Analysis of the Groundwater Quality Evolution and Pollution Source Identification over Multiple Periods in an Industrial Zone Based on the Hydrogeochemical-PMF Model

Abstract

Groundwater quality degradation in industrial zones presents a critical environmental challenge, as diverse and compositionally complex pollution sources severely constrain precise source identification and the formulation of effective remediation strategies. This study systematically investigates groundwater quality evolution and quantitatively apportions pollution sources in a representative industrial zone in southwestern China, employing an integrated framework of hydrochemical graphical analysis, dual-dimensional hierarchical cluster analysis, and Positive Matrix Factorization (PMF) receptor modeling, based on 189 groundwater samples collected across three hydroperiods (2022–2025) from 94 monitoring wells. (1) The study reveals that overall groundwater quality was unsatisfactory, with Class IV and V waters collectively accounting for 75%, 95%, and 91% across the three campaigns; primary exceedance parameters included ammonia nitrogen, total hardness, Mn, and sulfate. (2) Hydrogeochemical analysis revealed stable HCO₃-Ca type water at background monitoring points, slight contamination influence at diffusion points (occasional HCO₃·SO₄-Ca and Cl·SO₄-Ca types), and pronounced hydrochemical diversification at internal points, evolving from Ca-Cl dominance to the coexistence of Ca-Cl·SO₄, Ca-HCO₃, and other mixed types. (3) The PMF model consistently resolved five pollution sources across all campaigns: agricultural non-point source pollution, geological background, domestic wastewater, industrial emissions, and natural hydrogeochemical evolution. Anthropogenic sources (agricultural, domestic, and industrial) collectively contributed approximately 63.8% of the total contamination load, with individual average contributions of 18.6%, 26.2%, and 19.0%, respectively, indicating that groundwater contamination in the zone is influenced not only by industrial inputs but also substantially by agricultural non-point sources and domestic wastewater. This study reveals a coupled natural–anthropogenic driving mechanism and establishes a replicable integrated framework for pollution source identification and zoned precision management in comparable industrial zone settings.

1. Introduction

Groundwater constitutes the world’s largest accessible store of liquid freshwater, accounting for approximately 99% of the global liquid freshwater reserve [1]. It supplies drinking water to nearly half of the global population and provides roughly 43% of all consumptive agricultural irrigation water [2,3]. In China, groundwater contributes approximately one-third of the total water supply, with annual extraction exceeding 85 billion m³, making China one of the largest groundwater exploiters worldwide [4,5]; in many regions, groundwater serves as the primary—and often the sole—water source for urban, industrial, and agricultural uses [6]. Nevertheless, decades of rapid industrialization, agricultural intensification, and urban expansion have imposed progressively severe stresses on groundwater systems across the country [7,8]. According to the 2025 Bulletin on the State of China’s Ecological Environment, among 1865 national groundwater quality monitoring points, Category V water quality accounted for 23.2% of all sites, with the proportion reaching 25.7% in phreatic aquifers [9]. The principal anthropogenic drivers of groundwater quality deterioration include industrial wastewater discharge, agricultural non-point source pollution (synthetic fertilizers, pesticides, and intensive livestock operations), urban domestic sewage leakage, and leachate from solid waste disposal sites [10,11,12,13]. Among these, industrial zones constitute a particularly high-risk category of pollution sources: they concentrate multiple potentially hazardous activities—raw chemical storage, process wastewater generation, accidental spills, and hazardous waste management—within geographically confined areas, thereby generating spatially superimposed and compositionally complex contaminant mixtures [14,15,16]. The diversity, concealed nature, and temporal variability of pollution sources in such settings render source identification and targeted remediation exceedingly difficult, underscoring the urgent need for systematic source apportionment studies in industrial zone environments [17].

Source apportionment—the systematic identification and quantitative attribution of contributions from distinct pollution sources to a receptor—has become a cornerstone methodology in environmental forensics [18]. Three broad categories of source apportionment approaches have been developed: emission inventory methods, dispersion modeling methods, and receptor modeling methods [19,20]. Emission inventory approaches estimate pollutant loads by constructing detailed source-by-source emission databases; while conceptually straightforward, their accuracy depends critically on the completeness and quality of activity-level data and emission factors, which are frequently unavailable or subject to substantial uncertainty at regional scales [21]. Dispersion modeling methods simulate the transport, diffusion, and transformation of pollutants from emission sources through environmental media to receptor locations; however, their application to groundwater systems faces enormous challenges owing to the pronounced heterogeneity of aquifer hydraulic properties, the complexity of groundwater flow paths, and the difficulty of constraining contaminant attenuation parameters (adsorption, degradation, and dispersion coefficients) across different spatial scales [22]. Receptor models operate on an entirely different principle: starting from the chemical composition measured at the receptor, they employ statistical techniques to decompose the observed concentration matrix into source factor profiles and source contribution matrices, without requiring prior knowledge of source emission rates or pollutant transport mechanisms [18,23]. This operational advantage—inferring causes from effects—has made receptor models the most widely applied source apportionment framework in complex multi-source environmental systems [24].

Within the receptor model family, two broad subcategories have gained prominence in groundwater research: multivariate statistical techniques and factorization-based methods [25,26]. Multivariate statistical approaches include Principal Component Analysis coupled with Multiple Linear Regression (PCA-MLR) and Absolute Principal Component Scores-Multiple Linear Regression (APCS-MLR), which reduce the dimensionality of large water quality datasets by extracting independent orthogonal factors that explain maximum variance, and subsequently regress measured concentrations against these factors to estimate source contributions [27,28]. These methods have been successfully applied across diverse groundwater quality investigations—from the Songnen Plain in northeastern China [27] to the riverside groundwater systems of Beijing [28], and to the Poyang Lake Basin in southern China [29]. However, PCA-based methods have well-recognized limitations: they assume linear relationships among variables and normally distributed data, and the extracted factor scores may yield negative values that lack physical interpretability [30]. Concurrently, cluster analysis—including Hierarchical Cluster Analysis (HCA) and K-means clustering—has been widely adopted as an unsupervised pattern recognition tool for grouping groundwater sampling sites into categories with similar hydrochemical characteristics and for identifying associations among indicators that share common sources or geochemical behaviors [31,32,33]. When applied to multi-temporal datasets, cluster analysis can reveal spatiotemporal evolution patterns of water quality, such as tracking whether pollution ‘hotspots’ persist, migrate, or dissipate across different hydrological seasons [34,35]. Nevertheless, while cluster analysis excels at classification and pattern discovery, it does not inherently provide quantitative source contribution estimates.

Positive Matrix Factorization (PMF), first introduced by Paatero and Tapper in 1994 [36], has emerged as a particularly robust factorization-based receptor model that addresses several key limitations of conventional PCA approaches. Unlike PCA, PMF imposes non-negativity constraints on both factor contributions and factor profiles, yielding physically meaningful results consistent with the reality that source contributions and species concentrations cannot be negative [37]. Furthermore, PMF incorporates sample-specific, species-specific measurement uncertainty information through a weighted least-squares optimization framework, enabling it to appropriately down-weight data points with higher analytical uncertainty [38,39]. Originally developed for atmospheric aerosol source apportionment [40], PMF has increasingly been extended to groundwater quality studies in recent years. Zanotti et al. [41] demonstrated the combined application of PMF with GIS-based spatial analysis for characterizing groundwater and surface water quality in Northern Italy. Zhang et al. [39] employed PMF alongside PCA-APCS-MLR to identify and apportion groundwater pollution sources in a mixed land-use area of southwestern China. Chen et al. [40,41] applied PMF to assess groundwater quality, pollution sources, and associated health risks across the Songnen Plain. Zhu et al. [42] combined PMF with hydrochemical methods to resolve pollution sources in an industrial zone setting. Collectively, these studies have established PMF as an effective and increasingly indispensable tool for groundwater source apportionment [43].

Despite the aforementioned advances, several critical research gaps remain in the application of PMF to groundwater quality studies in industrial zone settings. First, the overwhelming majority of existing groundwater PMF investigations rely on single-period sampling campaigns, yielding only a static water quality snapshot at a particular point in time [44,45]. Such single-snapshot approaches are inherently incapable of capturing the spatiotemporal dynamics of pollution sources or elucidating the hydrogeochemical evolution of groundwater as it migrates along flow paths under varying hydrological conditions [46,47,48]. Second, although classical hydrochemical graphical methods are well-recognized for their value in revealing natural geochemical backgrounds and water–rock interaction processes, their integration with receptor models for the joint interpretation of anthropogenic pollution inputs superimposed on natural hydrogeochemical evolution remains insufficiently developed—few studies have systematically linked hydrochemical facies transitions to PMF-resolved source factor variations across multiple time periods [49,50,51]. Third, while dual-dimensional cluster analysis of both sampling sites and water quality indicators offers a powerful means of resolving spatiotemporal water quality differentiation patterns, its systematic coupling with PMF for cross-validation of source identification results has rarely been attempted. Fourth, existing PMF studies on groundwater have not adequately addressed the dynamic reassignment of indicator contributions across different pollution sources under changing hydrological regimes, nor have they quantitatively characterized the temporal evolution trajectories of pollution source structures, limitations that significantly constrain the model’s ability to inform time-sensitive, adaptive pollution management strategies in industrial zone environments.

To address the above research gaps, this study presents an integrated analytical framework that couples multi-period hydrochemical characterization, dual-dimensional hierarchical cluster analysis (of both sampling sites and water quality indicators), and PMF receptor modeling to investigate groundwater quality evolution and pollution source contributions in a representative industrial zone in southwestern China. The study is based on three systematic sampling campaigns encompassing 189 groundwater samples collected from 94 monitoring wells arranged along a gradient design. The specific research objectives are: (1) to assess the overall environmental status of groundwater in the study area and identify current pollution conditions; (2) to elucidate hydrogeochemical evolutionary patterns and spatiotemporal water quality differentiation through hydrogeochemical analysis and dual-dimensional hierarchical cluster analysis; (3) to apply the PMF receptor model in quantifying the relative contributions of distinct pollution sources and to track the temporal evolution trajectories of source structures across the three hydroperiods; and (4) to synthesize the above findings into a unified source identification framework that informs zoned, source-specific groundwater pollution prevention and control strategies for industrial zone.

2. Materials and Methods

2.1. Study Area Description

2.1.1. Geographic and Geological Setting

The study area is located in the southwestern plateau region of China, situated within a faulted basin belonging to the Nanpan River watershed of the Xijiang (Pearl River tributary) drainage system. Influenced by Cenozoic tectonic movements, the area exhibits well-developed folds and faults, forming complex terrain characterized by alternating faulted basins and karst mountains deeply incised by the Nanpan River and its tributaries. The overall topography slopes from southwest to northeast (Figure 1a), with elevations ranging from 1040 to 1120 m above sea level, classified as moderately to shallowly dissected high-mid-mountain plateau topography. The area experiences a subtropical plateau monsoon climate, with a mean annual temperature of 19.8 °C. Mean annual precipitation is approximately 800 mm, with pronounced spatiotemporal heterogeneity; the rainy season (May–October) accounts for over 82% of the annual total. Mean annual water surface evaporation is 1466 mm, with an annual aridity index of 1.8, indicating that evaporation far exceeds precipitation and that the area is characterized by significant aridity. A south-to-north flowing surface river traverses the central portion of the study area, with a reach length of approximately 25 km, a drainage area of 331 km², a mean annual discharge of 10.64 m³/s (maximum: 573 m³/s; minimum: 0.87 m³/s), and a mean annual runoff of 3.91 × 10⁸ m³. This river serves as the primary water source for regional industrial, agricultural, and domestic uses while simultaneously receiving groundwater discharge.

2.1.2. Hydrogeological Conditions

The study area is underlain predominantly by Quaternary Holocene alluvial-pluvial deposits (Q₄^al+pl), dolomite interbedded with shale, limestone intercalated with siltstone, and argillaceous sandstone (Figure 2). Contaminants generated at the surface infiltrate downward, directly impacting Quaternary unconsolidated sediment pore groundwater. The area comprises two adjacent hydrogeological units—eastern and western—both discharging toward the central surface river. Groundwater types primarily include unconsolidated sediment pore water and carbonate karst water. The depth to the groundwater table ranges from 1.4 to 15.5 m, with water level elevations between 1030 and 1080 m (Figure 1b). The water table is generally shallow, and groundwater is primarily recharged by atmospheric precipitation and surface water infiltration, exhibiting pronounced seasonal variability in both water storage and groundwater level dynamics.

2.1.3. Industrial Zone Overview

The study area encompasses a planned area of 17.7 km², with a built-up area of 3.31 km² (Figure 3). Sixteen enterprises have been established within the built-up area, with dominant industries encompassing food manufacturing, chemical raw materials and chemical products manufacturing, rubber and plastic products, non-metallic mineral products, fabricated metal products, comprehensive utilization of waste resources, and electric power generation. Based on environmental risk classification, four enterprises are designated as key pollution sources and twelve as non-key pollution sources.

2.2. Sample Collection and Analytical Methods

2.2.1. Sampling Design

Three systematic groundwater sampling campaigns were conducted across different hydroperiods: the first campaign (dry season, 2022) collected 16 samples; the second campaign (dry season, 2024) collected 94 samples; and the third campaign (wet season, 2025) collected 79 samples, yielding a total of 189 groundwater samples from 94 monitoring wells. It is worth noting that the monitoring wells in the first campaign were largely damaged due to long time intervals and social construction activities. As a result, sampling could not be conducted in the second and third campaigns. The monitoring wells used in the second campaign were basically sampled in the third stage. From the point distribution map (Figure 1), it can be seen that the monitoring wells in the second and third campaigns were basically densely arranged around the points of the first campaign, meaning they can represent the groundwater quality conditions of the damaged monitoring wells in the first campaign. The monitoring network was designed along a gradient configuration encompassing three zones: (1) comparison points located in the upgradient periphery of the industrial zone, representing regional natural groundwater quality; (2) internal points situated within the core industrial zone, directly reflecting the influence of industrial activities; and (3) diffusion points positioned in the downgradient periphery, capturing the spatial extent of contaminant plume migration.

2.2.2. Analytical Parameters and Methods

Groundwater samples were collected using 500 mL high-density polyethylene (HDPE) bottles. Before sampling, each well was purged of a minimum of three casing volumes (or until pH, temperature, and electrical conductivity stabilized), and the stabilized field parameters were recorded. Prior to sampling, bottles were rinsed three times with deionized water and subsequently flushed three times with the groundwater to be sampled at each site. Samples for dissolved constituents were filtered in the field through 0.45-µm membrane filters and preserved as follows: samples for cations and trace metals were acidified to pH < 2 with ultra-pure HNO₃; samples for anions were kept unpreserved and refrigerated; samples for ammonia nitrogen and nutrients were preserved with H₂SO₄ to pH < 2; and samples for petroleum hydrocarbons and volatile organics were collected without headspace and stored at 4 °C in amber glass vials. Samples were collected, immediately sealed, and transported to the laboratory for analysis within 48 h. Field-measured parameters included pH, dissolved oxygen (DO), oxidation-reduction potential (Eh), and electrical conductivity, determined using a multiparameter water quality meter (In-situ Smartroll, Fort Collins, CO, USA). Ca²⁺ and HCO₃⁻ were titrated on-site using hardness and alkalinity test kits. Laboratory analyses were performed as follows: cations were determined using an inductively coupled plasma optical emission spectrometer (IRIS Intrepid II XSP, Thermo Electron, Waltham, MA, USA) with an analytical precision within 1%; anions were determined using an ion chromatograph (ICS-2100, Dionex, Sunnyvale, CA, USA) with charge balance errors below 8%; heavy metals were analyzed using inductively coupled plasma mass spectrometry (ICP-MS, Elan DRC-E, PerkinElmer, Shelton, CT, USA) with relative standard deviations within ±5% and spike recoveries of 90–110%; total hardness was determined by Na₂EDTA titration. All sample collection, preservation, and pretreatment procedures strictly followed national technical specifications. Method detection limits (MDLs) for all parameters are listed in

Table 1. Statistical results of different types of groundwater detection indicators in the first period (April 2022).

Indicators	MDL	Comparison Point (n = 2)					Internal Point (n = 4)					Diffusion Point (n = 10)
		Min	Max	Ave	SD	CV	Min	Max	Ave	SD	CV	Min	Max	Ave	SD	CV
TH	0.5	244	458	351	151.30	0.43	639	2152	1063	686.60	0.65	84.6	719	497.3	193.70	0.39
TDS	8.6	256	564	410	217.80	0.53	921	3632	1717	1183.00	0.69	118	1113	713.8	300.90	0.42
pH	/	6.87	7.23	7.05	0.26	0.04	7.05	7.42	7.283	0.15	0.02	6.37	7.56	7.315	0.33	0.04
NO3	0.015	3.23	11.3	7.265	5.71	0.79	3.22	175	63.22	74.11	1.17	0.185	31.5	8.165	10.25	1.26
Cl	0.007	14.4	37.4	25.9	16.26	0.63	45.3	1332	380.8	561.20	1.47	3.02	259	71.03	77.17	1.09
F	0.004	0.172	0.26	0.216	0.06	0.29	0.238	0.612	0.395	0.14	0.37	0.154	1.02	0.324	0.24	0.73
SO4	0.017	1.05	144	72.53	101.10	1.39	340.8	406	380.9	29.94	0.08	0.606	274	176.8	96.55	0.55
I	0.002	0.005	0.005	0.0035	0.00	0.61	0.017	0.0765	0.03638	0.03	0.69	0.01	0.1158	0.03446	0.04	1.14
Ca	0.01	39.2	154	96.6	81.17	0.84	222	608	334	176.00	0.53	28.6	275	170.8	70.26	0.41
Fe	0.00055	0.14	0.285	0.2125	0.10	0.48	0.0507	0.544	0.202	0.21	1.03	0.0262	1.88	0.3256	0.56	1.72
Al	0.00096	0.0197	0.0518	0.03575	0.02	0.64	0.0153	0.0869	0.05116	0.03	0.56	0.0143	0.353	0.1068	0.10	0.97
Mn	0.0001	0.0295	0.246	0.1378	0.15	1.11	0.305	1.833	0.7516	0.67	0.89	0.00691	2.27	0.9676	0.79	0.82
VP	0.0002	ND	ND	/	/	/	ND	ND	/	/	/	0.004975	0.004975	0.000588	0.00	2.60
AS	0.02	0.06	0.08	0.07	0.01	0.20	0.035	0.51	0.2263	0.21	0.95	0.02	0.14	0.064	0.03	0.52
An	0.02	0.032	0.032	0.026	0.01	0.33	0.059	192	48.13	83.46	1.73	0.027	7.44	1.603	2.42	1.51
OC	0.4	0.6	2.4	1.5	1.27	0.85	1.2	4.1	2.119	1.26	0.59	0.4	4.2	1.33	1.12	0.84
Hg	0.00001	0.00007	0.00007	0.00004	0.00	1.06	0.00003	0.00004	0.000035	0.00	0.17	0.00003	0.00008	0.000042	0.00	0.40
TP	0.004	0.06	0.2	0.13	0.10	0.76	0.08	0.223	0.1583	0.07	0.42	0.03	0.4	0.1227	0.10	0.83
Ph	0.01	0.12	1.37	0.745	0.88	1.19	0.04	0.8	0.4113	0.37	0.90	0.04	0.22	0.089	0.06	0.62
Ti	0.00002	ND	ND	/	/	/	0.00002	0.00055	0.0001588	0.00	1.54	0.00003	0.00018	0.000054	0.00	0.95
As	0.00015	ND	ND	/	/	/	0.00015	0.00015	0.00015	0.00	0.00	0.00025	0.003	0.000435	0.00	2.05
Ni	0.00005	0.00068	0.0136	0.00714	0.01	1.28	0.00128	0.00277	0.001839	0.00	0.31	0.0001	0.015	0.002282	0.00	1.71
V	0.00008	0.00024	0.0003	0.00027	0.00	0.16	0.00032	0.00141	0.000848	0.00	0.53	0.00038	0.00195	0.001003	0.00	0.54
TN	0.03	3.18	10.6	6.89	5.25	0.76	4.24	335	105.2	146.20	1.39	0.42	33.6	10.68	11.22	1.05

For subsequent calculations. Descriptive statistics for each campaign are presented in Table 1 and Tables S1 and S2 in the Supplementary Information. Groundwater samples were collected, preserved, transported, and analyzed in accordance with the Technical specifications for environmental monitoring of groundwater (HJ/T 164-2020).

. Field blanks, travel blanks, and duplicate samples (≥10% of the sample number per campaign) were analyzed to assess contamination and reproducibility; all blanks were below the MDL. Non-detect values were substituted with half the method detection limit.

2.3. Data Analysis Methods

2.3.1. Water Quality Assessment

Groundwater quality was evaluated against the Standards for Groundwater Quality (GB/T 14848-2017) [52] to determine compliance status for each parameter. The comprehensive water quality category for each sampling point was assigned based on the worst-performing individual parameter. Exceedance rates for each parameter were calculated to identify the primary parameters exceeding regulatory thresholds and to characterize their spatiotemporal distribution patterns.

2.3.2. Hydrogeochemical Analysis

Based on the multi-campaign monitoring data, Piper trilinear diagrams, Durov diagrams, and Gibbs diagrams were constructed to analyze groundwater hydrochemical facies, compositional evolutionary trends, and their controlling mechanisms. Piper diagrams were employed to visualize the relative abundances of major cations and anions and to classify hydrochemical facies. Durov diagrams extend the Piper representation by incorporating TDS and pH as additional dimensions, thereby better revealing hydrochemical evolutionary pathways. Gibbs diagrams, relating TDS to Na⁺/(Na⁺ + Ca²⁺) and Cl⁻/(Cl⁻ + HCO₃⁻) ratios, were used to discriminate the dominant natural mechanisms governing groundwater chemical composition: precipitation dominance, rock weathering dominance, or evaporation-crystallization dominance. Additionally, the relationship between γ(Na⁺ − Cl⁻) and γ[(HCO₃⁻ + SO₄²⁻) − (Ca²⁺ + Mg²⁺)] was examined to elucidate water–rock interaction processes, including cation exchange and silicate weathering, that influence groundwater chemical composition.

2.3.3. Cluster Analysis

Cluster analysis (CA) is an unsupervised pattern recognition method that groups objects or variables sharing similar characteristics based on similarity (or distance) metrics, thereby revealing the underlying structure and classification of the data. In this study, CA was conducted using Euclidean distance as the similarity measure and Ward’s minimum variance method as the linkage rule. Cluster analysis was performed in two dimensions: (i) water quality indicators, to reveal intrinsic associations among parameters, and (ii) sampling sites, to delineate spatial classification patterns of monitoring points. Prior to clustering, data were standardized using Z-score transformation to eliminate scale differences among variables. Clustering results were visualized as circular dendrograms.

2.3.4. Positive Matrix Factorization (PMF)

The PMF model, first proposed by Paatero and Tapper in 1994, is a multivariate factor analysis receptor model. Its fundamental principle involves decomposing the sample concentration matrix X (n × m) into the factor contribution matrix G (n × p), the factor profile matrix F (p × m), and the residual matrix E (n × m). The underlying mathematical formulation is:

$$X_{(n\times m)}=G_{(n\times p)}{F}_{(p\times m)}+E_{(n\times m)}$$

(1)

where X is the sample concentration matrix, G is the factor contribution matrix, F is the factor concentration matrix, E is the residual matrix, n is the number of samples, m is the number of chemical species, and p is the number of resolved factors (pollution sources). The relationship between observed, predicted, and residual concentrations is expressed as:

$$e_{ij}=x_{ij}-{\displaystyle \sum _{k=1}^p{g}_{ik}{f}_{kj}}$$

(2)

where i = 1, 2, …, n; j = 1, 2, …, m; k = 1, 2, …, p; e_ij is the residual for the i-th sample and j-th species; x_ij is the observed concentration of the j-th species in the i-th sample; g_ik is the contribution of the k-th factor to the i-th sample; and f_kj is the concentration of the j-th species in the k-th factor profile.

The matrices G (n × p) and F (p × m) are obtained by minimizing the objective function Q, which serves as a measure of model fit, defined as follows:

$$Q={\displaystyle \sum _{i=1}^n{\displaystyle \sum _{j=1}^m{(e_{ij}/u_{ij})}^2}}$$

(3)

$$Q_{robust}={\displaystyle \sum _{i=1}^n{\displaystyle \sum _{j=1}^m(e_{ij}/(h_{ij}{u}_{ij}))}}$$

(4)

where h_ij = 1 when the residual exceeds a specified threshold and is otherwise calculated based on the uncertainty-weighted residual; u_ij represents the uncertainty associated with the measured concentration of species j in sample i.

Accurate determination of uncertainties is therefore critical. In this study, uncertainties were calculated as follows:

$$u_{ij}={\displaystyle \frac{5}{6}}MDL$$

(5)

$$u_{ij}=\sqrt{{(EF\times x_{ij})}^2+{(0.5MDL)}^2}$$

(6)

where MDL is the method detection limit for each species and EF (Error Fraction) is the error coefficient. When Conc ≤ MDL, Equation (5) is employed; when Conc > MDL, Equation (6) is used. Error fractions typically range from 0.1 to 0.6. For species with unstable concentrations or those approaching the detection limit, larger error fractions are assigned. Given that certain indicator concentrations in the study area exhibit substantial fluctuations, some approach detection limits, and the first campaign contains a small number of samples, EF was set to the upper bound of 0.6, which conservatively down-weights high-noise and near-detection-limit variables and thereby prevents unstable variables from dominating the solution. A sensitivity analysis in which EF was varied from 0.1 to 0.6 confirmed that the resolved number of factors, the identity of the dominant tracer of each factor, and the rank order of source contributions were essentially unchanged, indicating that the source apportionment is robust to the uncertainty model.

PMF model execution was performed using USEPA PMF 5.0 software. For each sampling campaign, water quality indicators with clear environmental significance and detectable exceedances were selected as model input variables. Factor numbers ranging from 3 to 6 were tested through multiple model runs. The optimal number of factors was determined by jointly evaluating the trends in Q values across factor numbers, scaled residual distributions, and the environmental interpretability of the dominant contributing species combinations within each factor. Finally, based on the species loading characteristics and spatial contribution distributions of each factor, in conjunction with land-use types, industrial layout, and hydrogeological conditions of the study area, each factor was assigned to a specific pollution source category, and the relative contribution of each source was quantified.

3. Result

3.1. Groundwater Quality Assessment

Based on the limit values of Class IV water standards in the Groundwater Quality Standards (GB/T 14848-2017), the water quality of the third-phase monitoring data in the study area was evaluated. The results are shown in Figure 4 and Figure S1 in the Supplementary Information. From the overall water quality perspective, the groundwater quality of the three phases in the study area is not optimistic. In the first phase, 25% of the 16 samples had Class IV water quality, 50% had Class V water quality, and only a few control monitoring points reached Class III or above water quality; in the second phase, 43% of the 94 samples had Class IV water quality and 52% had Class V water quality; in the third phase, 40% of the 79 samples had Class IV water quality and 51% had Class V water quality. Compared to the second phase during the dry season, the water quality has slightly improved, but it is still poor overall.

From the main exceeded indicators, the main exceeded indicators of the three phases are generally consistent, mainly including Ammonia nitrogen (An), Total dissolved solids (TDS), Total hardness (TH), Volatile phenols (VP), Total phosphorus (TP), Petroleum hydrocarbons (Ph), Total nitrogen (TN), Oxygen consumption (OC), Anionic surfactant (AS), pH, SO₄, Mn, Cl, F, Ca, NO₃, As, Al, I, V, Fe, Ti, Ni, Hg, etc. There are 24 indicators. Among them, the problems of exceeding An, TH, SO4, and Mn are the most prominent. The sample exceeding rates of An and TH in all three phases remain above 15%, and the exceeding rates of An and TH even exceed 25%.

Analyzed by monitoring point type, in Period 1, 0 comparison points, 5 diffusion points, and 3 interior points exceeded the Category IV threshold, accounting for 0%, 50%, and 75% of their respective type totals. In Period 2, 2 comparison points, 24 diffusion points, and 23 interior points exceeded Category IV, representing 40%, 57%, and 49% of their respective type totals. In Period 3, 0 comparison points, 19 diffusion points, and 21 interior points exceeded Category IV, accounting for 0%, 59%, and 49% of their respective type totals. These statistics demonstrate that groundwater quality in interior and diffusion monitoring zones was significantly inferior to that in background zones, with interior points exhibiting the highest exceedance proportions, directly reflecting the severe impact of industrial agglomeration activities on local groundwater quality.

3.2. Hydrogeochemical Characteristics and Evolution

Piper trilinear diagrams for the study area across different hydroperiods are presented in Figure 5. During Period 1 (2022 dry season), hydrochemical facies were relatively concentrated, predominantly HCO₃-Ca and HCO₃-Ca·Mg types, with a subset of internal points exhibiting HCO₃·SO₄-Ca and Cl·SO₄-Ca types, indicating that groundwater chemical composition during this period was primarily controlled by natural water–rock interactions such as carbonate dissolution, although select locations already exhibited perturbations from anthropogenic inputs. In Period 3 (2025 wet season), the distribution of sample points in the Piper diagram broadened substantially, with more diversified hydrochemical facies. In the cation ternary plot, the predominance of alkaline earth metals (Ca²⁺ + Mg²⁺) weakened slightly, with some samples shifting toward the Na⁺ + K⁺ apex, indicative of enhanced cation exchange and/or increased sodium source inputs. In the anion ternary plot, weak acid anions (HCO₃⁻) remained dominant, but the proportion of strong acid anions (SO₄²⁻ + Cl⁻) increased markedly in samples from the agglomeration interior and downstream diffusion zones. This evolutionary trend reveals the superimposed modification of groundwater chemical composition by industrial activities and agricultural non-point source inputs superimposed on natural hydrogeochemical evolution.

Durov diagrams (Figure 6) further elucidated hydrochemical evolutionary pathways. During the 2022 dry season, sample points were predominantly concentrated in the lower-left region of the Durov diagram (neutral pH, low TDS, HCO₃⁻- and Ca²⁺-dominated), representative of typical recharge-zone to flow-path carbonate dissolution-dominated hydrochemical signatures. By the 2025 wet season, a subset of samples migrated toward the upper-right region, with elevated TDS and pH and increased contributions from SO₄²⁻ and Cl⁻, indicating enhanced evaporative concentration and anthropogenic inputs. Internal points within the agglomeration exhibited a ‘dispersed’ distribution pattern in the Durov diagram, deviating markedly from natural evolutionary trends, further corroborating the intense perturbation of local groundwater chemical composition by anthropogenic activities.

Gibbs diagram analysis (Figure 7) indicated that the overwhelming majority of sample points fell within or near the upper boundary of the rock weathering dominance field, with TDS values ranging from 200 to 3000 mg/L, Na⁺/(Na⁺ + Ca²⁺) ratios predominantly below 0.5, and Cl⁻/(Cl⁻ + HCO₃⁻) ratios mostly below 0.3. This distribution confirms that water–rock interaction constitutes the dominant natural process controlling groundwater chemical composition in the study area. However, during the 2025 wet season, a subset of samples shifted toward the evaporation-crystallization dominance field, accompanied by elevated Cl⁻/(Cl⁻ + HCO₃⁻) ratios, indicating that the superimposed influence of anthropogenic inputs on groundwater chemical composition cannot be disregarded.

The relationship between γ(Na⁺ − Cl⁻) and γ[(HCO₃⁻ + SO₄²⁻) − (Ca²⁺ + Mg²⁺)] (Figure 8) was employed to resolve major ion sources and water–rock interaction types. During the 2022 dry season, most samples exhibited (Na⁺ − Cl⁻) values close to zero or slightly positive, indicating the combined effects of halite dissolution and minor silicate weathering. During the 2025 wet season, a subset of samples showed substantially positive (Na⁺ − Cl⁻) values, indicating enhanced cation exchange or sodium silicate mineral weathering; other samples displayed negative [(HCO₃⁻ + SO₄²⁻) − (Ca²⁺ + Mg²⁺)] values, suggesting the presence of Ca²⁺ and Mg²⁺ from non-carbonate/sulfate sources (e.g., silicate weathering) or HCO₃⁻ and SO₄²⁻ from non-mineral-dissolution sources (e.g., organic matter degradation and industrial wastewater inputs).

3.3. Cluster Analysis of Groundwater Quality Patterns

Cluster analysis was performed on the 24 water quality parameters exhibiting elevated concentrations across the three sampling Periods, and the resulting dendrograms are presented in Figure 9. In the 2022 dry season (Period 1), the cluster analysis partitioned the 24 indicators into two major groups. The first group primarily comprised macro-ionic constituents: TDS, TH, Ca²⁺, Mg²⁺, Cl⁻, and SO₄²⁻, reflecting the overall mineralization status and major ion composition of groundwater, with concentration levels predominantly controlled by natural hydrogeochemical processes. The second group encompassed trace constituents and pollution indicator parameters: Mn, An, TP, Fe, and As, representing characteristic components of reducing environments and anthropogenic contamination inputs. In the 2024 dry season (Period 2, 94 samples), the indicator clustering resolution improved substantially: macro-ionic bulk parameters (TH, TDS) and alkaline earth metals (Ca²⁺, Mg²⁺) formed separate sub-clusters, reflecting the differentiation of hydrochemical components from distinct sources. Heavy metal indicators (Mn, Fe, As, Al, Pb, and others) coalesced into an independent cluster group, indicating that trace metal contamination derived from industrial activities possesses relatively independent source characteristics and geochemical behaviors. Nutrient indicators (An, TP) clustered together with OC, indicative of the combined influence of organic pollution and domestic wastewater inputs. In the 2025 wet season (Period 3), the indicator clustering pattern was broadly similar to that of Period 2 but exhibited discernible seasonal variations. Industrial characteristic indicators such as Cl⁻ and SO₄²⁻ showed a stronger tendency to cluster with Na⁺ and TDS, reflecting the enhanced influence of surface-runoff-transported industrial emissions during the wet season. The clustering relationships of redox-sensitive indicators such as Mn and Fe underwent changes, reflecting alterations in redox conditions driven by water table rise during the wet season.

Cluster analysis of sampling sites across the three Periods (Figure 10) revealed distinct spatiotemporal clustering patterns. In the 2022 dry season, the 16 sampling points were partitioned into two clusters: Cluster 1 comprised 15 points (93.75% of the total), encompassing the majority of diffusion and internal points; Cluster 2 contained a single interior monitoring point (N01), whose hydrochemical characteristics differed fundamentally from those of all other sites, indicating strong anthropogenic contamination input at this location. In the 2024 dry season, the 94 sampling points were classified into four clusters: Cluster 1 (EQ33 and associated points), Cluster 2 (N03-2 and vicinity), Cluster 3 (N01-2, N01-4, and other points in the N01 vicinity), and Cluster 4 (the remaining majority of points). In the 2025 wet season, the 79 sampling points were likewise classified into four clusters: Cluster 1 (EQ33 and vicinity), Cluster 2 (N03-2, K10-4, EQ21, and associated points), Cluster 3 (N01-2, EQ34, and vicinity), and Cluster 4 (the remaining points). These clustering results consistently indicate the presence of significant and sustained anthropogenic contamination inputs in the vicinities of EQ33, N03-2, and the N01 well series. The fact that these specific locations persistently formed independent clusters, rather than merging with the majority cluster, across multiple hydroperiods constitutes strong evidence that anthropogenic contamination has fundamentally altered the local groundwater chemical signature, deviating substantially from the natural background range and forming discrete geochemical anomaly zones. This persistent, spatially confined clustering pattern is a diagnostic characteristic distinguishing industrial point-source pollution from natural variability or agricultural seasonal inputs. The increasing number of clusters from 2 (2022) to 4 (2024–2025) further indicates the progressive complication of pollution sources and water quality influencing factors over time, with wet-season non-point source inputs and water-table-fluctuation-driven seasonal pollution effects beginning to manifest.

3.4. PMF Source Apportionment

Building upon the spatiotemporal differentiation patterns revealed by hierarchical cluster analysis, the USEPA PMF 5.0 model was employed to quantitatively resolve pollution sources for the primary detected indicators in groundwater samples from each of the three sampling Periods. Concentration and uncertainty data for 19 detected parameters in each Period were used as model inputs. The optimal number of factors was determined to be five for all three Periods. The species contribution profiles for each factor are presented in Figure 11, and the detailed source identification results are presented below. The robustness of the selected five-factor solutions was evaluated using EPA PMF 5.0 diagnostics for the 2022, 2024, and 2025 campaigns, with Q(robust)/Q(true) ratios close to 1.0, indicating stable and converged solutions (Table S3). The percentage of scaled residuals outside ±3 was below 5% for all species, and the observed-versus-predicted concentrations showed strong correlations (R² > 0.7) for the dominant tracer species. Bootstrap (BS) analysis showed that the factors were mapped to the base-run factors in more than 90% of resamples for 2024 and 2025, confirming the stability of the source assignments in the large campaigns; for the 2022 campaign, the mapping rate was lower (about 60–70%), supporting our decision to treat the 2022 solution as exploratory. Displacement (DISP) analysis further indicated no significant rotational displacement.

In the 2022 dry season Period, Factor 1 was the dominant contributor of NO₃ (91.1%), TN (89.9%), AS (66.6%), and Cl (65.9%). Factor 2 was characterized by high loadings of Mn (60.5%) and Ni (56.8%). Factor 3 was the primary contributor of An (87.0%). Factor 4 exhibited elevated contributions of Fe (51.7%), Ph (50.0%), Al (41.2%), and F (31.9%). Factor 5 was the dominant source of TP (62.7%), SO₄ (42.3%), TH (36.9%), and Ca (36.6%).

In the 2024 dry season Period, the substantially increased sample size markedly enhanced the source resolution capability of the PMF model, with the tracer indicator assemblages of the five source types exhibiting considerably clearer chemical fingerprint characteristics. Factor 1 was dominated by NO₃ (89.5%) and TN (88.6%) as the overwhelmingly predominant loading indicators. Factor 2 was characterized by Mn (92.4%) and Ni (33.5%) as signature loading species. Factor 3 exhibited a broad-spectrum, multi-indicator anthropogenic pollution signature, with high loadings of An (85.9%), Cl (61.4%), F (57.0%), and SO₄ (55.7%). Factor 4 was characterized by TP (49.4%), Ca (39.8%), TH (30.4%), and TDS (29.0%). Factor 5 was dominated by Fe (72.6%), Al (71.8%), and V (56.3%).

In the 2025 wet season Period, Factor 1 was characterized by NO₃ (90.8%), TN (72.6%), and Cl (58.7%). Factor 2 was dominated by Mn (66.3%) and An (41.2%). Factor 3 exhibited an exceptionally strong petrochemical-organic composite signature, with Ph (96.3%), OC (89.3%), V (72.9%), An (54.2%), and Ni (21.6%). Factor 4 was characterized by TP (65.9%), As (60.6%), and F (55.1%). Factor 5 was dominated by Fe (77.3%), Al (73.2%), Ca (54.9%), SO₄ (54.1%), TH (51.2%), and TDS (46.2%).

4. Discussion

4.1. Identification of Groundwater Pollution Sources

Based on the five-factor PMF source apportionment results across the three sampling Periods, groundwater in the chemical industrial agglomeration is jointly controlled by five pollution source categories. Integrating the foregoing hydrochemical graphical analyses and cluster analysis results, the five sources are identified as: agricultural non-point source pollution, domestic wastewater, industrial emissions, natural geological background, and natural hydrogeochemical evolution. The identification criteria, tracer indicator characteristics, and cross-period evolutionary patterns for each category are elaborated below.

(1) Agricultural Non-point Source Pollution. This source category was characterized by NO₃ and TN as the most stable and definitive tracer indicators across all three Periods, with within-source contributions consistently maintained at 88–91%. The extraordinarily high and temporally stable NO₃ loading establishes this factor’s chemical fingerprint with near-unambiguous specificity. Agricultural non-point source pollution primarily reflects nitrogen fertilizer leaching from farmland surrounding the agglomeration entering groundwater via precipitation infiltration and irrigation return flow. In the 2022 dry season Period with limited sample size (n = 16), this factor additionally absorbed variance from AS (66.6%) and Cl (65.9%), inflating its apparent scope. Under the substantially larger sample sizes of 2024 (n = 94) and 2025 (n = 79), AS and Cl were successfully reassigned to the other factor, leaving NO₃ and TN as the exclusive core indicators of agricultural non-point source pollution. This indicator reassignment trajectory demonstrates that adequate sample size is a necessary condition for PMF to achieve fine-scale deconvolution of overlapping anthropogenic source signatures.

(2) Natural Geogenic Source. This source category was defined by Mn as its hallmark tracer, with within-source Mn loadings of 60.5%, 92.4%, and 66.3% across the three Periods, consistently paired with secondary Ni contributions. The dominance of Mn in a dedicated geogenic factor, rather than in an industrial factor, is a critical interpretive finding: it indicates the dominance of Mn in a dedicated factor, rather than in an industrial factor, suggests, consistent with the reducing conditions and carbonate-hosted aquifer of the region, that elevated Mn most likely originates from reductive dissolution of Mn-oxyhydroxide minerals, although this geogenic attribution should be regarded as the most parsimonious interpretation and would benefit from independent confirmation by mineralogical or isotopic evidence. The field-measured parameters support this interpretation: groundwater in the study area exhibited circumneutral pH and low DO with reducing-to-suboxic Eh, conditions under which Mn-oxyhydroxides are thermodynamically unstable and are reductively dissolved to release Mn²⁺ (and associated Fe²⁺) into solution, which is consistent with the elevated Mn (and moderate Fe) concentrations and with their co-occurrence with Ni in the geogenic factor. The temporal variation in Mn loading reflects the dynamic interplay between natural Mn release and anthropogenic processes: the very high 2024 loading suggests minimal interference from anthropogenic Mn inputs, while the decrease to 66.3% in 2025 coincides with the appearance of Mn in the industrial emission factor, indicating partial superimposition of industrial Mn on the natural geological background during the wet season. Nickel’s consistent co-occurrence with Mn in this factor is consistent with the geochemical affinity of Ni for Mn-oxyhydroxides in carbonate aquifer settings.

(3) Domestic Wastewater. An served as the principal tracer for this source category, with loadings of 87.0%, 85.9%, and (in conjunction with the industrial emission factor) 54.2% across the three Periods. In the 2022 Period, this factor was narrowly defined by An dominance; by 2024, its chemical signature expanded substantially to encompass a broad spectrum of indicators including Cl (61.4%), F (57.0%), SO₄²⁻ (55.7%), and additional contributions from AS, OC, TP, and As, reflecting the evolution of domestic wastewater influence from single-indicator organic pollution (characterized by An) to comprehensive hydrochemical modification. This broad-spectrum evolutionary pattern is attributable to the combined effects of inadequately treated domestic sewage from residential areas within and surrounding the agglomeration, septic tank leakage, and the co-mingling of domestic wastewater with diffuse industrial effluents in unsewered zones. In the 2025 wet season, the partial redistribution of An into the industrial emission factor reflects the spatial overlap of industrial ammonia-bearing wastewater plumes with domestic sewage-impacted groundwater under elevated water table conditions.

(4) Industrial Emissions. This type of source is the one with the most intense temporal evolution of the internal structure and tracer indicators’ combination among all the types of sources. Its three-stage evolution trajectory precisely corresponds to the development process of the chemical industrial park. In 2022 (16 samples), Factor 4 shows a “metal-petrochemical dual emphasis” composite feature: Fe (51.7%) and Ph (50.0%) are the highest load indicators, reflecting the composite emission characteristics of chemical raw material production, storage and transportation activities in the aggregation area. In 2024, Factor 4 transformed to mainly consist of indicators such as Ph (92.1%) and TP (49.4%), reflecting the continuous emission characteristics of chemical enterprises’ production activities. It is worth noting that Ph remains associated with the industrial-emissions factor in all three campaigns (Ph 50.0% in 2022, 92.1% in 2024, and 96.3% in 2025). The persistently high Ph loading confirms continuous, rather than episodic, loading.

(5) Natural Hydrogeochemical Evolution. This source category was characterized by SO₄, Ca, TH, and TDS as its fundamental loading indicators, representing uncontrollable natural water–rock interaction processes including carbonate mineral (calcite CaCO₃, dolomite CaMg(CO₃)₂) and sulfate mineral (gypsum CaSO₄·2H₂O) dissolution in the aquifer system. These processes constitute the fundamental determinant of regional groundwater natural chemical composition, serving as the geochemical background template upon which all anthropogenic pollution signals are superimposed. In the 2025 wet season, Fe (77.3%) and Al (73.2%) appeared in this factor with exceptionally high loadings, co-occurring with Ca, SO₄, TH, and TDS, indicating that silicate mineral weathering (feldspar, clay minerals, ferromagnesian minerals), rather than solely carbonate and evaporite dissolution, plays a significant role in establishing the natural Fe and Al backgrounds in groundwater. The 35.7% F loading in this factor further implicates fluorine-bearing mineral (fluorite, fluorapatite) dissolution as a contributor to the natural F background, providing a PMF-level quantitative basis for distinguishing anthropogenic F sources (industrial fluorine-containing wastewater) from natural F sources (mineral dissolution) in the study area.

Synthesizing the above five-source identification results, groundwater contamination in the study area exhibits a structural framework of ‘dual-layer geological background + three-layer anthropogenic input’: natural hydrogeochemical evolution (carbonate/evaporite dissolution and silicate weathering) and natural geological background (reductive Mn-Ni dissolution) jointly constitute the dual-layer, uncontrollable natural baseline of groundwater chemical composition; agricultural non-point source pollution, domestic wastewater, and industrial emissions, respectively characterized by NO₃-TN, An, and Ph as their chemical fingerprints, constitute the three-layer controllable anthropogenic input system. It is noteworthy that although the natural hydrogeochemical evolution source and the geological background source are classified as natural factors, secondary hydrogeochemical processes driven by human activities operate in localized zones: acidic industrial wastewater injection can accelerate carbonate dissolution (acid-enhanced dissolution effect), and microbial degradation of organic pollutants in contaminated zones generates CO₂, promoting carbonate mineral dissolution while creating conditions favorable for Fe-Mn oxyhydroxide reduction (reductive enhancement effect). This ‘natural–anthropogenic coupled enhancement’ mechanism represents the underlying cause of concurrent TH, TDS, and Mn exceedances at select monitoring locations.

4.2. Quantification of Source Contributions

The PMF model quantified the relative contributions of each pollution source to the overall contamination load of groundwater samples through the factor contribution matrix (G matrix). Based on the multi-indicator average contribution rates (the arithmetic mean, across the 19 input indicators, of the percentage of each indicator’s total concentration attributed to that factor) calculated for each factor across the three Periods, the contribution structure of the five source categories and their temporal evolution characteristics are presented in Figure 12.

In the 2022 dry season, the multi-indicator average contributions of the five source categories, ranked from highest to lowest, were: agricultural non-point source pollution (28.7%), natural hydrogeochemical evolution (19.6%), industrial emissions (18.1%), domestic wastewater (17.8%), and geological background (15.8%). Agricultural non-point source pollution ranked highest at approximately 29%; however, this elevated value should be interpreted with caution, as under the limited sample size of 16, this factor absorbed variance from indicators such as AS (66.6%) and Cl⁻ (65.9%) that more properly belong to other source categories, potentially overestimating the true contribution of agricultural non-point sources.

In the 2024 dry season, the substantially increased sample size led to a substantive restructuring of the multi-indicator average contribution profile: domestic wastewater rose to the top position at 31.1%, followed by industrial emissions at 18.2%, natural hydrogeochemical evolution at 17.9%, agricultural non-point source pollution at 16.0%, and geological background at 12.8%. The marked decrease in agricultural non-point source contribution relative to 2022 is primarily attributable to the successful separation of AS and Cl⁻ under the larger sample size, allowing the agricultural contribution estimate to regress toward a more realistic level. Domestic wastewater emerged as the single largest pollution source in this Period at approximately 31%, and its broad-spectrum chemical signature (high loadings of NH₃-N, Cl⁻, F⁻, SO₄²⁻, AS, OC, TP, and As) indicates that the influence of domestic wastewater had expanded from single-indicator organic pollution to comprehensive hydrochemical modification.

In the 2025 wet season, the multi-indicator average contribution structure underwent another shift: natural hydrogeochemical evolution rose to the top position at 26.4%, followed by agricultural non-point source pollution at 21.3%, domestic wastewater at 21.2%, industrial emissions at 19.8%, and geological background at 11.4%. The marked increase in natural hydrogeochemical evolution contribution is closely associated with elevated groundwater levels and intensified water–rock interactions during the wet season. Because the changes and differences from the dry season of 2022 to the dry season of 2024 were not so significant, it indicates that the rate of regional pollution diffusion was relatively slow. Moreover, the expected span from the dry season of 2024 to 2025 is not long either. Therefore, the impact of annual changes is relatively low. Thus, the main influencing factor for the significant difference between the dry season of 2024 and the rainy season of 2025 should be related to the seasonal effect of groundwater. The relatively stable contribution of industrial emissions across all three Periods (18.1%, 18.2%, and 19.8%) suggests continuous, uninterrupted industrial pollution input to groundwater over multi-year timescales, indicative of chronic industrial contamination rather than episodic pollution events. The similarly elevated contributions of agricultural non-point source pollution and domestic wastewater are significantly correlated with the presence of extensive farmland and residential settlements within and surrounding the agglomeration, indicating continuous and substantial agricultural and domestic pollution inputs.

The contribution ranges obtained here (agriculture 16–29%, domestic wastewater 18–31%, and industry 18–20%) are broadly consistent with those reported by Zhang et al. [23] and Zhu et al. [40], who found agricultural and domestic sources together accounting for 40–60% of groundwater pollution in mixed industrial–agricultural settings in southwestern China, and with Chen et al. [24], who reported industrial contributions of ~15–25% in an industrial area. The dominance of NO₃⁻ in the agricultural factor and of NH₃-N in the wastewater factor is likewise consistent with numerous receptor-model and isotope studies that identify NO₃⁻ as the principal fingerprint of fertilizer leaching and NH₃-N as the principal fingerprint of domestic sewage [23,42]. The substantial natural-background contribution (roughly 25–35% when the geogenic and hydrogeochemical-evolution factors are combined) is consistent with the widely documented predominance of carbonate weathering in controlling the baseline chemistry of karst groundwater [45].

Considering the potential instability of the first period contribution, the comprehensive analysis of the overall temporal evolution is based on the contribution of the second and third periods. It reveals that the total contribution of anthropogenic sources (the sum of the multi-indicator average contribution rates of the agricultural, domestic-wastewater, and industrial factors) across the second and third Periods was 65.4% and 62.3%, respectively, exhibiting a relatively stable pattern that indicates sustained and consistent-intensity anthropogenic influence on the study area, consistent with actual land-use conditions. The temporal fluctuations in natural source contributions (hydrogeochemical evolution + geological background) may be associated with enhanced mineral dissolution fluxes driven by increased precipitation infiltration during the wet season; additionally, changes in geological background source contribution do not solely reflect variations in the intensity of purely natural geological processes but also, to some extent, carry the imprint of human-activity-driven secondary hydrogeochemical processes.

4.3. Implications for Groundwater Pollution Management

The groundwater quality evolution processes and pollution source apportionment results obtained in this study carry clear policy implications for groundwater pollution prevention and control in chemical industrial agglomerations. Accordingly, the following four management recommendations are proposed. (1) Targeted prevention and control of industrial pollution sources. The persistent anomaly zones identified by cluster analysis—EQ33, N03-2, and the N01 well series—should be designated as priority groundwater pollution surveillance areas, subject to intensified monitoring and targeted source tracing investigations. The characteristic industrial pollution indicator Ph (petroleum hydrocarbons) demonstrates that the emission intensity of petrochemical organic pollutants demands serious attention; the construction of double-layer anti-seepage systems for storage tank areas and online pipeline leak monitoring systems should be continuously advanced, and an online groundwater quality monitoring and early-warning system should be established. (2) Coordinated source-specific management of agricultural non-point source pollution and domestic wastewater. The PMF five-factor model achieved effective separation of agricultural non-point source pollution (with NO₃⁻-TN as the core tracer) from domestic wastewater (with NH₃-N as the core tracer), indicating that nitrogen pollution control requires a dual-track strategy: NO₃⁻ reduction should be achieved through soil-test-based formula fertilization, water-fertilizer integration technology, and ecological interception ditch construction in agricultural zones peripheral to the agglomeration; NH₃-N reduction should be achieved through sewage collection pipeline networks and small-scale centralized treatment facilities in concentrated rural residential areas. (3) Differentiated management and attribution strategies for geogenic exceedances. PMF source apportionment results indicate that the geological background and hydrogeochemical evolution serve as the primary sources of Mn and Ni; however, secondary hydrogeochemical processes driven by human activities also operate concurrently. Consequently, in areas dominated by natural geogenic inputs, end-of-pipe treatment strategies at the water supply end (e.g., contact oxidation-filtration) should be prioritized; in areas dominated by anthropogenic inputs, source reduction should be the primary measure; in mixed-source areas, a combined strategy of source reduction plus end-of-pipe water supply treatment should be implemented. (4) Optimization and dynamic management of the seasonal monitoring network. The significant differences in PMF source apportionment results for certain indicators across the three Periods profoundly reveal the potential bias inherent in source apportionment based on single-season (particularly dry-season-only) monitoring data. It is recommended that a regularized monitoring system covering wet, normal, and dry seasons be established to continuously track long-term evolution trends in groundwater quality and pollution source structures. Furthermore, the introduction of multi-isotope tracing techniques is recommended to provide independent geochemical validation of PMF source apportionment results.

4.4. Limitations and Future Research

Several limitations should be considered when interpreting the present results. First, the first campaign (2022) contained only 16 samples relative to 19 PMF input variables, so its source contributions are exploratory and are not averaged with the larger campaigns. Second, only three measurement periods are available, and they differ in hydroperiod (two dry seasons and one wet season), which limits the ability to fully separate seasonal from interannual variation. Third, the PMF source assignments lack independent isotope/tracer validation and are therefore subject to rotational ambiguity. Finally, the monitoring network was progressively expanded, so spatial composition changed across campaigns. Future research should (i) maintain a fixed monitoring-well network across seasons (wet, normal, and dry), (ii) conduct multi-isotope and organic-marker (BTEX/PAH) measurements for independent source validation, and (iii) extend the PMF diagnostics (bootstrap, DISP) to longer time series to strengthen confidence in the quantitative source contributions.

5. Conclusions

This study investigated a representative chemical industrial agglomeration in southwestern China based on monitoring data from 189 groundwater samples collected across three hydroperiods (2022–2025). An integrated methodological framework incorporating hydrochemical graphical analysis, dual-dimensional hierarchical cluster analysis, and PMF receptor modeling was employed to systematically investigate groundwater quality evolution and quantitatively apportion pollution sources. The principal conclusions are as follows:

(1) The overall groundwater quality in the study area is unsatisfactory, with Category V water accounting for 50%, 52%, and 51% of samples across the three Periods, respectively. NH₃-N, TH, SO₄²⁻, and Mn were the most prominent parameters exceeding regulatory thresholds. Internal points within the agglomeration core exhibited significantly poorer water quality compared with comparison points and diffusion-area points.

(2) The dominant hydrochemical facies was HCO₃-Ca type; however, internal points within the agglomeration exhibited a clear evolutionary trend toward HCO₃·SO₄-Ca and Cl·SO₄-Ca types, indicating significant modification of hydrochemical composition by industrial activities. Gibbs diagrams and ion ratio analyses revealed a hydrochemical evolution mechanism jointly driven by natural water–rock interactions and anthropogenic inputs.

(3) Hierarchical cluster analysis revealed spatiotemporal differentiation patterns in groundwater quality. Monitoring points EQ33, N03-2, and the N01 well series persistently formed independent clusters across multiple hydroperiods, indicating sustained and continuous anthropogenic contamination inputs at these locations that have created discrete geochemical anomaly zones, a diagnostic spatial clustering characteristic distinguishing industrial point-source pollution from natural variability or agricultural seasonal inputs. The increasing number of clusters across Periods further signifies the progressive complication of pollution sources and influencing factors, with wet-season non-point source and water-table-fluctuation-driven seasonal pollution effects becoming increasingly evident.

(4) The PMF five-factor model consistently resolved the primary pollution sources in each Period as: agricultural non-point source pollution, geological background, domestic wastewater, industrial emissions, and natural hydrogeochemical evolution. The total contribution of the three anthropogenic sources (agricultural + domestic + industrial) remained relatively stable across the three Periods (mean 63.8%), with average contributions of 18.6%, 26.2%, and 19.0%, respectively, indicating that groundwater contamination in the industrial agglomeration is influenced not only by industrial inputs but also substantially by agricultural non-point sources and domestic wastewater. Effective groundwater pollution management and remediation therefore require prior identification of the primary pollution sources to enable targeted and efficient control measures.

This study reveals that groundwater pollution in chemical industrial agglomerations is characterized by a coupled natural–anthropogenic driving mechanism, and establishes a unified five-factor PMF-based source identification framework that elucidates the dual-layer architecture of the natural baseline (geological background and natural hydrogeochemical evolution) and the three-layer structure of anthropogenic inputs (agricultural non-point source, domestic wastewater, and industrial emissions). The integrated hydrogeochemical-PMF-cluster analysis methodological framework developed herein can be transferred to groundwater quality management applications in comparable complex contamination settings, providing a scientific basis for zoned and source-specific precision pollution prevention and control.