Predicting water quality variables using gradient boosting machine: global versus local explainability using SHapley Additive Explanations (SHAP)

Water quality assessment is critical for ensuring the health of aquatic ecosystems and managing water resources effectively. However, accurately predicting key water quality variables remains challenging due to the complex interactions between environmental factors and anthropogenic influences. In the present investigation, a new modelling framework is proposed for better prediction of three water quality variables, namely: (i) dissolved oxygen concentration (DO), (ii) water turbidity (TU), and (iii) water Chlorophyll a (Chl-a). Six machine learning models, i.e., adaptive boosting (AdaBoost), categorical boosting (CatBoost), histogram gradient boosting (HistGBRT), light gradient boosting machine (LightGBM), natural gradient boosting (NGBoost), and extreme gradient boosting (XGBoost), both applied and compared based on the combination of a large number of water quality variables. All models were developed using data collected from three stations: (i) USGS 05543010 Illinois River at Seneca, Illinois County, (ii) USGS 05586300 Illinois River at Florence, Illinois County, and (iii) USGS 05553700 Illinois River at Starved Rock, Illinois County, USA. The SHapley additive explanations (SHAP) was adopted in the present study for model interpretability and feature ranking. Furthermore, all models were compared using various numerical indices and graphical representations. From the obtained results we can draw the following conclusion. DO concentration can be predicted very well with high numerical performances, and the CatBoost model was found to be the best one exhibiting excellent numerical index: RMSE (0.430), MAE (0.326), R (0.980) and NSE (0.961), respectively. For Chl-a, all models were found to be less accurate and the best performances were obtained using the LightGBM with RMSE (5.916), MAE (4.294), R (0.892) and NSE (0.795), respectively. Finally, for water TU, none of the models were found to be accurate and very poor performances were obtained. Finally, the use of the SHAP has significantly helped in better understanding the overall contribution of the various water variables in the finale prediction.