Purpose <p>Sediment transport capacity significantly influences river morphological changes and flow characteristics. The nonlinear relationships between sediment discharge (Qs​) in rivers and its influencing factors pose challenges for accurate modeling.</p> Materials and methods <p>This study presents novel computational models aimed at improving Qs estimation accuracy. We developed and evaluated supervised deep neural networks (DNNs) models employing various optimization algorithms, including stochastic gradient descent (SGD), adaptive gradient algorithm (Adagrad), adaptive learning rate (Adadelta), root mean square propagation (RMSprop), and extended adaptive moment estimation (Adamax), for predicting Qs ​. Feature selection, as a classification, was applied to enhance the predictive power of these models. A dataset from 38 rivers and streams in South Korea, spanning from 2007 to 2020, was utilized, and data mining techniques were employed to classify input variables for modeling.</p> Results and discussion <p>Results demonstrate that models incorporating feature selection significantly outperformed those without. Specifically, feature selection enhanced Qs estimation by 13–47%, depending on the optimizer used in the DNN models. Channel width, average cross-sectional depth, water surface slope, and the geometric mean size of bed material had a high significant impact in the Qs​ modeling. Notably, the deep learning model with the Adamax optimizer and feature selection achieved an accuracy of 96.34%. A comparison between the proposed models and previous studies revealed that the accuracy of Qs estimation improved by 25% in terms of the correlation coefficient.</p> Conclusions <p>These findings indicate that the proposed techniques not only surpass previous studies but also provide a reliable framework for long-term sediment discharge estimation. This investigation highlights the utility of feature selection as a robust predictive approach for riverine Qs ​modeling.</p>