study.

Occlusion was assessed based on the overall brightness of the image. The mean pixel intensity of the grayscale image was calculated as

$${\mu }_I={\displaystyle \frac{1}{HW}}\sum _{i=1}^H\sum _{j=1}^W{I}_{ij}$$

(6)

where $H$ and $W$ denote the height and width of the image, respectively. An image was classified as occluded if ${\mu }_I$ was less than the threshold, set to 30 in this study, indicating possible obstruction by cloud shadows or surface objects.

Images that passed these automated criteria were subsequently subjected to manual verification by the researchers. This additional step was conducted to eliminate remaining low-quality images that might not have been detected through algorithmic screening, such as those affected by seasonal variations, artificial structures, or atypical surface patterns. This dual-stage process ensured the integrity of the dataset and contributed to the robustness and generalizability of the trained models.

2.6. Image–WQI Dataset Construction

For model implementation, a dataset was constructed by integrating RGB imagery with the corresponding WQI values calculated from water quality measurements. Each monitoring station was spatially matched with its associated RGB image patch, and the calculated WQI value was assigned as the continuous regression target. In this configuration, the RGB image served as the model input, and the WQI value served as the output variable. A total of 34,108 paired image–WQI samples were used for model training and evaluation.

The dataset was divided into training (80%), validation (10%), and testing (10%) subsets at the monitoring-station level. Monitoring stations were randomly assigned to one of the three subsets, and all image–WQI samples corresponding to a given station were allocated to the same subset. This spatially structured partitioning was applied to organize the data according to monitoring locations during model development and evaluation.

3. Methodology

3.1. Soft Computing Model Implementation

This study involves the analysis of a large and complex dataset containing details about river morphology and surrounding land use. Various neural network architectures from the computer vision field, such as CNNs, residual networks (ResNets) [29], residual networks with aggregated transformations (ResNeXt) [30], LeNet [31], and the convolutional block attention module (CBAM) [32], are used to process river images and estimate the WQI. These models are proficient at handling high-dimensional data and identifying crucial environmental indicators like vegetation cover, urbanization, and water turbidity, which are vital for evaluating water quality.

Using advanced neural networks, this study created a reliable tool to predict water quality accurately from river imagery. This method reduces the need for extensive physical water sampling, allowing for more frequent and widespread monitoring of water quality across larger areas.

3.1.1. KWEN

In this study, we proposed the Korea Water Evaluate Network (KWEN) model, which integrates the ResNeXt backbone with the CBAM, to estimate WQI from complex satellite image data (Figure 2). The overall architecture of the proposed framework is summarized in Figure 2a, and the internal ResNeXt–CBAM integration used for feature extraction is illustrated in Figure 2b. The attention mechanism is further decomposed across the subsequent panels. Figure 2c presents the CBAM structure, Figure 2d describes the channel attention module, and Figure 2e illustrates the spatial attention module. This architecture has been designed to effectively capture multi-layered spatial patterns characteristic of freshwater environments by combining the parallel grouped convolutional operations of ResNeXt with the attention mechanisms of the CBAM.

The input to the KWEN model is a 224 × 224 × 3 RGB satellite image. The first convolutional layer uses 64 filters with a kernel size of 3 × 3 to extract an initial feature map. This convolution operation is described by the following equation:

$$y\left[i,j\right]=\sum _{u=0}^{F-1}\sum _{v=0}^{F-1}x\left[i+u,j+v\right]·K\left[u,v\right]$$

(7)

In the convolutional operation described above, $x\left[i,j\right]$ represents the pixel value at position $(i, j)$ in the input image, $x$, while $K\left[u,v\right]$ denotes the weight value at position $(u, v)$ in the filter. The output, $y\left[i,j\right]$, is the result at position $(i, j)$ on the feature map. $F$ indicates the filter size. This convolutional process involves sliding the filter, $K$, across image $x$; computing the dot product of $K$ and the corresponding subregion of $x$ that it covers at each step; and assigning this result to the corresponding location in the output feature map $y$. This operation captures and translates the patterns and textures from the input image into useful features for further analysis, such as WQI estimation.

The feature map was processed after the convolution operation through normalization and the rectified linear unit (ReLU) activation function to enhance the model’s training speed and introduce non-linearity. The ReLU function [33] is calculated as follows:

$$\mathsf{\delta }\left(x\right)= max\left(0, x\right)$$

(8)

The output feature maps are then processed by a ResNeXt block. In ResNeXt, the input feature map is divided into several groups, and a separate convolution is applied to each group. These outputs are concatenated along the channel dimension. The grouped convolution operation is defined as

$$F_{\mathrm{group}}=\bigcup _{g=1}^G{\mathrm{Conv}}_g\left(F_g^{\mathrm{in}}\right)$$

(9)

where $F_g^{\mathrm{in}}$ denotes the input features of the $g$-th group, and ${\mathrm{Conv}}_g\left(·\right)$ is the convolution operation applied to that group. The outputs from each group are concatenated to form $F_{\mathrm{group}}$.

This result is then added to the identity shortcut $R$ through a residual connection:

$$F_{\mathrm{res}}=F_{\mathrm{group}}+R$$

(10)

Here, $R$ is typically an identity mapping or a 1 × 1 convolution if the dimensions do not match.

The output $F_{\mathrm{res}}$ is passed through the CBAM to emphasize informative features and suppress irrelevant ones. The CBAM consists of two sequential attention modules: channel attention and spatial attention. The channel attention map, $M_c$, is generated using both average pooling and max pooling operations, followed by a shared fully connected ($\mathrm{FC}$) layer and a sigmoid activation function, $\sigma \left(·\right)$, as shown below:

$$M_c\left(F\right)=\sigma \left(FC\left(AvgPool\left(F\right))+FC(MaxPool\left(F\right)\right)\right)$$

(11)

Next, spatial attention, $M_s$, is computed by concatenating the average-pooled and max-pooled features across the channel dimension, followed by a convolutional layer with a 7 × 7 kernel:

$$M_s\left(F\right)=\sigma \left(f_{7\times 7}\left(\left[AvgPool\left(F\right);MaxPool\left(F\right)\right]\right)\right)$$

(12)

The final attention-refined feature map is calculated by sequentially applying the channel and spatial attention mechanisms to the residual features:

$$F_{\mathrm{CBAM}}=M_s\left(M_c\left(F_{\mathrm{res}}\right)·F_{\mathrm{res}}\right)·\left(M_c\left(F_{\mathrm{res}}\right)·F_{\mathrm{res}}\right)$$

(13)

This process enhances important channels and spatial regions, making the model more sensitive to critical features related to water quality.

The refined feature map, $F_{\mathrm{CBAM}}$, is then transformed into a vector through global average pooling (GAP) [34], which averages each channel across spatial dimensions:

$$F_{\mathrm{GAP}}\left[k\right]={\displaystyle \frac{1}{H\times W}}\sum _{i=1}^H\sum _{j=1}^W{F}_{\mathrm{CBAM}}\left[k,i,j\right]$$

(14)

where $H\times W$ is the spatial resolution, and $k$ is the channel index.

This vector is input to a fully connected layer to estimate the WQI via regression:

$$\widehat{y}=W·F_{\mathrm{GAP}}+b$$

(15)

where $W$ and $b$ are the learnable weights and bias terms, respectively, and $\widehat{y}$ is the predicted WQI

Through this architecture, KWEN effectively extracts and emphasizes meaningful visual features from satellite images, allowing for accurate estimation of freshwater quality indices.

3.1.2. ResNet

Originally developed by Microsoft Research, ResNet has demonstrated exceptional performance in preventing overfitting in image classification and regression tasks [29]. A key feature of ResNet is its use of shortcut connections, which directly add inputs to outputs.

This approach addresses the problem of gradient vanishing as the model depth increases. In this study, the ResNet18 model integrated into the Torchvision library was used.

3.1.3. LeNet

This study used a modified version of the LeNet5 model, a concise and efficient CNN. This modified model is designed to process image inputs of size 224 × 224 with three channels. It consists of two convolutional layers for feature extraction, followed by three fully connected layers for making numerical predictions. The first convolutional layer has three input channels and six output channels, and the second layer has six input channels and 16 output channels, both with a kernel size of 5. The fully connected layers are adjusted based on the input feature map size, and the model outputs a single real value. ReLU activation function and max pooling operations are used in this process.

3.1.4. CNN

CNNs have been applied to image analysis for classification and regression tasks [31]. CNNs are deep learning models designed to process structured grid data, extracting hierarchical representations of the input data through interconnected layers [35]. These layers comprise convolutional layers, pooling layers, and fully connected layers [36]. Each filter in the convolutional layers is responsible for detecting specific features of the input [37].

A CNN architecture was demonstrated using Landsat-8 imagery to estimate water quality [9], and it was found that a four-layer CNN architecture performs the best. In our study, we also utilized a similar CNN architecture with four convolutional layers. The input images were 224 × 224 pixels with three channels. The first layer mapped these three channels to six, and then expanded the channel count to 12, 24, and finally 48. The output was a vector of size 9408, which fed into three fully connected layers to produce a single output value. The ReLU activation function was used to introduce non-linearity into the architecture.

3.2. Computing Experiment Environment

In this study, the proposed model was trained and tested using a specialized computing environment. The setup included a Ryzen 5700X CPU and an RTX 3060 GPU, supported by 32 GB of DDR4 memory. The software framework was based on Python version 3.9.1, with PyTorch version 2.0.1 used for model development and training. All processes were executed on Windows 10 as the operating system.

4. Results

4.1. Accuracy Evaluation

Five statistical indicators were applied in this study to evaluate model performance: index of agreement (IOA), scatter index (SI), root mean square error (RMSE), mean absolute error (MAE), and R². These indices are commonly used to assess regression models, such as WQI score estimation [12,38,39,40,41].

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

(16)

$$MAE={\displaystyle \frac{1}{n}}\sum _{i=1}^n\left|y_i-\widehat{y_i}\right|$$

(17)

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

(18)

$$IOA=1-{\displaystyle \frac{{\sum }_{i=1}^n{\left(y_i-\widehat{y_i}\right)}^2}{{\sum }_{i=1}^n{\left(\left|\widehat{y_i}-\overline{y}\right|+\left|y_i-\overline{y}\right|\right)}^2}}$$

(19)

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

(20)

where $\widehat{y_i}$ represents the WQI value predicted by KWEN, $y_i$ denotes the WQI value calculated from the observed data, and $n$ indicates the number of samples. The RMSE, MAE, and SI are metrics used to calculate the difference between the values predicted by the numerical models and the actual measured data, with lower values indicating better model performance.

The IOA [42] is a standardized scale for the degree of estimation error in numerical models. It ranges between 0 and 1, with values closer to 1 indicating better model performance. R² measures the fit between the water quality concentrations predicted by the model and the validation data, with values closer to 1 indicating stronger predictive capabilities of the model.

4.2. Statistical Analysis

In this study, model performance was comprehensively assessed using multiple statistical indicators, including the index of agreement (IOA), mean absolute error (MAE), root mean square error (RMSE), scatter index (SI), coefficient of determination (R²), and 5% accuracy.

Table 3. Statistical performance of WQI prediction models.

	Index of Agreement	Mean Absolute Error	Root Mean Square Error	Scatter Index	R2	5% Accuracy
Korea Water Evaluate Network	0.99	1.89	3.02	0.04	0.94	89.29%
Residual network	0.95	3.10	4.68	0.06	0.81	82.20%
LeNet	0.93	3.13	5.15	0.06	0.77	77.11%
Convolutional neural network	0.89	4.68	6.49	0.08	0.64	56.28%

presents the quantitative results obtained from the testing phase. Each metric plays a crucial role in evaluating the models’ predictive performance, and their corresponding contributions are described as follows:

  I.

IOA: This metric measures the degree of concordance between the predicted and observed values. A value closer to 1 indicates that the model predictions more closely match the observed outcomes, providing a useful gauge of overall model efficiency [42].

 II.

MAE: This represents the average of the absolute differences between the predicted and actual values. Lower MAE values signify more accurate predictions [43].

III.

RMSE: Calculated as the square root of the average of the squared differences between the predicted and actual values, RMSE represents the magnitude of prediction errors. This index emphasizes larger errors, making it particularly useful for assessing the sensitivity of a model to outliers [44].

IV.

SI: This measures how spread out the model predictions are compared with the actual values. A lower value indicates less variance in the predictions, implying higher model consistency [45].

 V.

R²: This indicates the proportion of variance in the observed data explained by the model. Higher values suggest that the model effectively explains the data; therefore, this tool is essential for evaluating a model’s explanatory power [46].

VI.

5% Accuracy: This metric indicates the proportion of model predictions within 5% of the actual values. A higher percentage indicates that the model provides highly accurate predictions.

In Table 3, the KWEN model significantly outperforms all others according to the statistical benchmarks. For instance, the KWEN model has an MAE value of 1.89, which is around 39% lower than that of the next best performer, ResNet, whose MAE is 3.1. This means that the KWEN model generally makes more accurate predictions with smaller errors. Similarly, for RMSE, KWEN shows a value of 3.02, which is approximately 35% lower than ResNet’s value of 4.68.

The LeNet and CNN models exhibit relatively low performance. LeNet’s MAE was 3.13, approximately 66% higher than that of KWEN, whereas CNN’s MAE was the highest at 4.68, 148% higher than that of KWEN. A similar trend is observed for RMSE, which is 5.15 for LeNet and 6.49 for CNN, which are 71% and 115% higher, respectively, than KWEN’s 3.02.

Regarding IOA, KWEN showed the highest score at 0.99, followed by ResNet at 0.95, LeNet at 0.93, and CNN at 0.89. The R² values were 0.94 for KWEN, 0.81 for ResNet, 0.77 for LeNet, and 0.64 for CNN. Finally, the 5% accuracy metric shows KWEN achieved 89.29% accuracy, significantly higher than ResNet’s 82.20%, LeNet’s 77.11%, and CNN’s 56.28% accuracy.

These quantitative results indicate that the KWEN model is more accurate and reliable than other models, making it a more dependable choice for predicting water quality in real-world environments.

4.3. Model Validation

Figure 3 presents scatter plots comparing the predicted and observed values of WQI for the four models: KWEN, ResNet, LeNet, and CNN. Each subplot illustrates the predictive performance of an individual model. A higher concentration of data points along the diagonal line indicates greater agreement between the predictions and the actual observations.

The KWEN model shows the most consistent alignment with the diagonal, demonstrating high predictive accuracy, as shown in Figure 3a. The ResNet model tends to overestimate, as indicated by the clustering of points above the diagonal (Figure 3b). In contrast, the LeNet and CNN models display more dispersed distributions, indicating larger deviations from the observed values, as shown in Figure 3c,d. The CNN model, in particular, exhibits the greatest spread, reflecting substantial prediction errors and variance, as illustrated in Figure 3d. These graphical findings align with the statistical analysis discussed in Section 4.2. Among the models, KWEN achieved the highest predictive accuracy, followed by ResNet and LeNet, while CNN showed the weakest performance.

Figure 4 provides residual histograms that illustrate the distribution of prediction errors for each model. The residuals, defined as the difference between predicted and observed values, offer insights into each model’s accuracy and stability. Standard deviation and skewness values are provided to support quantitative comparison.

The KWEN model yields the narrowest distribution of residuals, indicating minimal error and stable performance, as shown in Figure 4a. Its standard deviation is the lowest among all models, and its residuals are symmetrically distributed around zero. The ResNet model shows a moderately wider distribution and higher positive skewness, implying a tendency toward overprediction, as illustrated in Figure 4b. LeNet exhibits both increased spread and asymmetry, suggesting reduced predictive reliability, as presented in Figure 4c. CNN displays the highest standard deviation, indicating the largest variability in prediction errors, despite its relatively low skewness, as depicted in Figure 4d.

Overall, the KWEN model demonstrated the most accurate and consistent predictions, followed by ResNet and LeNet. The CNN model exhibits the least reliable performance, with significant variability across its predictions.

4.4. Gradient-Weighted Class Activation Mapping (Grad-CAM) Visualization Analysis

Grad-CAM is a technique used to visually interpret which parts of an input image a deep learning model focuses on when making predictions. Unlike quantitative evaluation metrics, Grad-CAM offers qualitative insights, enabling an intuitive understanding of model behavior. In this study, since the input consists of satellite RGB imagery, Grad-CAM visualization serves as an effective tool to examine the model’s learning focus and prediction logic.

Figure 5 illustrates the Grad-CAM results applied to the final convolutional layer of each model: (a) KWEN, (b) ResNet, (c) LeNet, and (d) CNN. The first row shows the original Google Earth satellite image, the second row presents the relative attention weight maps derived from Grad-CAM, and the third row displays the overlay of these heatmaps on the original images. These visualizations help to identify which regions of the input each model emphasized when estimating the WQI from the spatial land-use context.

Both KWEN and ResNet exhibit strong attention toward features associated with surrounding land use and urbanized areas adjacent to the river, indicating that these models incorporate broader spatial context beyond the river channel itself, as shown in Figure 5a,b. This pattern suggests that visible landscape features in adjacent areas are reflected in the estimation process. In contrast, LeNet and CNN focus predominantly on the geometric shape and structural characteristics of the river channel itself, indicating relatively strong emphasis on morphological features, as illustrated in Figure 5c,d.

Notably, KWEN demonstrates more localized and spatially coherent attention regions compared to ResNet. This observation indicates a more concentrated spatial response, which can be attributed to its architectural design combining a ResNeXt backbone with a convolutional block attention module (CBAM). By effectively capturing both spatial and channel-wise importance, KWEN places greater emphasis on land-use heterogeneity and urbanized surroundings.

Similarly, LeNet exhibits more continuous and structured attention along the river course than CNN, as evidenced by clearer activation patterns in its Grad-CAM visualization. This indicates a more continuous spatial focus along linear river structures. In contrast, CNN shows more fragmented and dispersed attention, indicating weaker spatial consistency. Overall, the Grad-CAM results show that the proposed framework utilizes a combination of river morphology and surrounding land-use characteristics during prediction.

5. Discussion

5.1. Comparison with Existing Studies and Ecological Implications

Understanding the extent to which a data-driven model captures environmentally meaningful patterns, rather than spurious statistical correlations, is central to evaluating its applicability. In this section, the spatial attention behavior identified through Grad-CAM and the predicted WQI distributions are examined in the context of established land-use water quality relationships reported for Korean and international river systems.

The underlying assumption of the proposed framework is that ecologically intact environments—such as forested headwaters and mountainous reaches with minimal anthropogenic disturbance—are associated with higher WQI scores, whereas areas subject to intensive agricultural activity or urbanization exhibit lower scores due to cumulative water quality degradation. This assumption is well supported by existing evidence. In the Han River basin, the largest and most densely populated watershed in South Korea, spatial variations in total nitrogen, total phosphorus, chemical oxygen demand, and suspended solids have been shown to be largely governed by combinations of agricultural land cover, urbanization intensity, and topographic characteristics, with forested upstream reaches consistently maintaining better water quality than urbanized downstream areas [47]. A broader assessment across 64 streams spanning the Han, Nakdong, Geum, and Yeongsan River basins corroborated this gradient, confirming that land-use patterns serve as a key determinant of water quality across the four major Korean watersheds [48]. The primary mechanism underlying this spatial gradient is non-point source pollution: agricultural fertilizers, livestock waste, and urban surface runoff are transported into rivers through diffuse pathways during precipitation events, progressively degrading water quality as watershed development intensifies [49,50]. The Grad-CAM analysis presented in this study confirmed that these expectations were broadly met. The model consistently attended to surrounding land-use characteristics—including agricultural fields, built-up areas, and impervious surfaces adjacent to river channels—rather than relying solely on in-channel pixel information. This correspondence suggests that the model operates through an indirect estimation mechanism: rather than detecting water quality constituents directly from the water surface, it leverages the spatial configuration of surrounding land use—which reflects the intensity and distribution of NPS pollution loading—to infer the overall water quality condition, as represented by the WQI.

This indirect, context-driven approach distinguishes the proposed framework from the majority of existing satellite-based water quality studies, which have predominantly relied on multispectral or hyperspectral sensors to retrieve individual water quality parameters through spectral signatures directly linked to optically active constituents. For instance, previous studies have coupled Landsat 8 OLI-TIRS spectral indices with AI models to estimate the WQI [16] and have evaluated machine learning algorithms for predicting optically active indicators such as turbidity, total dissolved solids, and chlorophyll-a from Landsat 8 and Sentinel-2 imagery [51,52,53]. These approaches are methodologically advantageous for the estimation of individual optically active parameters, as the spectral bands employed are physically related to the absorption and scattering properties of the target constituents. The present study, however, employs RGB imagery, which lacks the spectral resolution required for such direct retrieval. Instead, the framework relies on contextual landscape features that are visually discernible in high-resolution imagery. A critical distinction arises here: the WQI is, by definition, a composite indicator that integrates multiple physicochemical parameters into a single value, and its spatial variability is governed not only by in-stream optical properties but also by broader watershed-scale processes, most notably land-use-driven NPS pollution [49,50]. In this regard, the RGB-based contextual approach may capture a dimension of information—namely, the spatial configuration of anthropogenic pressures surrounding a river reach—that spectral methods focused on the water surface alone do not explicitly represent. Accordingly, the proposed framework should not be regarded as a substitute for spectrally rich remote sensing methods but may serve as a supplementary screening tool that provides spatially informed preliminary assessments, particularly in settings where archival water quality records and matched RGB imagery are available but multispectral data are not.

The Grad-CAM sensitivity patterns further inform the practical scope and limitations of this framework. The model’s consistent attention to agricultural and urbanized land-use features aligns with the well-documented understanding that NPS pollution from these sources constitutes a primary driver of water quality deterioration in Korean and international river systems [47,48,49]. It has also been noted that convolutional neural networks are increasingly being applied to water quality estimation owing to their capacity to capture complex, non-linear spatial dependencies that conventional regression models may not fully represent [54]. That the model highlights landscape features associated with known pollution drivers, rather than arbitrary image artifacts, supports the validity of its learned spatial representations. It should be acknowledged, however, that the model identifies spatial associations between landscape context and estimated WQI values rather than diagnosing specific pollution sources or causal pathways. Within this limitation, the proposed framework may complement conventional point-based monitoring programs by enabling preliminary identification of reaches where land-use pressures are visibly reflected in the surrounding environment—particularly in peri-urban and agricultural transition zones where diffuse NPS pollution sources remain difficult to characterize through spatially sparse in situ sampling networks alone [49,50].

5.2. Methodological Limitations and Scope

While Section 5.1 situated the findings within the existing scientific literature and ecological gradients, it is equally important to define the methodological scope within which the present results should be interpreted, and several limitations should be acknowledged. First, the modeling framework was developed using a single source of Google Earth RGB imagery, and its robustness under alternative image providers, spatial resolutions, acquisition conditions, and preprocessing workflows was not evaluated. Variability in image timing and environmental conditions may influence predictive stability. Second, because the target variable is a composite index derived from predefined parameters and weighting schemes, uncertainties inherent in WQI formulation are implicitly embedded in the training data. In particular, as noted in Section 2.3, the WQI used in this study was computed without turbidity and fecal coliform, which may limit its sensitivity to sediment-related disturbance and microbiological contamination. Third, although the dataset was partitioned at the monitoring station level to reflect spatial separation, some stations are geographically proximate, and residual spatial dependency among nearby locations cannot be completely excluded. Accordingly, the environmental and geographic scope represented in the dataset may constrain applicability beyond the examined domain. In addition, RGB imagery does not directly measure optically active water constituents such as chlorophyll or suspended matter, which are typically captured by multispectral or hyperspectral sensors. Instead, the present framework relies on visible contextual features, including surrounding land-use and watershed characteristics. This context-oriented approach enhances accessibility but defines a methodological boundary in which river condition is inferred indirectly rather than through direct spectral water-quality signals.

Despite these constraints, the findings indicate that RGB-based deep learning models can extract spatial patterns associated with river condition within the analyzed dataset and that such context-driven estimation remains systematically aligned with calculated WQI values. The consistent emphasis on surrounding land-use features suggests that visible imagery captures spatial context relevant to WQI variability. When interpreted within clearly defined methodological boundaries, the proposed framework offers a spatially informed analytical perspective that may complement conventional monitoring systems, particularly in data-supported environments where archival water quality records and matched imagery are available.

6. Conclusions

This study applied a convolutional neural network to RGB imagery in order to estimate river WQI values using spatial information derived from widely accessible images. By integrating image data with national water quality monitoring records, an image-based analytical framework was implemented and evaluated.

The results indicate that spatial features observable in RGB imagery can be utilized to approximate river condition within the study domain. Model attention analysis showed consistent emphasis on surrounding land-use characteristics and landscape configuration, suggesting that spatial context is reflected in the estimation process.

Overall, the proposed framework offers a spatially oriented perspective for river environmental assessment by linking landscape-scale features derived from RGB imagery with observed WQI values. However, its applicability depends on the availability of matched archival water quality records and corresponding image datasets. Within such data-supported contexts, it may nonetheless serve as a complementary tool for environmental evaluation and planning. Future validation across diverse regions and imaging conditions would be necessary to address current limitations and establish the practical applicability of the proposed framework.