Leveraging novel machine learning models in predicting groundwater irrigation suitability in southeastern Nigeria: A hydrogeochemical approach

Abstract

Groundwater suitability for irrigation is a crucial component of sustainable agriculture, particularly in regions like Obosi, southeastern Nigeria, where freshwater resources are limited and agricultural demands are high. Despite growing interest in improving prediction models for groundwater suitability, there remains a gap in employing advanced machine learning techniques alongside hydrogeochemical analysis for precise and reliable assessments. This study aims to address this gap by evaluating the performance of six novel machine learning models—Artificial Neural Networks (ANNs), Random Forest (RF), Support Vector Machine (SVM), CatBoost (CatB), AdaBoost (AB), and Gradient Boosting (GB)—in predicting irrigation suitability using 42 groundwater samples. The hydrogeochemical approach in this study employs Piper, Durov, and Schöller plots to analyze groundwater composition and geochemical processes. The findings reveal a sodium chloride (NaCl)-dominated water type, influenced by dissolution, mixing, and reverse ion exchange, with implications for water quality and management. The physicochemical analysis showed pH values ranging from 4.49 to 6.29, electrical conductivity (EC) between 8.16 and 101.7 µS/cm, and chloride concentrations from 8.43 to 32.65 mg/L. All measured parameters fell within permissible limits. However, irrigation indices such as sodium absorption ratio (SAR), sodium percentage (Na%), Kelly's ratio (KR), soluble sodium percentage (SSP), and potential salinity (PS) indicated the groundwater was unsuitable for irrigation. Spatial analysis showed distinct regional patterns, with SAR, Na%, and KR aligning NE–SW, while PS followed a S–N direction. Hydrogeochemical trends revealed Cl⁻ dominance in 71% of samples, with evidence of dissolution, ion exchange, and mixing processes. Principal component analysis highlighted complex inter-parameter relationships, with pH and EC explaining 14.94% of variance. ANN and RF consistently outperformed other models across multiple metrics, which include RMSE, MAPE, R², adjusted R², Akaike Information Criterion (AIC), and Bayesian Information Criterion (BIC), demonstrating strong predictive accuracy and generalization. By contrast, SVM, CatB, AB, and GB exhibited weaker performance and susceptibility to overfitting. This study underscores the potential of ANN and RF for reliable groundwater irrigation suitability assessments. Their integration with hydrogeochemical analysis offers a scalable, cost-effective approach applicable in similar regions globally, enabling enhanced precision in agricultural water management. The findings provide a scientific basis for policymakers, farmers, and water resource managers to implement targeted groundwater management strategies, ensuring sustainable irrigation practices. By integrating advanced machine learning with hydrogeochemical analysis, this study offers a cost-effective and scalable approach for assessing water suitability, which can be applied to similar agricultural regions worldwide.