Groundwater Vulnerability and Environmental Impact Assessment of Urban Underground Rail Transportation in Karst Region: Case Study of Modified COPK Method

Abstract

Urbanization always leads to increasing challenges to the groundwater resources in karst regions due to intensive land use, infrastructure development, and the rapid transmission potential of pollutants. This study proposed an improved groundwater vulnerability assessment (GVA) framework by modifying the widely used COP (Concentration of flow, Overlying layers, and Precipitation) model, through the integration of three additional indicators: urban underground rail transportation (UURT), land use and cover (LULC), and karst development (K). Guiyang, a typical urbanized karst city in southwest China, was selected as the case study. The improved COP model, namely the COPK model, showed stronger spatial differentiation and a higher Pearson correlation coefficient ($r$) with nitrate concentrations ($r$ = 0.4388) compared to the original COP model (R = 0.3689), which validates the effectiveness of the newly introduced indicators. However, both R values remained below 0.5, even after model modification, suggesting that intensive human activities play a role in influencing nitrate distribution. The pollution load index (PI) was developed based on seven types of pollution sources, and it was integrated with the COPK vulnerability index using a risk matrix approach, producing a groundwater risk map classified into five levels. Global Moran’s I analysis (0.9171 for COP model and 0.8739 for COPK model) confirmed strong and significant spatial clustering patterns for the two models. The inclusion of UURT and LULC improved the model’s sensitivity to urban-related pressures and enhanced its capacity to detect local risk zones. It is a scalable tool for groundwater risk assessment in urbanized karst areas and offers practical insights for land use planning and sustainable groundwater management.

1. Introduction

Karst aquifer systems (KASs) constitute a critical component of the global freshwater resources, which provide potable water for approximately 25% of the global population [1,2]. Due to their unique geological structures, such as complex networks of fissures and caves, KASs are highly sensitive to external changes in the water cycle, including rainfall events, seasonal fluctuations in recharge, land use alterations, and engineering disturbances [3,4,5,6]. Also, KASs are particularly vulnerable to pollution, where contamination events often result in extended and difficult recovery efforts [7,8,9]. In recent years, human activities such as excessive extraction [10], urban expansion [11], poor management practices [12], industrial discharges [13], and agricultural runoff [14] have increasingly threatened both the quantity and quality of groundwater in KASs. These pressures compromise the sustainability of groundwater, leading to risks to public health and ecosystem stability [15,16,17,18]. Therefore, employing advanced technologies such as remote sensing (RS), geographic information systems (GISs), and predictive modeling has become a crucial pursuit for global scholars aiming to achieve sustainable groundwater management [19,20].

Groundwater vulnerability assessment (GVA) serves as an essential management tool for identifying areas where pollutants from the surface are more likely to reach the groundwater system [21,22,23]. It is a critical initial step in hydrogeological studies and lays the foundation for future sustainable groundwater management [24,25]. Groundwater vulnerability (GV) can be categorized into intrinsic vulnerability, which considers the aquifer’s physical characteristics such as geology and hydrology, and specific vulnerability, which assesses the impact of specific contaminants and their interaction with the aquifer [26,27,28,29]. In karst regions, a range of GVA models have been specifically developed or adapted for KASs, including EPIK [30], REKS [31], RISKE [32], PI [33], KARSTIC [34], COP [35], PaPRIKa [36], IZDAN [37], PRESK [38], DRISTPi [39], LEPT [40], and IVAKY [41]. These models incorporate, either fully or partially, key geological, hydrological, and climatic features such as soil properties, surface conditions, karst development, characteristics of the saturated and unsaturated zones, and precipitation patterns to construct karst GVA models. In these models, the COP model is one of the most commonly used GVA models, focusing on three key factors: concentration of flow (C), overlying layers (O), and precipitation (P) [42,43,44,45], making it a practical and widely adopted framework for assessing intrinsic vulnerability in KASs.

Beyond direct applications, recent studies have frequently modified COP models to enhance their applicability and accuracy. For example, Li et al. [46] enhanced the COP model by incorporating sinkhole density, land use and cover (LULC), and karst network development, significantly improving GVA in karst regions. Cao et al. [47] and Qiu et al. [48] modified the COP method into the COPK approach by considering karst development to assess GV in northern China’s karst areas, notably improving model accuracy. Ghezelayagh et al. [49] introduced the COPKAT model, which enhances the traditional COP model by incorporating karstification, atmospheric conditions, and tectonics using the fuzzy logic method. Although these improvements to the traditional COP model enhance its performance by introducing new indicators, they still do not sufficiently address the changing surface conditions and hydrological dynamics under rapid urbanization. In particular, LULC such as impervious surfaces can significantly alter infiltration and runoff behavior. Also, urban underground rail transportation (UURT) introduces subsurface disturbances and concentrated drainage structures [50,51,52,53]. However, the two factors are rarely considered in existing models. In the context of urbanization, particularly in China, UURT and the increase in impervious surfaces are important aspects of city development [54]. Effectively integrating the impacts of UURT and LULC on groundwater into GVA models is crucial for more accurately identifying pollution-prone areas.

The aim of this study is to modify the traditional COP model by incorporating the impacts of UURT and LULC in Guiyang, a typical karst region, to accurately construct a GVA model. Additionally, the regional groundwater risk distributions are calculated, taking into account seven potential pollution sources (PPSs), including industrial, agricultural, mining activities, hazardous waste disposal sites, landfills, gas stations, and golf courses. This study is the first to incorporate UURT into GVA in karst areas, while also systematically integrating LULC. Combined with regional groundwater risk mapping, this constitutes the main innovation of the work. Nitrate (NO₃⁻), a commonly used groundwater quality parameter for validating GV models, is employed to compare the performances of the modified model before and after its improvement. The objectives of this study are twofold: first, to establish a GVA model for karst regions that takes into account the impacts of UURT and LULC within an urbanized context in Guiyang City; and second, to assess regional groundwater risks, considering multiple PPSs.

2. Study Area

Guizhou Province, located in the Yunnan–Guizhou Plateau, is one of the three largest karst regions in the world, with carbonate rocks accounting for 71.7% of its total area. The study area, Guiyang City, is located in the center of Guizhou Province, covering an area of 2527.32 km², with exposed carbonate rocks making up 74.68% (Figure 1). The study area has a subtropical humid and temperate climate, with an average temperature of 15.3 °C, extreme highs of 39.5 °C, and lows of −9.5 °C. It has an annual average of 1148.3 h of sunshine and records minimal snowfall. This area has an unevenly distributed annual precipitation averaging 1136.96 mm, peaking during the rainy season from April to October, while the dry season spans from December to March. The terrain is varied, being higher in the north and south, with elevations ranging from 800 m to 1600 m. Geologically, the study area is characterized by significant uplift during the Early and Middle Paleozoic, resulting in the widespread absence of the Devonian, Silurian, and parts of the Carboniferous and Ordovician strata across most of the region. In several locations, the Lower Carboniferous or Permian strata are in direct contact with underlying Cambrian or Ordovician rocks. The widespread distribution of carbonate rocks fosters active karst features like sinkholes and underground rivers, increasing the region’s susceptibility to surface pollution. As urbanization progresses, exemplified by three operational urban rail transit lines as of March 2024, the risks of groundwater contamination have heightened, highlighting the need for focused studies on GV in this karst-rich municipal area.

3. Methodology

Figure 2 illustrates the methodological framework of this study. Based on the traditional COP model, we introduced three additional factors: distance to UURT, LULC, and karst development (K). They were integrated into the vulnerability index calculation to formulate the improved COPK model. Nitrate (NO₃⁻), as a commonly used groundwater quality parameter, was employed to validate and compare the performance of the modified COPK model against the COP model. Finally, by combining the COPK results with various PPSs, regional groundwater risk was assessed, resulting in a spatial risk zoning map.

3.1. Data Source and Pre-Processing

Table 1. Data sources.

Data	Sources	Data Type	Scale
Soil	Harmonized World Soil Database v1.2	Raster	1 km
Lithology	No.111 Geological Party, GBGMED (http://www.gzdk111.cn/, accessed on 23 March 2025)	Shapefile (polygon)	1:50,000
NDVI	MOD13A3 (https://www.earthdata.nasa.gov/, accessed on 1 April 2025)	Raster	1 km
LULC	MNR of the P.R.C. (https://www.mnr.gov.cn/, accessed on 1 April 2025)	Shapefile (polygon)	1:50,000
DEM	Geospatial Data Cloud (http://www.gscloud.cn, accessed on 1 April 2025)	Raster	30 m
UURT	GPTIO Group (https://www.gyurt.com/, accessed on 6 April 2025)	Shapefile (line)	1:50,000
Precipitation	IGSNRR (http://english.igsnrr.cas.cn/, accessed on 6 April 2025)	csv.	/
Groundwater samples	No.111 Geological Party, GBGMED	/	/

Note: GBGMED: Guizhou Bureau of Geology and Mineral Exploration & Development; MNR: Ministry of Natural Resources; GPTIO group: Guiyang Public Transportation Investment and Operation Group Co., Ltd.; IGSNRR: Institute of Geographic Sciences and Natural Resources Research.

presents the data sources used in this study. The dataset includes vector, raster, and csv formats. All vector data, such as lithology, LULC, and distance to UURT, were converted into raster format with a spatial resolution of 30 m using vector-to-raster transformation in ArcGIS 10.8. For raster datasets such as normalized difference vegetation index (NDVI, 1 km), digital elevation model (DEM, 30 m), and soil (1 km), resampling techniques were applied to standardize them to a uniform resolution of 30 m. This process can facilitate the following overlay analysis and spatial calculations for GV map generations [46,55,56].

3.2. COP Model

The baseline COP model assesses GV by individually scoring the C, O, and P factors and then calculating their product (Equation (1)). A higher COP index indicates a stronger protective effect against contamination, corresponding to the lower GV of the KAS [23,35]. In this study, the calculation of the COP model was conducted in strict accordance with the methodology proposed by Vías et al. [35]. and the detailed scoring criteria for each factor are provided in Appendix A.

$$COP index=C score·O score·P score$$

(1)

where C score, O score, and P score represent the degree of flow concentration (C), the protective capacity of the overlying geological layers (O), and the precipitation intensity and distribution (P), respectively.

3.2.1. C Factor

The C factor (concentration of flow) in the COP model considers four elements: sinkholes, sinking streams, slope, and vegetation cover (Figure 3). According to the existing literature, areas that are close to sinkholes and sinking streams and have low vegetation cover and steep slopes tend to have lower scores, which indicates the weaker natural protection and higher GV of the KSA [45,48]. Figure 3e presents the spatial distribution of the C factor as calculated in the traditional COP model (Equation (2)).

$$C score=dh·ds·sv$$

(2)

where $dh$, $ds$, and $sv$ represent the distance to sinkholes, the distance to sinking streams, and the harmonized score of slope and vegetation cover, respectively.

3.2.2. O Factor

The O factor in the traditional COP model includes four key components: soil type, soil thickness, lithology, and confining conditions. In general, compact soil or lithologies, greater soil thickness, and confined aquifer conditions provide stronger protection against contamination [57,58]. The calculation for the O score is provided in Equation (3). The spatial distribution of the O factor and its four contributing indicators are shown in Figure 4.

$$O score=O_S+O_L$$

(3)

where ${\mathrm{O}}_{\mathrm{S}}$ represents the protective capacity of the soil layer, and ${\mathrm{O}}_{\mathrm{L}}$ is the protective capacity of the unsaturated lithological layer beneath the soil.

3.2.3. P Factor

The P factor in the COP model typically considers two main components: annual average rainfall and temporal distribution [45]. Precipitation can significantly affect the migration of contaminants in the subsurface environment [59]. The temporal distribution of precipitation affects the timing and duration of recharge events, which in turn influences the potential for surface pollutants to reach the aquifer. P score is calculated by Equation (4) and Figure 5 shows the distribution of P factor and two indicators.

Table 2. Precipitation information in Guiyang City.

Station Name	Latitude (°)	Longitude (°)	Average Annual Rainfall (mm/Year)	Average Annual Rainy Days	Temporal Distribution (mm/Day)
Xifeng	27.10	106.72	1104.03	179	6.17
Kaiyang	27.07	106.97	1172.29	188	6.24
Xiuwen	26.83	106.60	1098.03	181	6.07
Qingzhen	26.57	106.47	1193.88	179	6.67
Guiyang	26.58	106.73	1160.45	178	6.52
Baiyun	26.67	106.65	1134.03	179	6.34
Huaxi	26.42	106.67	1183.91	175	6.77
Wudang	26.63	106.77	1134.06	173	6.56

presents the information of all monitoring stations in the study area from 2009 to 2023. According to the COP model standard criteria outlined in Appendix A, all stations have P_Q values within 800–1200 mm (assigned a weight of 0.2) and P_I values below 10 mm/day (assigned a weight of 0.6). Therefore, there is a uniform p-value (0.8) throughout the study area.

$$P score=P_Q+P_I$$

(4)

where ${\mathrm{P}}_{\mathrm{Q}}$ and $P_I$ represent the quantity and temporal distribution of precipitation, respectively.

3.3. Modified COPK Model

To enhance the applicability of the COP model under urbanization, particularly in karst regions influenced by UURT and LULC, three additional indicators were introduced in this study: distance to UURT (DU), LULC, and karst development (K). The modified COPK model is calculated by Equation (5).

$$COPK index=\left(dh·ds·sl\right)·\left(O_S+O_L+O_{LU-DU}\right)·\left(P_Q+P_I\right)+K$$

(5)

where sl refers to the composite score of slopes and LULC, $O_{LU-DU}$ presents the composite score of LULC and distance to UURT, and K is the karst development. The meanings of the other variables are consistent with those defined previously.

In particular, $O_{LU-DU}$ is defined as follows:

$$O_{LU-DU} \left(x,y\right)=M\left(x,y\right)·O_{DU}\left(x,y\right)+[1-M\left(x,y\right)]·O_{LU}\left(x,y\right)$$

(6)

where ${\mathrm{O}}_{\mathrm{D}\mathrm{U}}\left(\mathrm{x},\mathrm{y}\right)$ is the score based on the distance to UURT at location $\left(\mathrm{x},\mathrm{y}\right)$, where lower values indicate higher GV due to proximity. $O_{LU}\left(x,y\right)$ is the LULC score at location $\left(x,y\right)$, and $M\left(x,y\right)$ is a binary spatial mask function, equal to 1 if the location is within the UURT influence zone, and 0 otherwise.

3.3.1. Modified C Factor

In the modified C factor, LULC replaces the NDVI component in the traditional COP model. The original binary classification of vegetation as simply “high” or “low” was considered insufficient to represent real surface conditions. Instead, LULC was integrated with a slope to form a composite variable (sl), with scoring criteria provided in

Table 3. The score of slopes and LULC (sl) in the C factor.

Slope	LULC	Score
≤8%	-	1.00
(8–31]	Shrubland, forest, or grassland	0.95
	Cultivated land	0.90
	Bare land	0.85
	Construction land	0.80
(31–76]	Shrubland, forest, or grassland	0.75
	Cultivated land	0.70
	Bare land	0.65
	Construction land	0.60
>76%	-	0.55

.

From the aspect of flow concentrations, areas categorized as construction land, cultivated land, or bare land are more likely to promote surface runoff and increase the concentration of flow, thus receiving lower scores in the C factor assessment [11,60]. These LULC types reduce surface infiltration and enhance the potential for rapid pollutant transport toward KASs. In contrast, areas covered by shrubland, forest, or grassland tend to enhance surface infiltration and reduce concentrated flow [61,62]. Figure 6 shows the sl indicator and spatial distribution of the modified C factor.

3.3.2. Modified O Factor

For the improvement of the O factor, the effects of UURT and LULC were integrated to better represent the protective capacity of the overlying layers. A composite indicator ($O_{LU-DU}$) was developed and the corresponding scores are shown in

Table 4. The scores of UURT and LULC (O_LU-DU) for the O factor.

Indicator	Classification	Score
Land use	Bare land	1
	Cultivated land	2
	Shrub or grassland	3
	Forest	4
	Construction land	5
Distance to metro line (m)	1000	1
	1500	2
	2000	3

. The areas closer to UURT lines were assigned lower scores, indicating a higher GV because of the increased disturbance of the natural hydrological system. Notably, construction land, as an impervious surface, has the highest protective capacity in terms of preventing vertical infiltration, and thus receives the highest scores in this component. In O_LU-DU, the influence of UURT is given priority over LULC when determining the overlaying layers. Specifically, scores from UURT are assigned to areas within its influence zone, replacing the original land use scores, while land use values are retained elsewhere. The modification process is shown in Figure 7.

3.3.3. K Factor

The K factor was introduced as an additional factor to represent the degree of karstification, which directly influences GV due to variations in conduit connectivity [63]. In this study, karst development was classified based on the Chinese standard (Code for Geotechnical Investigations of Urban Rail Transit of Guizhou Province, DBJ52/T099-2020 [64]). This classification standard integrates carbonate rock formations, karst development features, hydrogeological characteristics, and representative lithologies of Guizhou to categorize the K factor into three levels: strong, moderate, and poor [46]. These categories are assigned scores from 1 to 3, respectively, and other areas outside the classified karst zones are assigned a score of 4. The detailed classification method is provided in Appendix B, and Figure 8 shows the spatial distribution of the K factor.

3.4. Model Validation

To assess the performance of the two GV models, nitrate (NO₃⁻) was selected as the groundwater quality parameter for validation, given its widespread use as a tracer for anthropogenic pollution in KASs [65,66,67]. In this study, the Pearson Correlation Coefficient ($r$) was used to validate the performance of the original COP and modified COPK models [67,68,69]. This correlation was calculated for two models to quantify the correlation between the predicted GV scores and the measured nitrate levels, and it has been widely used for the validation of GV models [48,70]. Generally, a higher $r$ indicates a higher model performance for the GV model. The equation for r calculation is as follows:

$$r={\displaystyle \frac{{\sum }_{i=1}^n(X_i-\overline{X})(Y_i-\overline{X})}{\sqrt{\sum _{i=1}^n{(X_i-\overline{X})}^2}\sqrt{\sum _{i=1}^n{(Y_i-\overline{Y})}^2}}}$$

(7)

where $X_i$ is the GV index at the sampling point $i$; $Y_i$ is the corresponding nitrate concentration; $\overline{X}$ and $\overline{Y}$ are the mean values of GV and NO₃⁻, respectively; and $n$ is the total number of samples.

A total of 133 groundwater samples were used to assess the correlation between NO₃⁻ and the GV scores. Since the vulnerability indices generated by the COP and COPK model represented the protective capacity of the aquifer (i.e., higher values indicate lower GV), an inverse transformation of NO₃⁻ concentration was applied prior to validation. Specifically, the adjusted NO₃⁻ value $N_i^{\prime }$ for each groundwater sample was calculated as follows:

$$N_i^{\prime }=N_{max}-N_i$$

(8)

where $N_i$ is the original NO₃⁻ of sample $i$, and $N_{max}$ is the maximum NO₃⁻ concentration among all samples. This transformation ensures that higher adjusted NO₃⁻ corresponds to a higher GV, aligning with the model’s output logic.

3.5. Spatial Autocorrelation

Spatial autocorrelation was applied to examine the spatial structure and clustering characteristics of GV maps. It quantifies the degree of similarity between attribute values at neighboring spatial units [71,72]. In this study, both global and local spatial autocorrelation methods were employed. The global Moran’s I index was used to assess the overall spatial correlation across the study area, and local spatial autocorrelation analysis was further used to detect local spatial clusters. A positive Moran’s I value suggests spatial clustering of similar values, a negative value implies spatial dispersion, and a value near zero indicates a random spatial pattern [73,74]. The global spatial autocorrelation index is calculated as follows:

$$Moran’s I={\displaystyle \frac{n{\sum }_{i=1}^n{\sum }_{j=1}^n{W}_{ij}(x_i-\overline{x})(x_j-\overline{x})}{{\sum }_{i=1}^n{\sum }_{j=1}^n{W}_{ij}\left({\sum }_{i=1}^n{(x_i-\overline{x})}^2\right)}}$$

(9)

where $n$ is the total number of spatial units; $x_i$ and $x_j$ are the attribute values at cells $i$ and $j$, respectively; $\overline{x}$ is the mean of all values; and $W_{ij}$ is the spatial weight between units $i$ and $j$.

3.6. Groundwater Risk Calculation

To assess groundwater risk in the study area, seven categories of PPSs were identified, including industrial pollution sources, mining areas, hazardous waste disposal sites, landfills, gas stations, agricultural pollution sources, and golf courses. The identification of PPSs’ categories and their calculation was based on previous studies [75,76,77]. The pollution load risk index for an individual groundwater pollution source was calculated using Equation 10. The details of the $P_i$ calculation are shown in Appendix C.

$$P_i=T_i·L_i·Q_i$$

(10)

where $P_i$ is the pollution load risk index for pollution source $i$; $T_i$ is the toxicity of pollutants associated with source $i$; $L_i$ is the likelihood of release from source $i$; and $Q_i$ is the quantity of pollutants potentially released.

The final pollution source load index map was generated by overlaying the calculated pollution loads of all PPS categories. The comprehensive PI was calculated using the following equation:

$$PI=\sum _{i=1}^n{W}_i·P_i$$

(11)

where $PI$ represents the comprehensive pollution load index at a given location, and $W_i$ is the weight assigned to the $i^{th}$ category of pollution source (

Table 5. Weights for seven categories of PPSs.

No.	Classification	Score
1	Industrial pollution sources	5
2	Mining areas	5
3	Hazardous waste disposal sites	3
4	Landfills	4
5	Gas stations	3
6	Agricultural pollution sources	2
7	Golf courses	1

).

The final groundwater risk was assessed by integrating the results of the GV index and the PI. A matrix-based approach was employed to combine the five classes of GV with the five classes of pollution load, resulting in five levels of groundwater risk. The risk levels were classified as follows: very high, high, moderate, low, and very low.

4. Results

4.1. GV Map

Figure 9 shows the GV mapping results derived from the COP and COPK models, respectively. The COP model was classified based on the method proposed by Vías et al. [35], while the COPK model was classified using the natural breaks method. Overall, the COPK model had a large proportion of very high (27.73%) and high (27.52%) GV zones compared to the COP model, while the proportions of low (15.03%) and very low (6.33%) zones were significantly smaller than those in the COP model. Also, NO₃⁻ sampling points were shown on both GV maps. A general consistency was observed between the spatial distribution of high NO₃⁻ concentrations and the areas classified as highly vulnerable, supporting the validity of the model outputs.

4.2. Model Validation Results

To evaluate the performance of the GV models, both the COP and the modified COPK models were compared against transformed NO₃⁻ concentrations. As shown in Figure 10, scatter plots with linear regression fits were generated for each model, and PCCs were calculated to quantify the correlations. The results indicate a positive correlation between the two models and NO₃⁻ concentrations, which confirms the reasonability of model construction. The very small p-value (p < 0.001) implies that the result is robust and unlikely to result from random noise. Notably, the $r$ improved from 0.3689 in the COP model to 0.4388 in the COPK model, which suggests that the newly introduced indicators and scoring processes such as distance to UURT, LULC, and karst development significantly enhanced the model’s accuracy in urbanized karst areas. The modified COPK model is considered more suitable for this study area and can serve as a reliable basis for future groundwater risk assessments. However, it should be noted that even after these improvements, the $r$ value remains below 0.5, which may indicate the significant influences of human activities in the study area.

4.3. Spatial Autocorrelation Results

Table 6. Global Moran’s I calculation.

Model	Global Moran’s I	p-Value	Z
COP	0.9171	<0.001	16,785.4622
COPK	0.8739	<0.001	16,306.2645

shows the results of global Moran’s I for the COP and COPK models. The two models have strong and statistically significant positive spatial autocorrelation, which indicates the obvious and robust spatial clustering of the GV index (p-value < 0.001). The COP model yielded a Moran’s I value of 0.9171, while the COPK model resulted in a slightly lower value of 0.8739. Although the spatial autocorrelation of COPK is marginally weaker, it remains high and significant. The slight reduction in spatial autocorrelation may be attributed to the increased heterogeneity introduced by UURT, LULC, and karst development, which better differentiates localized vulnerability.

Figure 11 shows the results of local Moran’s I analysis for the COP model and the COPK model. Both models display significant spatial clustering, dominated by High–High (HH) and Low–Low (LL) types. Compared to the COP model, the COPK model exhibits a more dispersed pattern, reflecting the impact of UURT and LULC on local spatial structure. Additionally, HH clusters account for over 50% of the area in both models, confirming strong local spatial autocorrelation.

4.4. Groundwater Risk

Figure 12 presents the spatial distribution of groundwater risk across the study area, based on the integration of GV and pollution load. Figure 12a shows the spatial distribution of the PI. Overall, PI in Guiyang is relatively low and is scattered across the region in a fragmented pattern. The PI layer and the COPK vulnerability index were combined using the risk matrix approach, as shown in Figure 12c. Finally, a groundwater risk map was created in Figure 12d, including five levels: very low, low, moderate, high, and very high, classified by the natural breakpoint method. Overall, the groundwater risk in Guiyang is relatively low. Moderate- to high-risk areas are primarily concentrated in the central part of Guiyang City, and a closer comparison reveals a notable spatial correspondence with the distribution of the buffers of UURT. This alignment strongly supports the effectiveness of the UURT as an important indicator in capturing the impact of urban infrastructure on GVA.

5. Discussion

5.1. Effectiveness of the Modified COPK Model Under the Urbanization Context

In this study, the traditional COP model was modified and enhanced by introducing three additional indicators: UURT, LULC, and K. Specifically, UURT and LULC were integrated into the C and O factors with differentiated scoring approaches, allowing for a more precise assessment of anthropogenic impacts. Compared with existing modifications that primarily focus on improving the K factor [46,47,48], the modified COPK model offers a novel and comprehensive improvement, confirming that UURT, LULC, and K are essential for GVA in karst regions. Previous studies have pointed out that UURT can disrupt the natural surface cover and facilitate the rapid infiltration of pollutants into aquifers through construction-related fractures and subsurface disturbances [51,52,78]. Although modern UURTs are typically equipped with engineered drainage facilities, it is nearly impossible to completely avoid surface and subsurface disturbances [79,80]. Thus, incorporating UURT into GVA models provides a more realistic representation of urbanization impacts on groundwater conditions. Likewise, LULC has been widely recognized as a key factor in GVA under urbanization [81]. While in porous media settings, models such as DRASTICL have proven the value of the model improvement with the involvement of LULC [82,83,84,85], its application in karst GV assessments remains limited. Li et al. [46] incorporated the LULC indicator into the COP model to assess GV across karst regions in Guizhou Province, which was a valuable attempt. Building upon this, this study considers LULC in both the C (flow concentration) and O (overlying protective layer) factors, which proved to be highly effective in enhancing GV model accuracy. Additionally, we also noticed that the $r$ values of two models were both below 0.5, which was consistent with the relatively low $r$ in many previous studies [22,86,87,88]. This outcome highlights the substantial impact of anthropogenic activities, particularly the clustered pollution sources in the study area, which can obscure or weaken the statistical relationship between vulnerability and observed nitrate contamination [89,90]. Therefore, three major recommendations are derived from this work. Firstly, UURT and LULC should be systematically incorporated into the karst GVAs under the urbanization context. Secondly, urban rail development should prioritize the preservation of surface integrity and hydrological function to minimize the risk of contaminant migration, especially in vulnerable KASs. Finally, taking into account the impact of more complete human activities on GVA models remains a worthy direction for future research.

5.2. Analysis of Groundwater Risk Assessment

The groundwater risk assessment conducted in this study effectively integrated GVA and pollution loads, and the use of a risk matrix approach allowed for a structured classification of risk levels, which aligns with widely accepted practices in sustainable groundwater management. In recent years, such integrated approaches have been widely adopted in vulnerability risk modeling, as demonstrated in the works of Li et al. [91], Zhang et al. [92], and Li et al. [93], all of which successfully combined GV with pollution loading to derive spatial risk distributions. Compared to these studies, the present work offers a more comprehensive consideration of PPS categories. Specifically, seven major PPSs were considered in PI evaluation, including industrial pollution, mining activity location, hazardous waste disposal site, landfill, gas station, agricultural pollution, and golf courses. For each PPS category, three factors (pollutant toxicity, release probability, and expected contaminant quantity) were independently evaluated and integrated to generate the composite PI, as shown in Appendix C. This evaluation process was guided by the existing literature [75,76,77]; however, a standardized framework for PI calculation has yet to be established. Therefore, we recommend that future research should formalize the PI evaluation process and promote the application of the risk matrix approach used in this study. Also, from a management perspective, areas identified as high risk should be prioritized for monitoring, pollution source control, and infrastructure planning. The findings reinforce the importance of integrated risk assessment models in supporting proactive and spatially targeted groundwater protection strategies.

5.3. Urban Planning and Infrastructure Development for Sustainable Groundwater Management

This study presents a case-specific demonstration of how urban infrastructure (specifically, UURT) can be effectively incorporated into GVA to better reflect real-world anthropogenic pressures. While UURT represents one major component of urban development, it is important to recognize that a wide range of infrastructure projects, including road networks [94], land reclamation [95], and drainage systems [96], can also significantly alter hydrological conditions and increase risks to groundwater systems. The integration of UURT into the COP vulnerability framework in this study serves as a good example, illustrating the potential to incorporate other urbanization-related indicators into GV models. In the context of urbanization, GV should be viewed as a core constraint in spatial planning, particularly in ecologically sensitive areas such as karst regions where the aquifer system is highly vulnerable to disturbance. From a broader perspective, sustainable groundwater management requires a proactive and spatially informed planning approach, where vulnerability maps are used not only as scientific outputs, but as decision-support tools for urban planners, environmental regulators, and infrastructure designers. By identifying high-risk zones before construction, cities can implement preventive measures such as impermeable zone control, buffer zone planning, enhanced drainage designs, and pollution source zoning.

5.4. Limitations and Future Research

While this study presents an improved COP-based framework by incorporating UURT, LULC, and K to assess GV and risk in urbanized karst regions, several limitations should be acknowledged. First, the scoring and classification of the newly introduced indicators (particularly in UURT and LULC) still rely partly on expert judgment and semi-quantitative methods. Although spatial accuracy has been improved, the subjectivity in scoring may lead to uncertainty. Second, the model currently assumes static conditions for LULC and PPSs, which may not fully capture the temporal dynamics of urban development and contamination processes. Future studies should consider incorporating time-series data, such as changes in LULC, urban expansion stages, or seasonal nitrate variability, for more accurate GV model construction. Third, although the study area represents a typical urbanized karst region, the transferability of the model to other karst systems with different conditions requires further validation. In particular, more comparative applications across varying urbanization gradients and karst morphologies could enhance the generalizability of the approach. Fourth, this study focuses on nitrates as the primary validation parameter. While nitrate is widely accepted as a representative pollutant in vulnerability assessments, future research should expand validation efforts to include multi-contaminant indicators, such as heavy metals, organics, or microbial risks, to provide a more comprehensive evaluation of model performance under diverse pollution scenarios. Finally, future research should also explore coupling vulnerability risk models with decision-making systems, integrating them into urban planning platforms or environmental early-warning mechanisms, to facilitate real-time responses to emerging groundwater threats.

6. Conclusions

This study proposed an improved GV risk assessment framework designed for urbanized karst regions, using Guiyang City as a case study. The traditional COP model was enhanced by incorporating three critical indicators (UURT, LULC, and K), which were combined into the C and O factors to form the modified COPK model. This modification significantly improved the model’s performance and its correlation with observed transformed NO₃⁻ concentrations ($r$ = 0.4388) was higher than the COP model ($r$ = 0.3698). A comprehensive PI was constructed by integrating seven types of PPSs, including industrial, mining, hazardous waste disposal, landfills, gas stations, agricultural land, and golf courses. The COPK-based GV map and the pollution load map were combined using a risk matrix approach, resulting in a detailed spatial classification of groundwater risk into five levels. Moderate- to high-risk areas were primarily concentrated in the central part of Guiyang City, and a closer comparison revealed a notable spatial correspondence with the distribution of the buffers of UURT.

Overall, this work highlights the importance of integrating infrastructure-related indicators into GV models under urbanization contexts. The case of UURT reflects how surface disruption and subsurface alteration can exacerbate GV. Meanwhile, LULC proves to be a critical dual-role factor that influences both runoff concentration and protective-layer characteristics. This framework not only improves the scientific basis for groundwater risk assessment in karst regions, but also provides operational tools for urban planning, land use zoning, and environmental decision-making. The proposed approach offers high applicability and can be extended to other rapidly urbanizing karst areas worldwide.