Modeling the Ecological Impact of Phosphorus in Catchments with Multiple Environmental Stressors

Abstract

The broken phosphorus (P) cycle has led to widespread eutrophication of freshwaters. Despite reductions in anthropogenic nutrient inputs that have led to improvement in the chemical status of running waters, corresponding improvements in their ecological status are often not observed. We tested a novel combination of complementary statistical modeling approaches, including random‐effect regression trees and compositional and ordinary linear mixed models, to examine the potential reasons for this disparity, using low‐frequency regulatory data available to catchment managers. A benthic Trophic Diatom Index (TDI) was linked to potential stressors, including nutrient concentrations, soluble reactive P (SRP) loads from different sources, land cover, and catchment hydrological characteristics. Modeling suggested that SRP, traditionally considered the bioavailable component, may not be the best indicator of ecological impacts of P, as shown by a stronger and spatially more variable negative relationship between total P (TP) concentrations and TDI. Nitrate‐N (p < 0.001) and TP (p = 0.002) also showed negative relationship with TDI in models where land cover was not included. Land cover had the strongest influence on the ecological response. The positive effect of seminatural land cover (p < 0.001) and negative effect of urban land cover (p = 0.030) may be related to differentiated bioavailability of P fractions in catchments with different characteristics (e.g., P loads from point vs. diffuse sources) as well as resilience factors such as hydro‐morphology and habitat condition, supporting the need for further research into factors affecting this stressor–response relationship in different catchment types. Advanced statistical modeling indicated that to achieve desired ecological status, future catchment‐specific mitigation should target P impacts alongside multiple stressors. Core Ideas Soluble reactive P (SRP) alone was not the best indicator of diatom response. Total P (TP) association with diatoms was more spatially variable than SRP. Nitrate‐N and TP have a combined negative effect on the ecological response. Seminatural land use had the most important influence on ecological response. We recommend catchment‐specific mitigation of multiple stressors.

Probability maps of anthropogenic impacts affecting ecological status in European rivers

“…The importance of land use as major explaining variable affecting aquatic ecology is in line with other studies (e.g. Villeneuve et al, 2018 , Glendell et al, 2019 , Schmidt et al, 2019 , Tang et al, 2020 ). However, interpretation of regression coefficients must be taken cautiously as logistic regressions, as other data-driven empirical models, do not identify cause-effect relationships.…”

“…While other machine learning methods, like Regression Trees or Random Forest may often result in higher predictive performance (e.g. Grizzetti et al, 2017 , Glendell et al, 2019 ), logistic regressions maintain an explicit effect of explanatory variables on the probability outcome, i.e. the marginal effect, thus their interpretation is more straightforward than for machine learning approaches.…”

A review of the current state of process-based and data-driven modelling: guidelines for Lake Erie managers and watershed modellers

“…Statistical models encompass a wide range of statistical techniques to quantify measures of central tendency and uncertainty analysis with event-mean concentrations (Lee et al, 2011); assemble databases of average nutrient-export coefficients per land use/land cover (Reckhow et al, 1980;Harmel et al, 2008;Hertzberger et al, 2019); test BMP effectiveness with analysis of variance (Maniquiz et al, 2010a); conduct concentration-discharge analysis with linear regression techniques to identify chemostatic and chemodynamic patterns, threshold responses and hysteresis effects (Moatar et al, 2017); employ multivariate statistical methods for source identification (Mudge, 2007;Chen et al, 2013); develop multiple-linear regressions, spatial-stream network and geographically weighted-regression models to identify which landscape metrics correlate with nutrient pollution (Isaak et al, 2014;Scown et al, 2017); relate event-mean concentrations with rain intensity (Maniquiz et al, 2010b); compute daily, monthly, seasonal, annual and decadal nutrient loads (Lee et al, 2016); conduct socio-geospatial modelling studies (Wilson, 2015); investigate the disparity between overall nutrient loading reduction and a lack of improvement in ecological status of water bodies with regression trees for clustered data, compositional linear mixed models, and ordinary linear mixed models (Glendell et al, 2019); test multiple working hypotheses to investigate causality between main stressors and environment with structural equation modelling (Fan et al, 2016;Ryberg, 2017); apply Bayesian Belief Networks to reproduce non-linear impacts from BMP implementation and generate cascade effects on ecosystem services (Nash and Hannah, 2011;Landuyt et al, 2013).…”

Modeling Water Quality in Watersheds: From Here to the Next Generation

“…Both statistical and machine learning methods provide ways of estimating temporal dynamics and patterns of variability in constituent concentrations and loads, assisting the identification of system stressors ( Glendell et al, 2019 ), setting of regulatory targets ( Jung et al, 2020 ), and model simplification ( Jackson-Blake et al, 2017 ). However, extrapolation of fitted behavior beyond the ranges and environmental conditions of measured data requires extreme caution.…”