What question did this study set out to answer?

The aim is to evaluate interpretable machine learning approaches for predicting nitrate and phosphate concentrations under data-scarcity in the Albujón watershed.

April 4, 2026Open Access

Interpretable machine learning for nitrate and phosphate prediction in the Mar Menor watershed, Spain, under data-scarce conditions

Key Points

The aim is to evaluate interpretable machine learning approaches for predicting nitrate and phosphate concentrations under data-scarcity in the Albujón watershed.
Developed four tree-based algorithms: Decision Tree, Random Forest, Gradient Boosting, and XGBoost.
Used predictors such as streamflow, conductivity, year, and day-of-year.
Adopted a non-time-series modeling approach to handle data-scarcity.
Applied SHAP analysis to interpret the best-performing model and quantify predictor influence.
XGBoost achieved strong predictive accuracy (NSE = 0.66 for both nitrate and phosphate) on unseen data.
Cross-validation uncertainty was ±0.05 for nitrate and ±0.15 for phosphate.
Nitrate estimates were most influenced by conductivity and interannual trends.
Phosphate variability was associated with conductivity and seasonal patterns.

Abstract

The Albujón watershed (Spain) is a semi-arid basin dominated by intensive irrigated agriculture. It is the primary drainage system for the Mar Menor, Europe’s largest hypersaline lagoon, representing a critical ecosystem under severe anthropogenic stress. To address nutrient-monitoring challenges in data-scarce Mediterranean catchments, this study evaluates the interpretable machine learning (ML) against the traditional LOADEST framework. Four tree-based algorithms (Decision Tree, Random Forest, Gradient Boosting, and XGBoost) were developed to estimate nitrate and phosphate concentrations using predictors robust to irregular sampling (streamflow, conductivity, year, and day-of-year). A non-time-series modelling was adopted to overcome data-scarcity. SHAP (SHapley Additive exPlanations) was applied to interpret the best-performing model and quantify the relative influence of hydrological and temporal predictors. XGBoost achieved strong predictive accuracy on unseen data (NSE = 0.66 for both nitrate and phosphate), with cross-validation uncertainty of ±0.05 and ±0.15. SHAP analysis revealed nutrient-specific patterns in the statistical predictors. Nitrate estimates were strongly influenced by conductivity and interannual trends, while phosphate variability was associated with conductivity and seasonality, indicating the importance of ionic conditions and temporal patterns in this ephemeral system. These findings provide the first interpretable, data-driven characterisation of nutrient behaviour in the Albujón watershed and demonstrate how ML interpretability can support water-quality management in Mediterranean coastal basins where traditional approaches struggle with irregular-sampling and non-stationarity. • Nutrient estimates were reliable without time-series models, using temporal features. • Interpretable tree-based models identified nutrient-specific statistical predictors. • For nitrate, conductivity and interannual trends showed the strongest influence. • For phosphate, conductivity and seasonality were the key predictors.

Interpretable machine learning for nitrate and phosphate prediction in the Mar Menor watershed, Spain, under data-scarce conditions

Key Points

Abstract

Cite This Study