Interpreting machine learning models based on SHAP values in predicting suspended sediment concentration

Machine learning (ML) has become a powerful tool for predicting suspended sediment concentration (SSC). Nonetheless, the ability to interpret the physical process is considered the main issue in applying most of ML approaches. In this regard, the current study presents a novel framework involving four standalone ML models (extra trees (ET), random forest (RF), categorical boosting (CatBoost), and extreme gradient boosting (XGBoost)) and their combination with genetic programming (GP). Three metrics (coefficient of correlation ( r ), root mean square error (RMSE), and Nash–Sutcliffe model-fit efficiency (NSE)) and a more advanced interpretation system SHapley Additive exPlanations (SHAP) are used to assess the performance of these models applied to hydro-climatic datasets for prediction of SSC. The calibration process was based on data from 2016 to 2020, and the validation was done for 2021 data. Further description and application of the framework are provided based on a case study of the Bouregreg watershed. The results revealed that all implemented models are efficient in SSC prediction with NSE, RMSE, and r varying from 0.53 to 0.86, 1.20 to 2.55 g/L, and 0.83 to 0.91 g/L respectively. Box plot diagrams confirm the enhanced performance of these combined models, and the best-performing ones for the four hydrological stations being the combined RF+GP model at the Aguibat Ziar station, the combined XGBoost+GP model at the Ain Loudah station, the CatBoost model at the Ras Fathia station, and the RF model at the Sidi M ed Cherif station. The interpretability results showed that flow ( Q ) and seasonality ( S ) are the features most impacting SSC. These outcomes indicate that the applied models can extract accurate and detailed information from the interactions between the hydroclimatic factors and the generation of sediment by erosion (output). ML approaches illustrated the good reliability and transparency of the models developed for predicting SSC in a semi-arid setting, offered new perspectives for reducing ML models' "black box" character, and provided a useful source of information for assessing the consequences of SSC on water quality. The SHAP system and exploring other interpretable techniques are recommended to provide further information in future research. In addition, incorporating additional input data could enhance SSC predictions and deepen understanding of sediment transport dynamics.