.png){kind=icon}
.png){kind=icon}
Return
Parameter-specific classification of surface water quality using machine learning and explainable AI in heterogeneous watersheds
.png){kind=icon}
PRE
AI
H
M
M
DOI：10.1016/j.pce.2026.104413.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Marked spatio-temporal variability in surface water quality across watersheds with different land-use characteristics complicates conventional methods. This research proposes a parameter-specific machine learning framework based on individual water quality parameters to classify four nutrients (NO3-N, NH3-N, TKN, PO4-P) and seven dissolved trace elements (Al, Cr, Cu, Fe, Mn, Ni, Zn) quality classes across four watersheds. The framework was developed using 994 surface water measurements from 33 stations and 67 watershed-scale features derived from field campaigns, geospatial analyses, and historical datasets. Random Forest (RF), Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoost), Artificial Neural Network (ANN) were evaluated under four experimental scenarios. These scenarios incorporated mutual information-driven feature selection, iterative feature subset construction, and grid search hyperparameter optimization to improve parameter-specific classification performance. The optimized scenarios outperformed default configurations, yielding F-scores ranging from 0.81 to 0.96 across nutrients and from 0.59 to 0.80 for trace elements. XGBoost achieved the highest performance for 8 of the 11 parameters, including PO4-P (F-score = 0.839) and NH3-N (Fscore = 0.882). In contrast, ANN outperformed other models for parameters characterized by pronounced class imbalance and high variability, such as NO3-N, Zn, Mn. Model interpretability was assessed using SHAP for the best-performing models. SHAP analysis revealed that nutrient classes were primarily driven by land-use, agricultural activities, and point-source discharges, whereas trace elements were governed by more complex interactions involving industrial pressures and meteorological variability. Overall, the framework provides an interpretable basis for parameter-specific surface water quality classification. These insights support data-driven watershed management under contrasting land-use pressures.
Keywords:
Surface water quality
Watershed-scale analysis
Nutrients
Trace elements
Machine learning classification
Explainable artificial intelligence
Journal
IF:
4.1
Papers:
3.3K
Citations:
6.9K
Organization
Cited Papers