Interpretable machine learning for seasonal nitrate and total phosphorus concentrations integrating hydrologic, land-use, and ecological indicators

Authors: Daniels, Gbenga; Rajbhandari, Meghna; Carneiro Marques, Amanda

Source journal: Journal of environmental management (2026-11-19)

DOI: 10.1016/j.jenvman.2026.131408

Abstract

Diffuse nutrient pollution remains difficult to assess because nutrient concentrations reflect interacting watershed, hydrochemical, ecological, and hydrologic conditions that vary across space and time. Although machine-learning models incorporate multiple environmental predictors, the contribution of different predictor domains can be difficult to distinguish when all variables are evaluated simultaneously. This study developed an interpretable machine-learning framework to predict seasonal nitrate and total phosphorus concentrations across seven watersheds in southeastern Pennsylvania using observations from 2002 to 2024. Physicochemical, land-cover, ecological, climatic, and hydrologic information were progressively integrated through four predictor scenarios, allowing the added predictive value of each domain to be evaluated. Random Forest and Extreme Gradient Boosting (XGBoost) models were independently optimized for each scenario. Predictive performance improved as broader environmental information was incorporated, with the largest gain occurring after land-cover variables were added to the physicochemical baseline. The environmental-only M4 XGBoost models achieved test R(2) values of 0.858 for nitrate and 0.770 for total phosphorus, with RMSE values of 0.624 and 0.170 mg L(-1), respectively. Repeated site-stratified validation of the nitrate M4 model yielded a mean test R(2) of 0.834 +/- 0.026, indicating stable performance across alternative partitions. SHapley Additive exPlanations identified forest cover, developed land cover, and specific conductance as dominant predictors for both nutrients, while discharge contributed more strongly to nitrate and dissolved oxygen to total phosphorus prediction. These reveal that land-cover composition offers significant prognostic insights beyond in-stream conditions, whereas hydrochemical and hydrologic factors aid in differentiating nutrient-specific trends. The framework could guide the prioritizing of watershed monitoring and the evaluation of nutrient dynamics.

Read the full text

Related topics

Related open access articles

Abstract and bibliographic data from PubMed (U.S. National Library of Medicine). Copyright of the abstract belongs to the original publisher or authors.