Eutrophication, which is the existence of excessive nutrient loads, is a process that can endanger hydro-ecosystems like rivers. As two of the most effective water quality parameters, this study applies six advanced interpretable boosting machine learning (ML) models (AdaBoost, GBM, XGBoost, LightGBM, HistGBM, & CatBoost) to predict Total Phosphorus (TP) and OrthoPhosphate (OP) concentrations and also to assess the best boosting model in terms of accuracy, tendency, and computational cost. To build the boosting models, 12 water quantity (e.g., river discharge and sediment transport) along with water quality parameters (e.g., turbidity, dissolved oxygen, TKN, and nutrient concentrations) were examined to understand phosphorus dynamics. Preliminary outcomes of the study show that TP and OP values are negatively correlated with DO and pH values, while TKN has the greatest positive influence on them. Model interpretability analysis using local interpretable model-agnostic explanations (LIME) revealed that ammonia and nitrogen affect phosphorus levels differently depending on background nutrient conditions. At low TP and OP levels, these nutrients had a negative impact, whereas at higher phosphorus concentrations, they contributed positively to eutrophication. Also, the Shapley Additive Explanations (SHAP) analysis, as a global model interpretability method, highlights the impact of nitrate and ammonia (as chemical variables) and discharge (as a hydrological parameter) on the eutrophication process. The XGBoost achieved the highest predictive performance, GBM was the optimal model in terms of showing the least bias error, and LightGBM was the most effective model in terms of efficient computation.
扫码关注我们
求助内容:
应助结果提醒方式:
