📄 earth_science

Explainable GeoAI Framework for Groundwater Pollution Prediction, Vulnerability Mapping, and Sustainable Water Intelligence in Semi-Arid Regions

This study presents an integrated GeoAI–ML–XAI framework for semi-arid regions that utilizes high-accuracy machine learning models and novel explainability metrics to predict groundwater pollution, identify key contamination drivers, and generate sustainable, interpretable vulnerability indices for improved environmental decision-making.

Original authors: Yogesh Iyer Murthy

Published 2026-08-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yogesh Iyer Murthy

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Earth's underground water as a giant, invisible sponge hidden beneath our feet. This sponge doesn't just hold water; it acts like a massive, slow-motion filter that has been soaking up rain, river water, and even the chemicals we use on our farms and in our homes for thousands of years. Scientists call the study of this hidden world "hydrogeology," and when they start using super-smart computer programs to figure out what's happening inside that sponge, they call it "GeoAI" (Geospatial Artificial Intelligence). Usually, these computer programs are like "black boxes": you feed them data, and they spit out an answer, but they won't tell you why they gave that answer. It's like a magic 8-ball that says "Yes" or "No" but refuses to explain its reasoning. However, a new field called "Explainable AI" (XAI) is trying to crack open that box, acting like a translator that turns the computer's secret code into plain English so humans can understand the logic behind the prediction. This matters because groundwater is the lifeblood for drinking and farming in many dry places; if we don't know exactly what's poisoning it or where the trouble spots are, we can't fix the problem before it's too late.

In this study, a researcher named Yogesh Iyer Murthy decided to play detective with the underground water in a dry, semi-arid region of India called Beenagunj. He treated the groundwater like a complex puzzle, trying to figure out how dirty it was and, more importantly, why it was getting dirty. Instead of just guessing, he built a massive digital laboratory. He collected water samples from 50 different villages and measured a whole bunch of ingredients in the water, like saltiness (TDS), chalkiness (hardness), and various chemicals like nitrates and fluoride.

To solve the puzzle, he didn't just use one computer program; he threw 28 different types of machine learning algorithms at the data. Think of these algorithms as 28 different detectives, each with a unique way of solving crimes. Some were simple and linear, like a detective who only looks for straight lines. Others were complex and deep, like a detective who can see hidden patterns in a maze. The goal was to see which detective could best predict the "Pollution Index of Groundwater" (PIG)—a score that tells you how toxic the water is.

The results were fascinating. The "Trilayered Neural Network" detective turned out to be the superstar, predicting the pollution levels with almost perfect accuracy (a score of 0.996 out of 1). It was like this detective could see the future of the water quality with crystal clarity. However, the study didn't stop at just finding the best detective. The real magic happened when the researcher used "Explainable AI" tools, specifically SHAP and LIME. These tools acted like a spotlight, shining on the specific ingredients that were causing the pollution.

The spotlight revealed that the main culprits weren't always what people expected. The biggest drivers of pollution were Total Dissolved Solids (TDS), chlorides, conductivity (how well the water carries electricity, which indicates saltiness), and total hardness. These were the heavyweights pushing the pollution scores up. Interestingly, the study found that while things like nitrates and fluoride were present, they didn't have as huge of an impact on the overall pollution score in this specific area as the salty, hard-water ingredients did. The researcher also noticed that the pollution wasn't random; it was clustered. The southern and south-eastern parts of the region were like a "pollution hotspot," where the water was significantly dirtier, while the northern areas were much safer.

To make sure these findings were rock-solid and not just a fluke of one specific computer program, the researcher invented a new tool called the "Explainability Stability Index" (ESI). Imagine you ask 28 different experts for their opinion on a mystery. If they all say the same thing, you can trust the answer. The ESI measured how consistent the 28 different computer detectives were about which chemicals were the bad guys. The results showed that TDS and chlorides were consistently identified as the top troublemakers across almost every single model, giving the findings a high level of trust.

Finally, the researcher combined all this smart computer logic with the map of the area to create two new, super-smart scores: the "Explainable Groundwater Quality Index" (XGWQI) and the "Explainable Groundwater Vulnerability Index" (EGVI). These weren't just old-school scores; they were adaptive. They changed based on what the computer models actually saw in the data, rather than relying on fixed rules. The study found that these new, smart indices were much better at spotting the most dangerous "hotspots" than the traditional methods used before. For example, in villages like Mukundpur and Kousiya, the new indices flagged the water as "Very High" risk, confirming that these areas needed immediate attention.

In short, this paper didn't just predict where the water was dirty; it built a transparent, smart system that explains why it's dirty and where the danger is coming from. It proved that by using a team of 28 different AI detectives and asking them to explain their reasoning, we can get a much clearer, more reliable picture of groundwater health than ever before, helping communities in dry regions protect their most precious resource.

Technical Summary: Explainable GeoAI Framework for Groundwater Pollution Prediction, Vulnerability Mapping, and Sustainable Water Intelligence in Semi-Arid Regions

Problem Statement
Groundwater contamination in semi-arid regions, particularly in developing nations like India, is driven by complex hydrogeochemical interactions and anthropogenic activities such as intensive agriculture and urbanization. While traditional groundwater quality assessment relies on hydrochemical indices (e.g., Pollution Index of Groundwater or PIG), these methods often fail to capture nonlinear relationships between parameters and lack the ability to identify dominant contamination drivers. Conversely, while Machine Learning (ML) offers high predictive accuracy, many models function as "black boxes," limiting their utility in environmental policy and decision-making where transparency and scientific explanation are paramount. Existing studies often focus on single model families or predictive accuracy without rigorous model diagnostics, spatial intelligence, or robust explainability across multiple algorithms.

Methodology
The study proposes an integrated GeoAI–Machine Learning–Explainable Artificial Intelligence (GeoAI–ML–XAI) framework applied to 50 villages in the Beenagunj Block of Madhya Pradesh, India. The methodology proceeds through the following stages:

Data Acquisition and Preprocessing: Groundwater samples were collected from borewells during the pre-monsoon season. Key hydrochemical parameters included TDS, chlorides, fluoride, nitrates, total hardness, conductivity, alkalinity, turbidity, pH, calcium, and magnesium. The Pollution Index of Groundwater (PIG) was computed as the target variable.
Machine Learning Modeling: Twenty-eight distinct regression algorithms were employed to predict PIG values. These spanned seven model families: Linear Models, Decision Trees, Support Vector Machines (SVM), Gaussian Process Regression (GPR), Neural Networks (ANN), Kernel Regression, and Ensemble methods. The dataset was split into 70% training, 15% testing, and 15% validation, with hyperparameters optimized via Bayesian optimization.
Explainable AI (XAI) Integration: To address the "black box" issue, the study utilized SHapley Additive exPlanations (SHAP) for global feature importance and Local Interpretable Model-Agnostic Explanations (LIME) for local instance interpretation.
Novel Metric Development:
- Explainability Stability Index (ESI): A new metric was introduced to quantify the consistency of hydrochemical feature contributions across all 28 algorithms, calculated as $ESI_j = 1 - (\sigma_j / \mu_j)$ , where $\sigma$ and $\mu$ are the standard deviation and mean of SHAP values, respectively.
- Explainable Groundwater Quality Index (XGWQI): An adaptive index integrating SHAP-derived weights and ESI stability scores to dynamically adjust parameter importance.
- Explainable Groundwater Vulnerability Index (EGVI): A composite index combining PIG, spatial vulnerability scores (VSI), and ESI to assess environmental risk.
Spatial Analysis: Inverse Distance Weighting (IDW) interpolation and spatial autocorrelation (Global and Local Moran's I) were used to map contamination hotspots and clustering behaviors.

Key Results

Predictive Performance: Among the 28 algorithms, the Trilayered Neural Network achieved the highest predictive accuracy with an $R^2$ of 0.996 and RMSE of 0.009 on the testing set. Gaussian Process Regression (GPR) models (specifically Rational Quadratic and Exponential kernels) demonstrated robust and stable performance with $R^2$ values exceeding 0.90. Linear models like Stepwise Linear Regression also performed well ( $R^2 \approx 0.96$ ), while complex tree-based and high-degree polynomial SVMs showed signs of overfitting.
Dominant Contamination Drivers: SHAP and LIME analyses consistently identified TDS, chlorides, conductivity, and total hardness as the primary drivers of groundwater pollution. TDS exhibited the highest positive contribution (approx. 32–37% influence). In contrast, parameters like pH, calcium, and magnesium showed lower or negative contributions, suggesting they are more influenced by geogenic equilibrium than anthropogenic pollution.
Spatial Distribution: Spatial analysis revealed significant heterogeneity. The southern and south-eastern regions (e.g., Mukundpur, Kousiya, Pathariya) were identified as severe contamination hotspots with PIG values $\ge$ 1.0, characterized by high TDS and chloride levels. The northern regions exhibited lower pollution levels. Global Moran's I ( $0.63, p < 0.001$ ) confirmed significant positive spatial autocorrelation, validating the presence of clustered contamination patterns.
Stability and Indices: The ESI confirmed that TDS and chlorides possess "Very High" stability (ESI > 0.85) across different model architectures, reinforcing their reliability as contamination indicators. The proposed XGWQI and EGVI successfully identified high-risk zones (e.g., Mukundpur, Kousiya) with greater sensitivity than conventional indices (WQI, PIG, EWQI), particularly in areas dominated by ionic enrichment.

Significance and Claims
The paper claims that the proposed GeoAI-XAI framework represents a significant advancement over conventional groundwater assessment by transitioning from static, expert-weighted indices to adaptive, data-driven, and interpretable systems. Key contributions include:

Enhanced Interpretability: By integrating SHAP, LIME, and the novel ESI, the framework provides transparent, scientifically reliable explanations for model predictions, overcoming the limitations of black-box ML in environmental management.
Robust Decision Support: The development of XGWQI and EGVI allows for vulnerability assessment that explicitly accounts for the stability of contamination drivers, enabling targeted management strategies (e.g., salinity control vs. nitrate mitigation).
Spatial Intelligence: The framework effectively combines geospatial interpolation with ML to identify not only where contamination is severe but why (via dominant drivers) and how stable these drivers are across different modeling approaches.
Applicability: The study positions this framework as a transferable tool for sustainable water resource management in semi-arid, agriculturally stressed regions, supporting goals related to public health protection and climate resilience.

The author acknowledges limitations, including the reliance on a single season's data, the use of village-level rather than precise borewell coordinates for spatial interpolation, and the computational complexity of running 28 models with XAI. However, they assert that the framework offers a substantial improvement in transparency, hotspot identification, and adaptive assessment capabilities compared to traditional methods.

More like this