The influence of pH and temperature on benthic chlorophyll-a: Insights from SHAP-XGBoost and random forest models

Résumé fourni par la source

Biological threats to river health relate to algal biomass, for which benthic chlorophyll– a (chl– a ) is an indicator; consequently, predicting chl– a helps understand ecosystem dynamics. There is little information on machine learning predictive models of benthic chl– a and input parameters in lotic ecosystems, and to fill this gap, we predict benthic chl– a levels in China's Thousand Islands Lake (TIL) watershed using machine learning algorithms. Water samples for nutrient and metal analysis were collected across 147 sites in the TIL catchment. We employed Random Forest (RF), eXtreme gradient boosting (XGBoost) and SHAP-enhanced eXtreme gradient boosting (SHAP XGBoost) models, alongside Support Vector Regression (SVR), to predict chl– a levels in diverse reaches and identify the key determinants. The XGBoost outperformed the RF model in the test, training and validation datasets. In the SHAP XGBoost, pH was the most important characteristic, followed by mean average temperature (AT). The SVR demonstrated that AT is vital for the upper and middle catchment reaches, while pH is more important in the lower reaches. In partial dependence plots, the chl– a concentration depended highly on pH and AT. High pH and AT released P from stream colloids, lowered colloid adsorption, increasing chl– a concentration. We concluded that the SHAP XGBoost model could be used to identify the key determinants of chl– a from chemical and physical variables in the lotic system. • Water pH and temperature predict benthic chlorophyll- a in freshwater. • SHAP XGboost performed better than the random forest model. • Temperature and water pH predict chlorophyll-a in upper, middle, and lower reaches.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.