Estimating monthly suspended sediment load of the Brahmaputra River by integrating machine learning models with unsupervised outlier detection

Suspended sediment load (SSL) modeling is crucial in hydrology and river engineering for infrastructure design, river basin management, flood mitigation, and channel change studies. Monitoring SSL often is costly and labor-intensive, especially in the Brahmaputra-Jamuna River, a complex braided river with a nonlinear relation between discharge and sediment. In this study, the potential of six standalone machine learning (ML) models—Random Forest (RF), Support Vector Regression (SVR), Adaptive Boosting (AdaBoost), Elastic Network (EN) Regression, Gradient Boosting Regression Tree (GBRT), and eXtreme Gradient Boosting (XGB)—is investigated for estimating SSL by integrating antecedent hydro-sediment information and utilizing the Isolation Forest data outlier detection technique, which influences model robustness and accuracy. Monthly observed SSL, discharge (Q), and ERA5 reanalysis rainfall (R) data from 1976 to 2023 were used to calibrate and validate the model. Various input scenarios were developed by combining SSL, discharge, and rainfall with different temporal lags. The results reveal that discharge is the best predictor for modeling SSL (Pearson correlation coefficient, r = 0.72), and the most effective input combination is the discharge and rainfall in the current month and two previous months, and the SSL in the two previous months (Q_t, Q_t-1, Q_t-2, R_t, R_t-1, R_t-2, SSL_t-1, SSL_t-2). All applied models yield good performance (Nash-Sutcliffe Efficiency, NSE ranging from 0.71 to 0.82), except SVR (NSE = 0.68), which showed relatively weak performance, while XGB outperformed the other methods (NSE = 0.82). A Taylor diagram and reliability analyses also confirmed XGB’s superiority in estimating SSL. Using outlier detection, the results show performance improvements of 9.76% for XGB, 7.59% for GBRT, 7.41% for RF, 6.67% for AdaBoost, 5.48% for EN, and 4.41% for SVR, indicating that outliers significantly impacted the models. The current study identified the best ML model for SSL estimation at Badadurabad, Bangladesh, and showed how to use antecedent hydro-sediment data to estimate the current sediment load more easily and affordably in braided rivers with weak hydro-sediment coupling, supporting sustainable sediment management in similar rivers.