Luận án tiến sĩ phát triển phương pháp máy học cho phân loại nước ngầm

MINISTRY OF EDUCATION VIETNAM ACADEMY OF SCIENCE AND TRAINING AND TECHNOLOGY GRADUATE UNIVERSITY OF SCIENCE AND TECHNOLOGY Michael Omar AN APPROACH OF ENSEMBLE SPATIAL MACHINE LEARNING FOR GROUNDWATER DRINKABILITY CLASSIFICATION DOCTORAL DISSERTATION ON INFORMATION SYSTEM Hanoi - 2025 BỘ GIÁO DỤC VIỆN HÀN LÂM KHOA HỌC VÀ ĐÀO TẠO VÀ CÔNG NGHỆ VIỆT NAM HỌC VIỆN KHOA HỌC VÀ CÔNG NGHỆ Michael Omar NGHIÊN CỨU PHÁT TRIỂN PHƯƠNG PHÁP HỌC MÁY KẾT HỢP THÔNG TIN KHÔNG GIAN CHO BÀI TOÁN PHÂN LOẠI CHẤT LƯỢNG NƯỚC NGẦM LUẬN ÁN TIẾN SĨ MÁY TÍNH NGÀNH: HỆ THỐNG THÔNG TIN MÃ SỐ: 9 48 01 04 Xác nhận của Học viện Người hướng dẫn 1 Người hướng dẫn 2 Khoa học và Công nghệ (Ký, ghi rõ họ tên) (Ký, ghi rõ họ tên) Hà Nội - 2025 MINISTRY OF EDUCATION VIETNAM ACADEMY OF SCIENCE AND TRAINING AND TECHNOLOGY GRADUATE UNIVERSITY OF SCIENCE AND TECHNOLOGY Michael Omar AN APPROACH OF ENSEMBLE SPATIAL MACHINE LEARNING FOR GROUNDWATER DRINKABILITY CLASSIFICATION DOCTORAL DISSERTATION ON COMPUTER MAJOR: INFORMATION SYSTEMS CODE: 9 48 01 04 Graduation University of Advisor 1 Advisor 2 Science and Technology’s confirmation (Signature, Full Name) (Signature, Full Name) Hanoi - 2025 DECLARATION I hereby declare that the results presented in the discussion are my re- search work under the guidance of the guidelines. The data and results presented in the project discussion are completely honest and have not been published in previous works. Reference data have been fully cited. The thesis uses cited in- formation from various sources, and all citations are clearly acknowledged.

The research findings co-authored with other researchers have been included in the thesis with their consent. This thesis was completed during my time as a doctoral student at the Vietnam Academy of Science and Technology. year 2025 Thesis Author Michael Omar Acknowledgments Throughout the process of researching and completing my thesis, I have been fortunate to receive guidance, assistance, invaluable feedback, and encour- agement from a diverse group of individuals, including scientists, educators, peers, and family members. First and foremost, I would like to extend my deepest gratitude to As- soc.

Tran Thi Ngan, Assoc. Nguyen Long Giang and Assoc. Le Hoang Son. Their dedicated mentorship and support have been pivotal through- out my research journey.

I also would like to express my sincere appreciation to the faculty and scientists at the Institute of Information and Technology, the Institute of Infor- mation Technology VNU. Their insightful suggestions significantly contributed to the advancement of my research. My gratitude extends to the Board of Directors and the Training De- partment at Graduate University of Science and Technology, Vietnam Academy of Science and Technology for providing me with the conducive environment necessary to fulfill my research objectives. Lastly, I must acknowledge my colleagues, family, and friends, whose un- wavering encouragement, shared experiences, support, and assistance have been instrumental in helping me overcome challenges and achieve the results presented in this thesis.

year 2025 Signed by: Michael Omar i Contents Symbols and Abbreviations iv List of Tables v List of Figures viii Introduction 1 Chapter 1.Groundwater Drinkability Classification 8 1.1 Introduction to Groundwater Drinkability Classification .2 ML/DL Methods .3 Hybrid Spatial Models .4 Gaps and Summary .5 Research Method to Address Gaps .3 Study Areas: India and Vietnam .1 Mekong Delta, Vietnam .3 Hydrological Context & Site Rationale .4 Evaluation Metrics & Scenario .1 Encoding of Spatial Coordinates .2 Derived Features from Raw Measurements .3 Incorporating Domain Knowledge into Feature Creation .7 Generalization and Transferability to Other Geographical Regions .Proposed Ensemble Spatial Machine Learning Methods 45 2.4 Pros and Cons .Results and Evaluations 80 3.1 Objective of the Evaluation .2 Validation of AI-LGBM .1 Datasets and Preprocessing .2 Hyperparameter Optimization and Tuning .3 Pros and Cons .4 Performance Evaluation and Comparison .5 Appended (Post-Optimization) ML Results: AI-LGBM .3 Validation of PSO-SCNN .1 Datasets and Preprocessing .2 Hyperparameter Optimization and Tuning .3 Performance Evaluation and Comparison .4 Appended (Post-Optimization) Result — PSO–SCNN .4 Model’s Performance Comparison .1 Failure Case Analysis .2 Implications for Groundwater Quality Classification .3 Feature Importance and Future Directions .1 AI-LGBM Findings .2 PSO-SCNN Findings. 128 Conclusion and Future Development 130 APPENDIX A: REPRODUCIBILITY 134 Symbols and Abbreviations No. Abbreviation Description 1 AI Artificial Intelligence 2 AI-LGBM Auto Immune Light Gradient Boosting Machine 3 CNN Convolutional Neural Network 4 DL Deep Learning 6 GIS Geographic Information System 7 GWQ Groundwater Quality 8 IPCC Intergovernmental Panel on Climate Change 9 ML Machine Learning 10 PSO Particle Swarm Optimization 11 PSO-SCNN Particle Swarm Optimization-Spatial Convolutional Neural Network 12 R² Coefficient of Determination 13 RMSE Root Mean Square Error 14 SCNN Spatial Convolutional Neural Network 15 SELU Scaled Exponential Linear Unit 16 SHAP Shapley Additive Explanations 17 SVM Support Vector Machine 18 TDS Total Dissolved Solids 19 LightGBM Light Gradient Boosting Machine 20 XGBoost Extreme Gradient Boosting 21 LSTMs Long Short-Term Memory 22 MAE Mean Absolute Error 23 AUC Area Under Curve 24 ANOVA Analysis of Variance 25 WQI Water Quality Index 26 WQC Water Quality Class 27 GWQ Groundwater Quality v List of Tables 1.1 Summary of Classical Hydrological Methods .2 Observed Values of Water Quality Parameters .3 Assigned Weights to Water Quality Parameters .4 Calculated Sub-Indices for Water Quality Parameters .5 Water Quality Classification Based on WQI Values .6 Machine Learning Methods for Hydrological Water Quality Assess- ment .7 Deep Learning Methods in Hydrology .8 Descriptive Statistics of Groundwater Parameters in the Mekong Delta (Vietnam) .9 Descriptive Statistics of Groundwater Parameters in Odisha (India) 27 1.10Comparison of Hydrological Characteristics .11Justification for Selecting Odisha and the Mekong Delta .12Evaluation Metrics for Model Performance .13 Hybrid Spatial-AI Models Used in Groundwater Classification .14Baseline Models for Groundwater Quality Classification .15Dataset Overview and Column Types .16Dataset Overview and Column Types for Indian Water Quality Dataset 33 1.17Data Preprocessing Steps .1 Benefits of Combining Components in the AI-LGBM Model .2 Hyperparameter Search Space and Final Values for AI-LGBM .3 Performance Comparison: Default vs Optimized AI-LGBM .4 AI–LGBM strengths, caveats, and recommended mitigations.5 Model Input Analysis .6 Comparison of Learning Optimizers .7 Key PSO-SCNN Hyperparameter Values .8 Effect of PSO Parameter Settings on PSO-SCNN Performance (Val- idation Set) .9 Hyperparameter Tuning for Groundwater Models .10PSO–SCNN strengths, caveats, and recommended mitigations.1 Hyperparameter Search Space and Final Values for AI-LGBM .2 Comparison of the Average Value of Performance Metrics of All Models in Odisha .3 Comparison of the Average Value of Performance Metrics of All Models in Vietnam .4 Comparison of AI-LGBM with Baseline Models .5 Comparative Performance of the Models .6 Comparison of Proposed Models with Advanced Methods .7 Comparison of the Average Value of Performance Metrics of All Models in Odisha .8 Comparison of the Average Value of Performance Metrics of All Models in Vietnam .9 Comparison of AI-LGBM with Baseline Models .10Performance Metrics for Various Models in Odisha Dataset .11Performance Metrics for Various Models in Vietnam Dataset .12Model Comparison (Avg.15Comparison of AI-LGBM Vs DL, (Open source) Datasets .16Model Performance Vietnam Dataset Comparison with Log Loss .17Comparison of PSO-SCNN with AI-LGBM and Baseline Models .18Effect of PSO Parameter Settings on PSO-SCNN Performance (Val- idation Set) .19PSO controller configuration used for SCNN tuning.20PSO–SCNN hyperparameter search space.21Effect of PSO controller settings on PSO–SCNN (validation set illustration).22Hyperparameter optimization method comparison.24Aggregate comparison of proposed models vs.25Held-out testing on the Vietnam dataset.26Held-out testing on the India (Odisha) dataset.27Cross-validation results (mean ± SD) of proposed models.28Metric Analysis Performance .29Model Comparison Table: Deep Learning Models .30Model Comparison Table: Machine Learning Models .31Convergence Epochs and Time to Convergence .33PSO–SCNN validation metrics after optimization.34PSO–SCNN post-optimization results on held-out test sets.35PSO–SCNN cross-validation summary (mean ± SD).36Advance Model Performance Vietnam (Testing Set) post-run .37 Metric Analysis Performance .38Comparison of Proposed Models with ML Baseline Models .39Cross-Validation Results (Mean ± SD) of Proposed Models .40Ablation Study: Quantitative Impact of Components .41Convergence Epochs of Ablation Models .42Training Time and Memory Consumption Comparison for AI-LGBM and PSO-SCNN Models .43One-way ANOVA comparing model performance metrics.44One-way ANOVA comparing performance across regions.45Significance Test Results for Methods and Datasets. 124 viii List of Figures 1.1 Traditional ML flow diagram for water quality analysis and classi- fication .2 Geographical Context of Study Areas (a) Location of the Mekong Delta (Source: Mekong River Commission); (b) Provincial Extent within the Mekong Delta.3 Hydro-geological map of the Odisha study area.4 Box-Plot analysis .5 Scatter Plot Analysis .6 AI-LGBM Model Feature Importance .7 SHAP Summary Plot for AI-LGBM Model .8 SHAP Interpretation and Implications .9 Spatial Visualization of Groundwater Quality Classification .10 Evolution from Traditional Methods to Hybrid Spatial-Aware ML Framework .1 Proposed System Model of the Artificial Intelligence Framework .2 Proposed AI-LGBM Methodological Flowchart .3 PSO-SCNN Spatial Model Architectures .4 PSO-SCNN Flowchart .5 Sensitivity Analysis – AUC vs Kernel Size .6 Sensitivity Analysis – AUC vs Number of Filters .7 Sensitivity Analysis – AUC vs Learning Rate .8 Parameter Validation Results: A table showing model performance with different hyperparameters.1 Optuna Optimization History (Objective: Weighted F1-Score) .2 Hyperparameter Importance Analysis via Optuna .3 SHAP Summary Plot for Optimized AI-LGBM Model .4 Model Loss and Accuracy on Vietnam Dataset .5 Mean Error and K-Value Comparison .6 Comparative analysis of model performance in Vietnam and India 87 3.7 Bivariate Analysis and Data Outlier (1) .8 Bivariate Analysis and Data Outlier (2) .9 Comparative analysis and performance of K-NN and SMOTE for Vietnam and India .10 Sensitivity Analysis of Learning Rate and Number of Leaves.

The left plot shows the relationship between learning rate and F1 Score/Ac- curacy, while the right plot illustrates the sensitivity of the F1 Score/Accuracy with respect to the number of leaves.12 Classification of water quality in Vietnam based on the model’s classification.13 Training Time vs AUC for All Models .14 PSO-SCNN Training and Validation Loss .15 PSO-SCNN Training and Validation Accuracy .16 Validation Loss - PSO-SCNN vs Deep Baselines .17 Spatial visualization of groundwater quality classification .18 Comparison of Water Quality Visualizations in Odisha .19 Side-by-side performance visualizations for PSO-SCNN.21 SHAP Summary Plot for AI-LGBM Model .22 Overlay of predicted unsafe zones with actual contamination areas 115 3.23 Feature importance highlighting key factors in water quality clas- sification .24 Averaged p-values for each feature in water quality classification .25 Ablation Study Results on the Impact of Removing Model Com- ponents .26 Ablation Study: AUC Scores of Model Variants .27 PSO-SCNN Prediction Grid (Longitude vs Latitude) .28 Feature Range Differences (Correct vs Error) .29 Confusion Matrix — PSO-SCNN .31 Feature Distribution for Misclassified vs Correctly Classified Samples121 1 Introduction Research Context Ensuring the safety of drinking water is a paramount global challenge, essential for public health, environmental sustainability, and economic devel- opment. This need is critically amplified by a growing global population that intensifies pressure on finite water resources [1–3]. An estimated two billion peo- ple still lack access to safely managed drinking water, making the advancement of robust water quality assessment methods a global health imperative [4]. Con- taminated sources are a primary vector for waterborne diseases and expose pop- ulations to chemical and pathogenic contaminants, creating a persistent public health crisis [5].

The urgency for a new assessment paradigm is compounded by mounting environmental pressures from industrial and agricultural runoff [6], as well as the spatiotemporal variability of water quality, which is being exacerbated by climate change [7]. Traditional water quality monitoring, which relies on manual field sam- pling and laboratory analysis, is increasingly ill-suited to address the scale of this challenge. These methods are inherently inefficient, slow, and unscalable, par- ticularly in resource-constrained regions [8], necessitating a shift towards more advanced, automated solutions. Problem Statement In response to these limitations, modern machine learning (ML) and deep learning (DL) offer powerful tools for water quality prediction [9–13], their ap- plication is often undermined by a critical flaw: spatial blindness.

Most stan- 2 dard models fail to account for spatial autocorrelation the principle that nearby samples are related which leads to unreliable predictions [14]. This issue is com- pounded by naive validation protocols, such as random k-fold cross-validation, which are statistically unsound for geospatial data and yield overly optimistic performance metrics [15, 16].