A study of uncertainties in groundwater vulnerability modelling using Bayesian model averaging (BMA)

A machine learning framework for spatio-temporal vulnerability mapping of groundwaters to nitrate in a data scarce region in Lenjanat Plain, Iran

The temporal aspect of groundwater vulnerability to contaminants such as nitrate is often overlooked, assuming vulnerability has a static nature. This study bridges this gap by employing machine learning with Detecting Breakpoints and Estimating Segments in Trend (DBEST) algorithm to reveal the underlying relationship between nitrate, water table, vegetation cover, and precipitation time series, that are related to agricultural activities and groundwater demand in a semi-arid region. The contamination probability of Lenjanat Plain has been mapped by comparing random forest (RF), support vector machine (SVM), and K-nearest-neighbors (KNN) models, fed with 32 input variables (dem-derived factors, physiography, distance and density maps, time series data). Also, imbalanced learning and feature selection techniques were investigated as supplementary methods, adding up to four scenarios. Results showed that the RF model, integrated with forward sequential feature selection (SFS) and SMOTE-Tomek resampling method, outperformed the other models (F1-score: 0.94, MCC: 0.83). The SFS techniques outperformed other feature selection methods in enhancing the accuracy of the models with the cost of computational expenses, and the cost-sensitive function proved more efficient in tackling imbalanced data issues than the other investigated methods. The DBEST method identified significant breakpoints within each time series dataset, revealing a clear association between agricultural practices along the Zayandehrood River and substantial nitrate contamination within the Lenjanat region. Additionally, the groundwater vulnerability maps created using the candid RF model and an ensemble of the best RF, SVM, and KNN models predicted mid to high levels of vulnerability in the central parts and the downhills in the southwest.

Uncertainty-based saltwater intrusion prediction using integrated Bayesian machine learning modeling (IBMLM) in a deep aquifer

Data-driven machine learning approaches are promising to substitute physically based groundwater numerical models and capture input-output relationships for reducing computational burden. But the performance and reliability are strongly influenced by different sources of uncertainty. Conventional researches generally rely on a stand-alone machine learning surrogate approach and fail to account for errors in model outputs resulting from structural deficiencies. To overcome this issue, this study proposes a flexible integrated Bayesian machine learning modeling (IBMLM) method to explicitly quantify uncertainties originating from structures and parameters of machine learning surrogate models. An Expectation-Maximization (EM) algorithm is combined with Bayesian model averaging (BMA) to find out maximum likelihood and construct posterior predictive distribution. Three machine learning approaches representing different model complexity are incorporated in the framework, including artificial neural network (ANN), support vector machine (SVM) and random forest (RF). The proposed IBMLM method is demonstrated in a field-scale real-world "1500-foot" sand aquifer, Baton Rouge, USA, where overexploitation caused serious saltwater intrusion (SWI) issues. This study adds to the understanding of how chloride concentration transport responds to multi-dimensional extraction-injection remediation strategies in a sophisticated saltwater intrusion model. Results show that most IBMLM exhibit r values above 0.98 and NSE values above 0.93, both slightly higher than individual machine learning, confirming that the IBMLM is well established to provide better model predictions than individual machine learning models, while maintaining the advantage of high computing efficiency. The IBMLM is found useful to predict saltwater intrusion without running the physically based numerical simulation model. We conclude that an explicit consideration of machine learning model structure uncertainty along with parameters improves accuracy and reliability of predictions, and also corrects uncertainty bounds. The applicability of the IBMLM framework can be extended in regions where a physical hydrogeologic model is difficult to build due to lack of subsurface information.

Hydro-chemical based assessment of groundwater vulnerability in the Holocene multi-aquifers of Ganges delta

Determining the degree of high groundwater arsenic (As) and fluoride (F-) risk is crucial for successful groundwater management and protection of public health, as elevated contamination in groundwater poses a risk to the environment and human health. It is a fact that several non-point sources of pollutants contaminate the groundwater of the multi-aquifers of the Ganges delta. This study used logistic regression (LR), random forest (RF) and artificial neural network (ANN) machine learning algorithm to evaluate groundwater vulnerability in the Holocene multi-layered aquifers of Ganges delta, which is part of the Indo-Bangladesh region. Fifteen hydro-chemical data were used for modelling purposes and sophisticated statistical tests were carried out to check the dataset regarding their dependent relationships. ANN performed best with an AUC of 0.902 in the validation dataset and prepared a groundwater vulnerability map accordingly. The spatial distribution of the vulnerability map indicates that eastern and some isolated south-eastern and central middle portions are very vulnerable in terms of As and F- concentration. The overall prediction demonstrates that 29% of the areal coverage of the Ganges delta is very vulnerable to As and F- contents. Finally, this study discusses major contamination categories, rising security issues, and problems related to groundwater quality globally. Henceforth, groundwater quality monitoring must be significantly improved to successfully detect and reduce hazards to groundwater from past, present, and future contamination.

Modelling groundwater vulnerability in a vulnerable deltaic coastal region of Sundarban Biosphere Reserve, India

Groundwater is the most reliable source of freshwater for human well-being. Significant toxic contamination in groundwater, particularly in the aquifers of the Ganges delta, has been a substantial source of arsenic (As). The Sundarban Biosphere Reserve (SBR), located in the southwestern part of the world's largest Ganges delta, suffers from As contamination in groundwater. Therefore, assessment of groundwater vulnerability is essential to ensure the safety of groundwater quality in SBR. Three data-driven algorithms, i.e. "logistic regression (LR)", "random forest (RF)", and "boosted regression tree (BRT)", were used to assess groundwater vulnerability. Groundwater quality and hydrogeochemical characteristics were evaluated by Piper, United States Salinity Laboratory (USSL), and Wilcox's diagram. The result of this study indicates that among the applied models, BRT (AUC = 0.899) is the best-fit model, followed by RF (AUC = 0.882) and LR (AUC = 0.801) to assess groundwater vulnerability. In addition, the result also indicates that the general quality of the groundwater in this area is not very good for drinking purposes. The applied methods of this study can be used to evaluate the groundwater vulnerability of the other aquifer systems.

Bayesian merging of numerical modeling and remote sensing for saltwater intrusion quantification in the Vietnamese Mekong Delta

Saltwater intrusion has become one of the most concerning issues in the Vietnamese Mekong Delta (VMD) due to its increasing impacts on agriculture and food security of Vietnam. Reliable estimation of salinity plays a crucial role to mitigate the impacts of saltwater intrusion. This study developed a hybrid technique that merges satellite imagery with numerical simulations to improve the estimation of salinity in the VMD. The salinity derived from Landsat images and by numerical simulations was fused using the Bayesian inference technique. The results indicate that our technique significantly reduces the uncertainties and improves the accuracy of salinity estimates. The Nash-Sutcliffe coefficient is 0.74, which is much higher than that of numerical simulation (0.63) and Landsat estimation (0.6). The correlation coefficient between the ensemble and measured salinity is relatively high (0.88). The variance of the ensemble salinity errors (5.0 ppt2) is lower than that of Landsat estimation (10.4 ppt2) and numerical simulations (9.6 ppt2). The proposed approach shows a great potential to combine multiple data sources of a variable of interest to improve its accuracy and reliability wherever these data are available.

Designing dynamic groundwater management strategies through a composite groundwater vulnerability model: Integrating human-related parameters into the DRASTIC model using LightGBM regression and SHAP analysis

Groundwater nitrate contamination has emerged as a pressing global concern. Given its potential for long-term impacts on aquifers, protective measures should primarily focus on prevention. Drawing on the theory of groundwater vulnerability (GV), the original DRASTIC model and parameters related to human activities are employed as inputs and integrated with the LightGBM regression algorithm to facilitate nitrate index (NI) prediction tasks. The SHAP analysis is conducted to effectively examine the contribution of parameters to the NI prediction and interpret the issue of parameter interactions. In addition, to mitigate the limitations of the intrinsic GV model, a composite nitrate index (CNI) is developed by linearly combining the DRASTIC index with the NI. The framework presented in this study provides adaptive strategies for managing groundwater resources over different time periods. A representative region for arid and semiarid climates, the Yinchuan region, is studied using the framework. As compared to 2012, the intrinsic GV index has changed spatially in 2022. Human activities have increased the influence of the nitrate concentration as shown by the Pearson correlation coefficient of -0.082 between the DRASTIC index and nitrate concentration. A significant increase in pollution levels was predicted by NI, ranging from -0.116 to 0.968. According to SHAP analysis, the significant increase in NI levels in 2022 was mainly due to high-value industrial and agricultural production. In 2022, 12.02% of the areas had an increase of at least 0.549 in the CNI. 42.1% of the areas were classified as moderate or high CNI levels. The farm was identified as a high-contributing source to nitrate pollution. The small-scale agricultural and livestock activities in non-urban areas also contribute to groundwater pollution. Dynamic groundwater management strategies need to be implemented in high-growth and high-level CNI areas.

Implementation and evaluation of different techniques to modify DRASTIC method for groundwater vulnerability assessment: a case study from Bouficha aquifer, Tunisia

Groundwater vulnerability assessment has nowadays evolved into an essential tool towards proper groundwater protection and management, while the DRASTIC method is included among the most widely applied vulnerability assessment methods. However, the high uncertainty of the DRASTIC method mainly associated with the subjectivity in assigning parameters ratings and weights has driven many researchers to apply various methods for improving its efficiency. In this context, in the present study, different techniques were implemented with the aim of modifying the DRASTIC framework and thus enhancing its performance for groundwater vulnerability assessment in the Bouficha aquifer, Tunisia. In a first stage, the land use type (L) was incorporated as an additional parameter in the typical DRASTIC framework, thus taking into consideration the impact of anthropogenic activities on groundwater vulnerability. Subsequently, the rating and weighting systems of the developed DRASTIC-L framework were modified through the application of statistical methods (DRASTIC-L-SA) and genetic algorithms (GA) (DRASTIC-L-GA) in an attempt to investigate and compare both linear and nonlinear modifications. To evaluate the various vulnerability frameworks, correlation between vulnerability values and nitrate concentrations, expressed as Spearman's rank correlation coefficient (ρ) and Correlation Index (CI), was examined. The results revealed that the DRASTIC-L-GA framework developed by applying a fully GA-based optimization procedure provided the highest values in terms of the performance metrics used, making it the most suitable for the study area. In addition, the aquifer under study was found to be less vulnerable to pollution when employing the typical DRASTIC framework instead of the modified ones, leading to the conclusion that the former substantially underestimates pollution potential in the study area.

Groundwater vulnerability assessment of elevated arsenic in Gangetic plain of West Bengal, India; Using primary information, lithological transport, state-of-the-art approaches

Deterioration of groundwater quality is a long-term incident which leads unending vulnerability of groundwater. The present work was carried out in Murshidabad District, West Bengal, India to assess groundwater vulnerability due to elevated arsenic (As) and other heavy metal contamination in this area. The geographic distribution of arsenic and other heavy metals including physicochemical parameters of groundwater (in both pre-monsoon and post-monsoon season) and different physical factors were performed. GIS-machine learning model such as support vector machine (SVM), random forest (RF) and support vector regression (SVR) were used for this study. Results revealed that, the concentration of groundwater arsenic compasses from 0.093 to 0.448 mg/L in pre-monsoon and 0.078 to 0.539 mg/L in post-monsoon throughout the district; which indicate that all water samples of the Murshidabad District exceed the WHO's permissible limit (0.01 mg/L). The GIS-machine learning model outcomes states the values of area under the curve (AUC) of SVR, RF and SVM are 0.923, 0.901 and 0.897 (training datasets) and 0.910, 0.899 and 0.891 (validation datasets), respectively. Hence, "support vector regression" model is best fitted to predict the arsenic vulnerable zones of Murshidabad District. Then again, groundwater flow paths and arsenic transport was assessed by three dimensions underlying transport model (MODPATH). The particles discharging trends clearly revealed that the Holocene age aquifers are major contributor of As than Pleistocene age aquifers and this may be the main cause of As vulnerability of both northeast and southwest parts of Murshidabad District. Therefore, special attention should be paid on the predicted vulnerable areas for the safeguard of the public health. Moreover, this study can help to make a proper framework towards sustainable groundwater management.

Quantifying the groundwater total contamination risk using an inclusive multi-level modelling strategy

This paper investigates aggregated risks in aquifers, where risk exposures may originate from different contaminants e.g. nitrate-N (NO₃-N), arsenic (As), boron (B), fluoride (F), and aluminium (Al). The main goal is to develop a new concept for the total risk problem under sparse data as an efficient planning tool for management through the following methodology: (i) mapping aquifer vulnerability by DRASTIC and SPECTR frameworks; (ii) mapping risk indices to anthropogenic and geogenic contaminants by unsupervised methods; (iii) improving the anthropogenic and geogenic risks by a multi-level modelling strategy at three levels: Level 1 includes Artificial Neural Networks (ANN) and Support Vector Machines (SVM) models, Level 2 combines the outputs of Level 1 by unsupervised Entropy Model Averaging (EMA), and Level 3 integrates the risk maps of various contaminants (nitrate-N, arsenic, boron, fluoride, and aluminium) modelled at Level 2. The methodology offers new data layers to transform vulnerability indices into risk indices and thereby integrates risks by a heuristic scheme but without any learning as no measured values are available for the integrated risk. The results reveal that the risk indexing methodology is fit-for-purpose. According to the integrated risk map, there are hotspots at the study area and exposed to a number of contaminants (nitrate-N, arsenic, boron, fluoride, and aluminium).

Novel machine learning algorithms to predict the groundwater vulnerability index to nitrate pollution at two levels of modeling

The accurate mapping and assessment of groundwater vulnerability index are crucial for the preservation of groundwater resources from the possible contamination. In this research, novel intelligent predictive Machine Learning (ML) regression models of k-Neighborhood (KNN), ensemble Extremely Randomized Trees (ERT), and ensemble Bagging regression (BA) at two levels of modeling were utilized to improve DRASTIC-LU model in the Miryang aquifer located in South Korea. The predicted outputs from level 1 (KNN and ERT models) were used as inputs for ensemble bagging (BA) in level 2. The predictive groundwater pollution vulnerability index (GPVI), derived from DRASTIC-LU model was adjusted by NO₃-N data and was utilized as the target data of the ML models. Hyperparameters for all models were tuned using a Grid Searching approach to determine the best effective model structures. Various statistical metrics and graphical representations were used to evaluate the superior predictive performance among ML models. Ensemble BA model in level 2 was more precise than standalone KNN and ensemble ERT models in level 1 for predicting GPVI values. Furthermore, the ensemble BA model offered suitable outcomes for the unseen data that could subsequently prevent the overfitting issue in the testing phase. Therefore, ML modeling at two levels could be an excellent approach for the proactive management of groundwater resources against contamination.

Formulating Convolutional Neural Network for mapping total aquifer vulnerability to pollution

Aquifer vulnerability mapping to pollution is topical research activity, and common frameworks such as the basic DRASTIC framework (BDF) suffer from the inherent subjectivity. This paper formulates an artificial intelligence modeling strategy based on Convolutional Neural Network (CNN) to decrease subjectivity. This formulation considers three definitions of intrinsic, specific, and total vulnerabilities. Accordingly, three CNN models are trained and tested to calculate IVI, SVI, and TVI, respectively referring to the intrinsic, specific, and total vulnerability indices. The formulation is applied in an unconfined aquifer northwest of Iran and delineates hotspots within the aquifer. The area under curve (AUC) values derived by the receiver operating curves evaluate the vulnerability indices versus nitrate concentrations. The AUC values for BDF, IVI, SVI, and TVI are 0.81, 0.91, 0.95, and 0.95, respectively. Therefore, CNNs significantly improve the results compared to BDF, but IVI, SVI, and TVI have approximately identical performances. However, the visual comparison between their results provides evidence that significant differences exist between the spatial patterns despite identical AUC values.