Assessment of groundwater vulnerability based on the modified DRASTIC model: a case study in Baicheng City, China

Accurate assessment of groundwater vulnerability objectively reflects an area’s potential for groundwater pollution and provides a reference basis for pollution control and prevention. The main objective of this study was to modify the original DRASTIC model to improve the consistency of groundwater vulnerability assessment results with regard to the actual conditions of the study area. To optimize the assessment objectivity, two additional factors that are influenced by human activities (land use and degree of groundwater extraction) were added to form the DRASTICLE model. Then, based on the correlation between all factors and measured nitrate concentrations, the improved three-scale analytic hierarchy process (AHP) and the weights of evidence (WOE) methods were used to reassign the factor weights of the DRASTICLE model. The area under the receiver operating characteristic (ROC) curve, denoted as AUC, was used to quantitatively evaluate the accuracy of all four models (original DRASTIC model AUC: 0.62). By modifying the factors and weights, the three new models showed better performance, AUC values were 0.64, 0.73 and 0.85 for the DRASTICLE, AHP-DRASTICLE, and WOE-DRASTICLE models, respectively. These results indicate that the modified models could more accurately convey groundwater vulnerability in the study area. The WOE-DRASTICLE model, which had the best performance, was then used to assess groundwater vulnerability in 2000 and 2010 and these were compared to 2018. In 2000, 2010, and 2018, the proportion of areas with very high groundwater vulnerability was 5%, 6%, 8%, respectively. Meanwhile, the proportion of areas with very low vulnerability decreased from 73 to 75%, and then rose to 82%, indicating that the spatial distribution of groundwater vulnerability has changed over time. Findings of this study are expected to provide a new theoretical basis for the Baicheng City municipal government in China to better manage and exploit groundwater resources.


Introduction
Groundwater is an important drinking water resource because of its good seasonal storage capacity, stable temperature, nonsusceptibility to pollution, and convenience for exploitation. Polluted groundwater conditions are not easy to identify, and once present, such pollution is difficult to control and remediate. Therefore, to improve the sustainable development and utilization of groundwater, it is necessary to develop more effective prevention and control programs for groundwater pollution. Groundwater vulnerability assessment is an important tool for identifying areas that are especially prone to groundwater contamination, for facilitating rational development, and for improving land-use planning and groundwater resource management measures.
Determining the vulnerability of groundwater to contamination is particularly important in Baicheng City, which is an important commodity grain base in China, and where groundwater provides about 90% of the city's water supply. In recent years, with the increase in population, the development of industry and agriculture as well as urban expansion have led to an increase in the extraction of groundwater, and further, the quality of groundwater has also deteriorated. Nitrate pollution is one of the main characteristics of groundwater pollution (Almasri 2008;Jhariya 2019), and is mainly caused by chemical fertilizers and industrial wastewater pollution.
Three main methods for groundwater vulnerability assessment include: process-based simulation (Huan et al. 2016), statistical methods (Bonfanti et al. 2016), and indexoverlay methods (Gogu and Dassargues 2000;Huan et al. 2012). There is a wide global application of index-overlay methods for vulnerability assessment. Examples of these include the DRASTIC, SINTACS, AVI, and GOD methods (Ferreira and Oliveira 2004;Ghazavi and Ebrahimi 2015).
The DRASTIC model (Aller et al. 1987) is currently the most widely used vulnerability assessment method, owing to its ease of operation and ease of obtaining parameters (the baseline parameters are listed in Sect. 3.2). Moreover, its groundwater vulnerability assessment results can be presented in the form of a vulnerability map, which intuitively explains the distribution of groundwater vulnerability (Al-Adamat et al. 2003;Victorine Neh et al. 2015). However, in recent years, scholars have found the DRASTIC model to have certain deficiencies in the selection of model factors and weights. For example, the method fixes the weight and rates of each factor, without adequately considering the actual real-world conditions and the effects of local hydrogeological conditions (Barzegar et al. 2015;Brindha and Elango 2015). Another limitation of DRASTIC is that it does not consider the effects of human activities on groundwater pollution. Therefore, it has become necessary to optimize the DRASTIC model such that more objective results can be achieved in combination with the actual conditions of the study area.
Theoretically, optimization of the DRASTIC model can be divided into two main aspects. First is the optimization of the assessment factors, which can be carried out by modifying the rating of factors, removing existing factors (An and Lu 2018), or adding other parameters (Khosravi et al. 2018;Omotola et al. 2020). Gogu and Dassargues (2000), Babiker et al. (2005) and Wang and Yang (2008) have all suggested the addition of land use into the evaluation model because this factor is strongly related to groundwater vulnerability. Accordingly, Kazakis and Voudouris (2015) and Wu et al. (2018) added land-use factors to modify the model, and were able to improve the model's accuracy. At the same time, for this study, different degrees of groundwater over-exploitation are known to exist in certain parts of the study area. Therefore this study did not only include land use (L) as a vulnerability assessment factor (Khan and Jhariya 2019;Sener and Davraz 2012), but also considered the spatial distribution of groundwater extraction rates (E) (Abu-Bakr 2020) into the DRASTIC model to obtain a new model called DRASTICLE.
The second aspect is to optimize the weight of the parameters (Barzegar et al. 2018;Sahoo et al. 2016). To do this, we refer to Pacheco et al. (2015), who adopted five methods to modify the weights of the DRASTIC model, and Khosravi et al. (2018) who applied four objective methods to modify the original DRASTIC model. Other modification strategies include using the analytic hierarchy process (AHP) (Bai et al. 2012;Sener and Davraz 2012;Thirumalaivasan et al. 2003), the artificial neural network, the fuzzy logic method (Rezaei et al. 2013), the logistic regression method (Antonakos and Lambrakis 2007), and the weights of evidence (WOE) method (Khosravi et al. 2018) to optimize the weights of parameters in the DRASTIC models.
In this study, two objective methods were used to modify the weights in a DRASTIC-based vulnerability assessment model for Baicheng City: the improved three-scale AHP and the WOE methods. The analytic hierarchy process (Saaty and Kearns 1985) is simple to operate and has strong practicability and adaptability. The traditional AHP determines the judgment matrix using the nine-scale principle, yet with this method it is challenging to determine a reasonable judgment matrix because of the difficulty in determining the relative importance of each of the parameters. Therefore, for this study, the three-scale method (Zuo 1988) was adopted to simplify the model, which is conducive for comparing the relative relationships between parameters. In the improved three-scale AHP, the correlation between each parameter and nitrate concentration in the model were calculated and compared to determine the importance of the parameters, which overcomes the artificial subjectivity of the AHP and improves the accuracy of groundwater vulnerability assessment.
Similarly, WOE is a geological statistical method that can be used in the spatial analysis of information from multiple sources. This analysis method is increasingly being used due to the increasing sophistication and widespread availability of geographic information system platforms and corresponding expansion modules. This is especially the case for metallogenic prediction (Zhang et al. 2016), but the methodology has also been extended to other similar fields, particularly for risk analyses such as landslide sensitivity analyses (Hong et al. 2017) and the assessment of karst collapse risk zones (Perrin et al. 2015). Barber et al. (1998) first applied WOE to regional groundwater vulnerability assessments with good results. Because of the high solubility and mobility of nitrate in water, this compound is easily transported in water that infiltrates through the vadose zone to groundwater and is a good reflection of the degree of groundwater pollution (Shrestha et al. 2016;Voutchkova et al. 2021;Wang and Yang 2008). Consequently, nitrate was selected as the response factor in this study.
The main objective of this study was to use the DRASTIC model to evaluate groundwater vulnerability in the study area. First, two additional parameters, land use and degree of groundwater extraction, were added to obtain the DRAS-TICLE model. Second, the improved three-scale AHP and WOE methods were used to optimize the parameter weights of the DRASTICLE model.

Study area
Baicheng City is located in the northwestern part of Jilin Province in China. It is located to the west of Songnen Plain, and to the east of the Horqin Grassland, between longitudes 121° 38′ 06″-124° 23′ 56″ E and latitudes 44° 13′ 57″-46° 18′ 15.8″ N (Fig. 1). This region covers an area of about 25,600 km 2 . The region has a temperate continental monsoon-type climate with obvious seasonal changes. The average annual precipitation is approximately 400 mm, which has an uneven distribution throughout the year. The average annual pan evaporation rate is 1340 mm, and the annual average temperature is 4.7 °C.
Low hills are situated in the northwestern part of the study area that have elevations of 300-662.6 m above mean sea level. The elevations of the northeastern and southeastern plains are 130-140 m above mean sea level; in the southwest, a latent desert area is situated 150-180 m above mean sea level. From the northwest to the southeast, the terrain of Baicheng comprises successive low mountains, hills, and plains, and is slightly uplifted in the southwest (Feng 2019). The strata mainly include rocks of the Carboniferous, Permian, Jurassic, Cretaceous, Neogene, and Quaternary ages.
Groundwater occurs in four main aquifer types within the study area: loose rock pore unconfined aquifers, unconsolidated rock pore confined aquifers, clastic rock porefissure aquifers, and bedrock fissures. The lithology of the unconfined aquifer in the valley plain is sand and sand/ gravel with a thickness of 20-40 m. In Taoer River alluvial fan, the unconfined aquifers consist of gravels and sands with a small amount of clay, and a thickness of about 10-20 m. The lithology of the unconfined aquifers in the alluvial lacustrine plain is mainly alluvial sand and loesslike sandy loam, with a thickness of about 5-10 m.
The unconsolidated rock pore confined aquifers are distributed in the low plain. Their lithology is mainly composed of gravels and sands, with a thickness of about 5-15 m, and can be up to 30 m. The clastic pore-fissure aquifers are mainly distributed under Quaternary strata in the low plain area. The lithology is mainly sandstone and glutenite, with a thickness of about 30 m. The main groundwater type in the hilly area is bedrock fissure water. Overall, groundwater is mainly stored in the structural fracture zones of various bedrock lithologies, in the weathered fissure zones, and in the alluvial sand and gravel layers of the river valley. These sediments have an uneven distribution and vary greatly in water content. Among these, the object of this study is the unconfined aquifer in the Baicheng city area.
In 2018, the Baicheng sub-center of the Jilin Water Environment Monitoring Center performed sampling of 51 shallow groundwater monitoring wells in the city plain area during April and September, which were tested for water quality parameters. The main chemical parameters that exceeded the Chinese water quality standard were nitrate, ammonia-nitrogen, manganese, fluoride, and arsenic. Elevated concentrations of these parameters suggest that groundwater in the area has suffered from different degrees of pollution. Associated land-use types in the area are mainly cultivated land, accounting for approximately 60% of the study area, followed by grassland and salinealkali land. From 2000 to 2018, the area of cultivated land and artificial surface both increased gradually, and the proportions of grassland and forest accordingly gradually decreased.

Source of data
The meteorological data and hydrogeological data used in this study were collected and derived from the results of field measurements and sampling analysis by the project team. The data for groundwater depth, hydrochemistry, and groundwater exploitation were either provided by the Baicheng Water Resources Management Center, or measured in the field by the project team. The groundwater depth data were collected from long-term monitoring wells in 2000, 2010 and 2018, with each well dataset covering January-December, measured every five days. The number of monitoring wells that were measured varied from 120 to 160 per year. Nitrate data were collected from 202 wells in November 2017 by the project team and 52 wells were tested by the Baicheng Water Resources Management Center in April 2018. In addition, a total of 205 hydrogeological borehole logs were collected from this area.
Land-use type data grids of Baicheng City in 2000, 2010, and 2018 with a resolution of 30 × 30 m were downloaded from the National Catalogue Service for Geographic Information (http:// www. webmap. cn/ main. do? method= index) and GLOBELAND30 (http:// www. globa lland cover. com/ home. html? type= data).

The DRASTIC and DRASTICLE models for groundwater vulnerability
The DRASTIC model is mainly aimed at assessing the vulnerability of an unconfined aquifer. The model selects seven factors affecting groundwater flow and pollutant transport as vulnerability assessment parameters: depth to groundwater (D), net recharge (R), aquifer media (M), soil media (S), topography (T), impact of the vadose zone (I), and hydraulic conductivity of the aquifer (C). For this study, each parameter was classified according to its range of variation and internal attributes, and the corresponding vulnerability ratings were given; the larger the rating, the higher the vulnerability grade. The rating and weight of each factor for groundwater vulnerability have been previously described by Aller et al. (1987). The groundwater vulnerability index (VI) was calculated using Eq. 1: where the subscripts r and w represent the ratings and weights of the parameters, respectively.
Based on the original DRASTIC model, the two aforementioned factors affected by human activities, the type of land use (L) and degree of groundwater extraction (E), were added to generate the DRASTICLE model, and the VI was calculated according to Eq. 2. Bojórquez-Tapia et al. (2009) indicated that five categories of groundwater vulnerability should be appropriate for conveying meaningful information to planners, decisionmakers, and stakeholders. Therefore, in this study, the groundwater vulnerability index was divided into five categories: very low, low, moderate, high, and very high.

Preparation of the DRASTICLE parameters
All factors used in the model for this study are described as follows, and the classes and ratings of all factors are listed in Supplementary Table A. Depth to groundwater (D): The depth to groundwater determines the contact time between the surface pollutants and aeration zone media before entering the aquifer. Generally speaking, the shallower the depth to groundwater, the shorter the time it takes for pollutants to reach the aquifer, and the more susceptible the groundwater is to pollution. The inverse distance weight tool of ArcGIS was used to process the groundwater table depth data of 120 long-observation wells in the research area. These data were then used to develop a depth to groundwater map for the study area. (1) Page 5 of 14 230 The groundwater depth was divided into five depth groups: 1.2-1.5, 1.5-4.6, 4.6-9.1, 9.1-15.2, and 15.2-17.3 m, and the corresponding ratings were given. The results are presented in Fig. 2a.
Net recharge (R): Contaminants on the surface or soil can be transported vertically to groundwater through recharge water and transported within the aquifer. The greater the amount of recharge, the greater the possibility of pollutants reaching the aquifer (i.e., there is an increased vulnerability of groundwater to contamination). Many recharge sources of groundwater exist in the study area, including precipitation infiltration, lateral recharge in mountainous areas, river channel leakage, and well irrigation return recharge. The net recharge ranged from 51.2 to 236.8 mm and was divided into three categories, the results of which are shown in Fig. 2b.
Aquifer media (A): The pore characteristics of the aquifer media determine the velocity of groundwater flow and affect the adsorption, diffusion, and dispersion of pollutants. In general, larger aquifer medium particles and more pores leads to the presence of better permeability, and a lower probability of the pollutants being diluted and attenuated; thus, the higher the groundwater vulnerability. The target of this study was an unconfined aquifer, for which the fissure water aquifer was dominated by granite, and the pore water aquifer was dominated by gravelly pebbles, loess sub-sand, and silty-fine sand. The Taoer River alluvial Fan phreatic aquifer is mainly composed of gravel and pebble gravel. Other shallow aquifers in the study area include sandy and silty sediments beneath an alluvial and lacustrine plain, and sandy loess sediments beneath swamps and a saline-alkali valley (Fig. 2c).
Soil media (S): The soil medium affects the amount of surface water infiltrating underground, as well as the ability of pollutants to infiltrate into the vadose zone beneath the ground surface. The surface of the study area is overlain by shallow sediments of Pleistocene and Holocene age. The underlying sediments of Lower Pleistocene age consist of glacial sediments with clay content. The moraine deposits of the Upper Pleistocene Zhenxi Ice Age are mainly distributed in the Taoer River alluvial Fan and Jiaoliu River valleys. The upper ice-water accumulation layer is mainly composed of gravel, sand, and gravel with minimal clay, and is interbedded with clay, loam, and fine sand lenses. The alluvium of the Guxiangtun Formation, which is mainly distributed in the alluvium plain, is a loess-like sandy soil. Holocene alluvial deposits, which are distributed in the floodplain of the Taoer, Jiaoliu, and Hanhe rivers, as well as along the periphery of the fan front, are mainly composed of gravels, and sands and feature a sandy loam layer. The associated aeolian sediments are comprised of yellow-and white-coloured fine sands (Fig. 2d).
Topography (T). Slope mainly affects the infiltration of atmospheric precipitation. The lower the slope, the more infiltration is generated; thus, the higher the potential for pollution. In this study, the slope distribution was extracted from digital elevation model data of Baicheng City using the ArcGIS surface analysis tool. The terrain in the study area is relatively gentle overall, though the hilly area in the northwest part is steeper, with a topographic slope of up to 70%. The slopes in the area were classified into five categorized: 0-2%, 2-6%, 6-12%, 12-18%, and > 18% (Fig. 2e).
Impact of the vadose Zone (I): The vadose zone controls the length of the infiltration path and seepage path of the surface water. The soil layer of the vadose zone has a remarkable ability to adsorb and block the entry of pollutants into groundwater. The better the sorting, the finer the particles and the higher the clay content of the vadose zone medium; thus, the worse the permeability, the stronger the adsorption and purification ability, the stronger the pollution prevention performance, and the weaker the groundwater vulnerability, hence the greater the vulnerability of the groundwater. The main lithologies in the vadose zone of the study area were found to consist of, clayey and silty sediments intermixed to various extents with sands and gravels. In the Taoer River alluvial Fan area, the lithology of the vadose zone was mainly gravel, while the valley area was loam, and in the western low plain, the vadose zone was mainly comprised of silts and sands (Fig. 2f).
Hydraulic conductivity of the aquifer (C): This parameter assesses the ability of sediments to transmit water and dissolved pollutants through pore spaces. The hydraulic conductivity of aquifers in the study area varies significantly. Using the obtained permeability coefficient values of 157 wells, the distribution of hydraulic conductivity values in the area was obtained by simple kriging difference values, and the distribution pattern was roughly the same as the lithological distribution, ranging from 0.65 to 470.36 m/day. Sediments in the Taoer River alluvial fan had the largest hydraulic conductivity values (Fig. 2g).
Land use (L): The range of pollutants that have the potential to be discharged to soils depends on the dominant land use activities in a region, as does the extent to which soil pollution by these chemical constituents. Additionally, all the characteristics of the surface cover can also be expected to have a great impact on the interception capacity of pollutants and the manner in which pollutants enter the aquifer. Typically, the vulnerability of groundwater in industrial areas is high, mainly because the distribution of factories is concentrated, and if the production wastewater is not discharged according to regulatory standards or is treated improperly and leakage occurs, the wastewater will likely become a potential source of groundwater pollution. In agricultural areas, the wide-spread application of a variety of pesticides, fertilizers, and livestock and poultry manure can easily produce surface pollution sources, which pose a threat to groundwater quality. In green belt areas, such as Fig. 2 Maps of groundwater vulnerability conditioning factors: a depth to groundwater, b net recharge, c aquifer media, d soil media, e topography, f impact of the vadose zone, g hydraulic conductivity of the aquifer, h land use, and i degree of groundwater extraction grasslands and forests, surface plants have the function of reducing surface runoff, reducing soil erosion, and reducing recharge due to transpiration. Consequently, a dense vegetation cover on soils may have a certain protective effect on groundwater, and the vulnerability of groundwater in these areas is lower than areas with little vegetation. The land-use types in the study area were divided into eight categories: artificial surfaces, forest, water bodies, wetlands, shrubland, cultivated land, grassland, and bare land. (Fig. 2h).
Degree of groundwater extraction (E) Groundwater exploitation intensity is another major factor affecting groundwater vulnerability. Excessive exploitation of groundwater leads to an increased drop in groundwater levels and the increased formation of cones of depression within the water table, resulting in an increased hydraulic gradient. The degree of groundwater extraction is the ratio of the actual extraction amount to the recoverable amount. The degree of groundwater extraction in the Taobei District was 94%, while the proportion of extraction in Taonan City in the Huolin River Basin was 14%, and most of the other areas was between 20 and 50% (Fig. 2i).

Factors correlation test
The evaluation factors in the model were expected to be relatively independent and have no strong correlation with each other. Therefore, a multi-collinearity assessment procedure was performed among the evaluation factors (Arabameri et al. 2019; O'Brien 2007) to derive the tolerance and variance inflation factor (VIF), with the thresholds of tolerance < 0.1 and VIF > 10, indicating strong multi-collinearity. The random point creation tool of ArcGIS, which was adopted to create 10,000 random points in the research area, extracted the rating values of nine parameters corresponding to each point, and calculate the tolerance and VIF using SPSS software.

Use of a three-scale analytic hierarchy process (AHP)
1. The AHP was used to modify the weights of the DRAS-TICLE model. This was carried out using a three-scale method instead of the original nine-scale method to better construct the judgment matrix. The AHP assessment was required to compare and judge parameters, and to determine their order of importance. The significance scale comparison is presented in Table 1. To determine the magnitude of the relationship, a correlation analysis between each parameter and the actual nitrate concentration was carried out. It was assumed that the higher the correlation with nitrate, the more important the parameter. In this way, the degree of importance of each parameter in the vulnerability assessment could be determined.

Use of the weights of evidence assessment approach
The weights of evidence method (WOE) (Agterberg 1989) is a geostatistical quantitative prediction method based on binary (existing or non-existent) images and Bayes' rule under the assumption of independent conditions. Assuming that the study area is A(T) km 2 , and the study area is divided into cells of area U km 2 , the total number of cells in the study area is N(T) = A(T)/U. Assuming that there are N(D) cells with response factor (D) distribution, the probability of the occurrence of a response factor in any cell selected in the research area is P(D) = N(D)/N(T), which is called the prior probability. It is assumed that the prior probabilities of each cell are equal throughout the study area. Then, the prior probability is expressed in terms of the odds (O): The weights are calculated as follows: where B is the model factor, and D is the response factor (nitrate concentration). The weight contrast is C = W + − W − , and the standard deviation of the weight difference is = √ 2 (W + ) + 2 (W − ) , where 2 W + and 2 (W − ) are the  The ci factor is less important than the cj factor 1 The ci factor is as important as the cj factor 2 The ci factor is more important than the cj factor Environmental Earth Sciences (2022) 81:230 variances of W + and W − , respectively. The final weight is W = C (C) . The WOE requires that the distribution of predictors relative to the response factor satisfy the condition of independence. For n predictors, if all are conditionally independent with respect to the response factor, the logarithm of the odds is: Finally, using the formula P = R 1+R , the logarithm of the posterior odds can be transformed into the posterior probability.

Comparison and validation of models
To validate and compare the accuracy of the five models, the ROC curve (Mukherjee and Singh 2020) was used to evaluate and compare the results of different models. This methodology takes each value of the predicted results as the possible judgment threshold and calculates the corresponding sensitivity and specificity accordingly. The false-positive rate (1 − specificity) is taken as the horizontal coordinate, and the true positive rate, that is, sensitivity, is drawn as the vertical coordinate. The area under the ROC curve, the AUC value, is a good measure of the model's predictive accuracy, and ranges in value from 0.5 to 1. The larger the value, the stronger the judgment of the model.
In this study, a ROC curve was drawn based on the groundwater nitrate concentration and the groundwater vulnerability index values. The most effective model was identified and then used to evaluate the groundwater vulnerability of the study area in 2000 and 2010 and further analyze the associated temporal and spatial distribution of groundwater vulnerability.

Multi-collinearity diagnosis
The results of the multi-collinearity diagnosis for each evaluation factor are presented in Table 2. The tolerance and VIF values were 0.35-0.97 and 1.04-2.84, respectively. Both of these parameters met the conditions of tolerance > 0.1 and VIF < 10, which indicates that there was no overlap among the nine evaluation factors. Consequently, these parameters were considered to be independent and suitable for use in the model evaluation.

Groundwater vulnerability assessment using the DRASTIC and DRASTICLE models
According to the original DRASTIC model, the minimum and maximum values of the groundwater vulnerability assessment index in the study area were 94 and 193, respectively. Groundwater vulnerability was classified into five categories based on the Jenks method in ArcGIS, and a groundwater vulnerability distribution map was drawn (Fig. 3a). Using this process, areas with a very low vulnerability to contamination (I) vulnerability region accounted for about 9% of the study area, mainly distributed in the hilly area in the northwest. The areas with a low (II) vulnerability were the most distributed, accounting for about 41% of the study area. The moderate (III), high (IV), and very high (V) vulnerability regions accounted for about 29%, 15%, and 5% of the study area, respectively. Based on this, the overall groundwater vulnerability in the study area is generally relatively low. The weight of evaluation factors in DRASTICLE model are shown in Table 3.The vulnerability assessment results calculated according to Eq. 2 are shown in Fig. 3b. These results indicate that the vulnerability of the study area (considering the effects of groundwater abstraction) is mainly moderate (III) and high (IV), accounting for about 35% and 22% of the study area respectively.

Groundwater vulnerability assessment using the three-scale AHP methodology
To verify the consistency of the judgment matrix, the consistency index values of the AHP-DRASTICLE model were found to be 0.047. As these values are less than 1, this indicates that the normalized weight values passed the consistency test. The weights of the evaluation factors are listed in Table 3.
According to the weights in Table 3, the evaluation factors were weighted and superimposed to obtain the distribution map of groundwater vulnerability in the study area, as shown in Fig. 3c. According to the AHP-DRASTICLE model, the groundwater vulnerability in the study area was mainly low (II) and moderate (III).

Groundwater vulnerability assessment using the WOE methodology
Taking the nitrate concentration as the response factor, there were 254 nitrate points in the study area. Monitoring wells with concentrations of NO 3 − ≥ 50 mg/L were selected as the response factor occurrence point, and there were 26 points with NO 3 − ≥ 50 mg/L. The probability of the occurrence of the response factor was P(D) = 0.102, and the prior probability odds were O(D) = P(D) 1−P(D) =0.114. The weights calculated according to Eqs. 4 and 5 are shown in Supplementary Table  S. According to the weights in Supplementary Table A, among the factors affecting groundwater vulnerability, depth to groundwater (D) at 4.6-9.1 m had the highest impact (1.668) on groundwater vulnerability. Other factors that were found to have significant impacts on groundwater vulnerability in the study area were: net recharge (R) at 177.8-254 mm; the aquifer media (A) where sands and gravels were present; the soil medium (S) where loamy soils were present; topography (T) where slopes were in the range of 2-6%; the impact of the vadose zone (I) where sand and gravels were present; the hydraulic conductivity of the aquifer (C) when this was greater than 81.5 m/day; land use (L) where this was artificial surface in urban areas; and the degree of groundwater extraction (E) where this was in the range of 50-70%.
According to Eq. (6), the posterior probability of the response factors of the WOE-DRASTICLE model was calculated. The results were classified into five categories according to the inflection point of the frequency distribution histogram, as shown in Fig. 3d. The assessment of groundwater vulnerability by WOE showed that the vulnerability was mainly very low (I) and low (II), accounting for 52% and 20% of the total area, respectively.

Evaluation and comparison of model results
As can be seen from Fig. 4, the original DRASTIC model had the lowest AUC value of 0.62, but this was increased to a value of 0.64 after the addition of the two additional factors to create the DRASTICLE model. The Kruskal-Wallis nonparametric test was performed on the results of the two models, and the incremental significance value was 0.000, that is, less than 0.01, indicating that the results of the DRASTIC and DRASTICLE models were significantly different. This suggests that the evaluation result of the model is improved after adding two factors. Similarly, the calculated AUC for the DRASTICLE model after the assessment using the three-scale AHP method was 0.73. The AUC of the DRASTICLE model using the WOE assessment method had the largest value of 0.85. This indicates that after adding two factors and modifying weights by two methods (AHP and WOE), the evaluation results of the model are better than the traditional DRASTIC model.

Changes in groundwater vulnerability over time
The WOE-DRASTICLE model was used to investigate in more detail the spatial and temporal variability of groundwater vulnerability in the study area. This assessment was carried out using data from the years 2000 and 2010. The distribution of groundwater vulnerability that was produced by this assessment is shown in Fig. 5. In 2000, the very low (I) and low (II) vulnerability zone accounted for about 72% of the area, and the very high (V) vulnerability zone accounted for only about 4% of the area, mainly distributed in the Taoer River Fan area. In 2010, the areas of very low (I) and very high (V) groundwater vulnerability accounted for about 21% and 11% of the area, respectively. However, the very low (I) vulnerability zone accounted for about 52% of the area and the very high (V) vulnerability zone for about 5% of the area in 2018. From 2000 to the present, the temporal and spatial distribution of groundwater vulnerability has changed greatly, which is related to groundwater depth and land use type change.

Discussion
The concept of the DRASTIC model was used to evaluate groundwater vulnerability in the area. The standard DRAS-TIC model was first used to evaluate the inherent vulnerability of the aquifer. This assessment indicated that much of the study area had a low (II) vulnerability (41%) to contamination, with high (IV) and very high (V) vulnerability zones accounting for only about 20% of the area. As can be seen from Fig. 3a, 50% of the points with nitrate > 50 mg/L fall into high (IV) and very high (V) vulnerability zones, and the remaining 50% of these values are mainly distributed in low (II) vulnerability zones. Finally, about 49% of the nitrate < 50 mg/L points are distributed in the very low (I) and low (II) vulnerability zones.
The seven parameters in the standard DRASTIC model only consider the geological and hydrogeological conditions of the study area and do not consider the effect of human activities on groundwater. However, the effects of human activities are considered in the two added parameters that were used to generate the DRASTICLE model. In the DRASTICLE evaluation results, only about 20% of the nitrate concentrations that are less than 50 mg/L points fell in the high (IV) and very high (V) vulnerability zones, and 31% of the nitrate concentrations > 50 mg/L were located in the very low and low vulnerability zones. Therefore, after adding the two parameters, the accuracy of the evaluation results was further improved.
The vulnerability assessment was further improved by using the three-scale AHP and WOE assessments to modify the weights of the parameters. The evaluation results of the AHP-DRASTICLE model showed good dispersion, with about 19% of nitrate concentrations > 50 mg/L points falling in the low and very low vulnerability zones, and 82% of nitrate concentrations < 50 mg/L points falling in the moderate, low, and very low vulnerability zones. Similarly, about 81% of the nitrate concentrations > 50 mg/L in the WOE-DRASTICLE fell in the high and very high vulnerability zones, only 3% of the nitrate < 50 mg/L points fell in the very high vulnerability zone. These results show that the evaluation results are more reliable after the weight of parameters was modified by the use of the AHP and WOE methods.
The WOE results show that depth to groundwater at 4.6-9.1 m had the highest impact (1.668) on vulnerability, implying that shallower groundwater levels are not necessarily more vulnerable to pollution. Moreover, with increased net recharge to the aquifer, the chance of contaminants entering the groundwater becomes greater and the degree of protection against pollution became smaller, with net recharge at 177.8-254 mm having the greatest impact. In general, for aquifer media, the coarser the particles or the greater the fissure density in the aquifer matrix, the higher the permeability and, consequently, the greater the vulnerability of the aquifer. This was reflected in the model results, which indicated that the presence of sands and gravels in the aquifer matrix had the highest ratings. Considering soil characteristics, in general, the type of soil, amount of expansion and shrinkage, and size of the media particles are the factors which determine the magnitude of soil contamination susceptibility. Soils with relatively thick layers, high organic matter content, and smaller-sized particles have a high capacity to absorb contaminants, as well as good antifouling properties, and a low vulnerability. The results of this study show that soil media had a great influence on loam, but not gravel. In general, for topography, the topographic slope affects the surface runoff volume, runoff velocity, runoff direction, and residence time. The gentler the slope, the more pollutants are able to leach, and thus, the more likely they will leach into the groundwater. The results of this study showed that topography had the greatest effect at 2-6%, not 0-2% as was expected from research by Khosravi et al. (2018).
Considering the vadose zone, in general, if the clay gravel content in the vadose zone is higher and the granules are finer, the amount of water that would infiltrate through the soil profile to groundwater would be lower. In this study, the WOE results show that the impact of the vadose zone had the greatest influence when this zone contained sands and gravels. In general, for hydraulic conductivity, the hydraulic conductivity coefficient of an aquifer reflects the aquifer's hydraulic permeability, which determines the migration rate of pollutants in the aquifer and the degree of difficulty of pollutants entering the groundwater system. In this study, the results showed that when the hydraulic conductivity of the aquifer was > 81.5 m/day, the groundwater vulnerability was high.
The presence of artificial surfaces had the greatest influence on the groundwater vulnerability assessment, likely because the farmers use practices such as raising livestock and grow crops and vegetables on their land, thus the associated animal waste, pesticides, and fertilizers would be expected to lead to substantial pollution of the groundwater. The degree of groundwater extraction at 50-70% had the greatest effect on groundwater pollution probability.
The AUC of the ROC curve shows that both the improved DRASTICLE, AHP-DRASTICLE and WOE-DRASTI-CLE models had better evaluation results than the standard DRASTIC vulnerability model. Overall, the findings show that the reliability of the model evaluation results was improved in two ways: by adding the two factors which assess the effects of human activity on groundwater contamination, and by improving the weighting of the factors in the DRASTICLE model.

Conclusions
This study evaluated the vulnerability of groundwater to contamination in Baicheng City in China. This was undertaken by optimizing the standard DRASTIC assessment method to better represent the actual vulnerability distribution of groundwater. Both the factors and weights of the DRASTIC method were altered to improve the overall vulnerability assessment in the study area. This modified vulnerability model is known as DRASTICLE.
In consideration of the influence of human activities on groundwater, two factors, land use and degree of groundwater extraction, were added to the evaluation factors. The three-scale AHP and WOE methods were also used to improve the weight of the factors. The results of the four models (DRASTIC, DRASTICLE, AHP-DRASTICLE, WOE-DRASTICLE) were compared using the ROC curve method. The results of this assessment showed that the model improvements proposed in this study had obvious improvements: the model evaluation results were more accurate and had higher correlations with nitrate concentrations in groundwater in the study area.
The WOE-DRASTICLE model was selected to evaluate the groundwater vulnerability of Baicheng City in 2000 and 2010. From 2000 to 2018, the spatial and temporal distribution of groundwater vulnerability in the region changed greatly. The areas of high vulnerability were distributed in the Taoer River fan in the northwest of the study area. Importantly, it should be noted that groundwater pollution can still occur in very low or low vulnerability areas, though compared with areas of high vulnerability, these areas are less vulnerable to human activities and natural environmental pollution. Overall, it is recommended that government departments should facilitate reasonable control of groundwater pollution prevention and extraction according to changes in groundwater vulnerability in the region. This study, therefore, provides a theoretical basis for the Baicheng City municipal government to manage and exploit groundwater resources.