Toro et al. (2025) Groundwater quality prediction for drinking and irrigation uses in the Murcia region (Spain) by artificial neural networks

Identification

• Journal: Applied Water Science
• Year: 2025
• Date: 2025-09-01
• Authors: Eva Maria García del Toro, M. Isabel Más-López, Luis F. Mateo, M. Ángeles Quijano
• DOI: 10.1007/s13201-025-02605-z

Research Groups

• Departamento de Ingeniería Civil: Hidráulica, Energía y Medio Ambiente, Universidad Politécnica de Madrid, Spain
• Departamento de Matemática E Informática Aplicadas a Las Ingenierías Civil y Naval, Universidad Politécnica de Madrid, Spain
• Centro I+D+I en Infraestructuras Civiles Inteligentes y Sostenibles (CIVILIS), Universidad Politécnica de Madrid, Spain

Short Summary

This study developed and evaluated artificial neural network models (RProp-MLP and PNN DDA) to predict groundwater quality for drinking and irrigation in the semiarid Murcia region, Spain, using two defined quality indices (DWQI and IWQI) and demonstrating the superior performance of RProp-MLP.

Objective

• To develop and evaluate artificial neural network (ANN) models (Resilient Backpropagation Multilayer Perceptron (RProp-MLP) and Probabilistic Neural Network with Dynamic Decay Adjustment (PNN DDA)) for predicting groundwater quality in the Murcia region, Spain, for both drinking and irrigation purposes, using newly defined quality indices (DWQI and IWQI) with a minimal set of input parameters.

Study Configuration

• Spatial Scale: Regional scale, covering 11,180 km² within the Segura River Basin, specifically 38 groundwater bodies in the Region of Murcia, Spain.
• Temporal Scale: 24 years (2000–2023).

Methodology and Data

• Models used:
  • Artificial Neural Networks (ANN): Resilient Backpropagation Multilayer Perceptron (RProp-MLP), Probabilistic Neural Network with Dynamic Decay Adjustment (PNN DDA).
  • Statistical methods: Principal Component Analysis (PCA) for parameter weighting and dimensionality reduction, Spearman correlation analysis for input variable selection, Kolmogorov–Smirnov test for normality, Chi-square (χ²) test for model validation.
  • Framework: KNIME analytics platform.
• Data sources:
  • Groundwater samples: 1962 samples from 159 sampling stations across 38 groundwater bodies in the Region of Murcia, Spain.
  • Data provider: Segura Hydrographic Confederation (CHS) electronic office (MITECO, n.d.).
  • Parameters: pH, electrical conductivity (EC), concentrations of bicarbonate (HCO₃⁻), calcium (Ca²⁺), chloride (Cl⁻), magnesium (Mg²⁺), nitrate (NO₃⁻), potassium (K⁺), sodium (Na⁺), and sulfate (SO₄²⁻).
  • Derived irrigation indicators: Kelly ratio (KR), Magnesium hardness (MH), Potential salinity (PS), Sodium absorption rate (SAR), Soluble sodium percentage (%Na).

Main Results

• Principal Component Analysis (PCA) for the Drinking Water Quality Index (DWQI) resulted in a two-component model explaining 68.768% of the total variance, with Factor 1 dominated by parameters related to salinity and natural mineralization (EC, Na⁺, Cl⁻, Mg²⁺, SO₄²⁻, Ca²⁺).
• PCA for the Irrigation Water Quality Index (IWQI), including irrigation indicators, resulted in a three-component model explaining 76.982% of the total variance, with Factor 1 related to saline intrusion and mineralization (SO₄²⁻, PS, Ca²⁺, EC, Mg²⁺, Cl⁻) and Factor 2 to sodium and magnesium impact on soil (%Na, Na⁺, KR, SAR, MH).
• Groundwater quality classification indicated that 58.2% of samples were unsuitable for drinking and 48.2% were unsuitable for irrigation, with the poorest quality predominantly in the southern and southeastern areas due to high salinity and nitrate contamination.
• Input variables selected for ANN models were Ca²⁺, Cl⁻, EC, Mg²⁺, NO₃⁻, Na⁺, SO₄²⁻ (7-input configuration) and EC, Mg²⁺, NO₃⁻, Na⁺, SO₄²⁻ (5-input configuration); 3-input configurations were found to be inaccurate.
• Both RProp-MLP and PNN DDA models demonstrated high accuracy in predicting DWQI and IWQI with 7- and 5-input configurations.
• The RProp-MLP model consistently outperformed the PNN DDA model across all evaluated metrics, achieving coefficients of determination (R²) greater than 0.99 and minimal prediction errors.
• The 7-input RProp-MLP model yielded the best performance, with approximately 80% of predictions exhibiting a relative error below 10% when validated against independent data.
• The 5-input RProp-MLP model also showed acceptable performance, with 60% of predictions having a relative error below 10%, offering a suitable balance between accuracy and computational cost.

Contributions

• Development and validation of two novel groundwater quality indices (DWQI and IWQI) specifically tailored for drinking and irrigation uses in a semiarid Mediterranean region.
• Application and comparative evaluation of two ANN models (RProp-MLP and PNN DDA) within a KNIME framework for groundwater quality prediction using minimal input data.
• Demonstration of the superior predictive performance of RProp-MLP, particularly with a 5-7 input variable configuration, for groundwater quality assessment in a complex hydrochemical environment.
• Provision of a practical and cost-effective tool for groundwater management in semiarid regions prone to salinization and agricultural pollution.
• Utilization of a substantial, long-term dataset (2000-2023) covering a wide geographical area and diverse aquifers, enhancing the robustness and generalizability of the findings within the study region.

Funding

This research received no external funding.

Citation

@article{Toro2025Groundwater,
  author = {Toro, Eva Maria García del and Más-López, M. Isabel and Mateo, Luis F. and Quijano, M. Ángeles},
  title = {Groundwater quality prediction for drinking and irrigation uses in the Murcia region (Spain) by artificial neural networks},
  journal = {Applied Water Science},
  year = {2025},
  doi = {10.1007/s13201-025-02605-z},
  url = {https://doi.org/10.1007/s13201-025-02605-z}
}