<p>Bridge failures resulting from local scour at piers often occur during flood events where flows are unsteady and their durations are usually insufficient for the scour process to reach equilibrium. This study employs extreme gradient boosting (XGBoost), random forest (RF), and support vector machine (SVM) algorithms to develop machine learning models for predicting the temporal evolution of scour depth at piers under unsteady flow conditions. The models were trained and validated with systematically collected flume experiment data, with particular emphasis on evaluating the influence of input parameters on predictive performance. Results indicate that hydraulic parameters serve as more suitable inputs than the scour depth. Among these parameters, water depth demonstrates slightly stronger predictive influence than flow velocity. Meanwhile, expressing flow velocity in terms of effective bed shear stress or effective stream power yields better predictive performance. Furthermore, incorporating time rates of change of water depth and flow velocity can enhance model accuracy under unsteady conditions. The combination of water depth, flow velocity, and their temporal derivatives yields optimal predictions. With this input configuration, XGBoost and RF models outperform both SVM and established deterministic formulas, while maintaining a non-zero optimal forecast horizon. The findings offer guidance on developing efficient and accurate machine learning models for scour prediction.</p>