Advancing SDG 6.3.2 with machine learning-based virtual sensors for high-frequency nutrient monitoring

Abstract

Reliable monitoring of Nitrogen and Phosphorus in ambient waters is critical for achieving Sustainable Development Goal (SDG) indicator 6.3.2, yet the high cost of in-situ nutrient sensors limits global data coverage, especially in low-middle-income countries (LMICs). This study presents a novel virtual sensing framework that replaces expensive nutrient sensors with Machine Learning models trained on affordable Baseline-Features (Dissolved Oxygen, pH, Electrical Conductivity) and enhanced with low-cost features (Turbidity, Temperature, Flow). To our knowledge, this is the first study to integrate the REFORMS checklist into the end-to-end development of virtual sensing for SDG 6.3.2 nutrient monitoring, ensuring transparency, reproducibility, and policy relevance. Using Extra Trees as the best performing model, rigorously benchmarked through LazyPredict, spot checking, and hyperparameter tuning (Grid Search, Randomized Search, Bayesian Optimization), the framework achieved state-of-the-art predictive accuracy (R² up to 0.98) across contrasting urban (The-Cut) and rural (River-Enborne) catchments. SHAP analysis further demonstrated interpretable feature contributions, with Electrical Conductivity and Turbidity consistently emerging as dominant drivers. The results establish that Baseline-Features are sufficient for stable rural systems, while urban systems require additional features to achieve SDG-compliant accuracy. Beyond technical performance, the study contributes policy recommendations for UNEP and LMIC agencies, including equivalency testing guidelines and capacity-building for national monitoring programs. This framework advances virtual sensing from research concept to an operationally viable tool for bridging nutrient data gaps in SDG 6.3.2 reporting.