Machine Learning Models Predict Arsenic Contamination

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Aug 27 2024

In an article published in the journal Water, researchers focused on predicting arsenic (As) contamination in groundwater, a significant health threat in Asia. Using hydro-chemical, geological, and soil parameters, the authors applied multiple linear regression (MLIR) and random forest (RF) models.

The RF models outperformed MLIR in estimating As concentrations and predicting contamination risks in China's Hetao Basin and Bangladesh, demonstrating their robustness in managing As contamination through key environmental predictors.

Background

Geogenic As contamination in groundwater is a critical environmental health issue, particularly in South and Southeast Asia. Despite extensive research, predicting As concentrations remains challenging due to the complex interplay of geochemical processes, hydrological factors, and limited data availability.

Traditional models like MLIR have been widely used but often fall short in accuracy due to their linear nature. To address these gaps, this study employed RF and multivariate logistic regression (MLOR) to model As contamination in the Hetao Basin and Bangladesh. By integrating hydro-chemical, soil, and geological data, this research aimed to improve prediction accuracy and provide insights into the varying contamination mechanisms across different geographical regions.

Study Area and Analytical Methods

The researchers focused on assessing As contamination in groundwater within the Hetao Basin and the Bengal Delta, employing a range of hydrogeochemical, geological, and soil parameters. The Hetao Basin, characterized by its complex sedimentary structure and varying groundwater levels, contrasted with the Bengal Delta, shaped by massive sediment deposition from the Ganges–Meghna–Brahmaputra river system.

High As concentrations were influenced by factors such as the potential of hydrogen (pH), oxidation-reduction potential, and the presence of ions like calcium ion (Ca²⁺) and chloride ion (Cl⁻), alongside soil properties like organic carbon density and clay content.

Data were collected from groundwater wells in both regions, followed by preprocessing to ensure data integrity. Statistical analyses were conducted to evaluate key metrics, and models were developed to predict As concentrations. Multicollinearity was assessed using variance inflation factor (VIF) and Pearson’s correlation coefficients, ensuring that the predictive models were reliable.

Feature selection for modeling revealed that factors such as pH and dissolved organic carbon (DOC) were consistently significant across both regions. The RF regression and MLIR models were used to predict As contamination, while RF classification and MLOR models assessed the probability of high-risk contamination. The models were validated using a subset of the data, confirming their effectiveness in predicting As contamination in groundwater.

Comparative Analysis and Model Performance

The authors investigated the hydrochemical and geological characteristics of groundwater in the Hetao Basin, China, and three regions in Bangladesh (Rajshahi, Dhaka, Chittagong), focusing on their impact on As contamination. In the Hetao Basin, groundwater was marked by high salinity and elevated concentrations of ions like Cl− and sulfate ion (SO₄²⁻), resulting in alkaline pH conditions and high total dissolved solids (TDS).

This environment promoted the desorption of As from mineral surfaces, particularly under high ionic strength and alkaline conditions. Conversely, Bangladesh's groundwater exhibited lower salinity, and more neutral pH, and was predominantly influenced by rainfall recharge, leading to lower concentrations of dissolved ions.

In terms of organic content, Bangladesh showed higher organic carbon density, which, combined with reducing conditions, significantly influences As mobility. In contrast, the Hetao Basin, characterized by a lacustrine deposition environment, has a higher soil organic carbon (SOC) content and greater cation exchange capacity (CEC), affecting contaminant retention and As behavior.

The study's modeling results revealed that the RF regression model outperformed the MLIR model in predicting As concentrations, capturing spatial variability more effectively in both regions. Additionally, the RF classification model showed superior accuracy in classifying groundwater As contamination probabilities compared to the MLOR model, demonstrating robustness across different geographic contexts. The researchers concluded that the interplay of redox conditions, organic matter degradation, and competitive adsorption processes played a crucial role in controlling As mobility in groundwater.

Conclusion

In conclusion, the researchers underscored the critical role of hydro-chemical and geological factors in predicting As contamination in groundwater. The RF regression model significantly outperformed traditional MLIR models in accuracy, demonstrating its effectiveness in managing As contamination.

The authors highlighted the importance of incorporating advanced predictive models like RF regression into environmental management strategies to improve predictions and mitigate risks. Recommendations included ongoing model calibration, comprehensive monitoring, strategic management, and fostering interdisciplinary research and international collaboration. Addressing data quality and incorporating anthropogenic factors will further enhance predictive accuracy and safeguard public health.