Study region: The Albujon 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. Study focus: 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. New hydrological insights for the region: 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 Albujon watershed and demonstrate how ML interpretability can support water-quality management in Mediterranean coastal basins where traditional approaches struggle with irregular-sampling and nonstationarity.