Objective: This study focuses on the critical task of predicting the number of people to be evacuated (i.e., relocation number) during flood disasters. Accurate predictions of relocation numbers are vital for ensuring timely resource allocation and efficient disaster management, particularly in flood-prone areas where rapid decision-making can drastically mitigate the adverse impacts of the disaster. Methods: This research developed a robust relocation number prediction framework that combines feature selection and data augmentation techniques using the extreme gradient boosting (XGBoost) model, a widely used gradient-boosting machine learning algorithm. The model was built using historical data from flood events across China between 2014 and 2018. These events included meteorological and geographical features and the relocation number during each disaster. Feature selection was accomplished using Shapley additive explanations (SHAP), a game theory method for measuring the contribution of each feature to the model predictions. The selected features were then fed into the XGBoost model for training. A data augmentation strategy was also introduced to handle the challenge of limited training samples. This strategy involved the injection of Gaussian noise using a weighted k-nearest neighbors method to generate synthetic data points that preserved the local structure of the data, thereby enhancing the model's robustness and generalization ability. Results: The study demonstrates that the XGBoost model performs well with the selected features and augmented data. Initially, the model is trained on a small dataset, leading to satisfactory accuracy but limited generalization ability. However, after applying data augmentation, the model's performance significantly improves, especially for extreme values in the data. The testing phase reveals that R² improves from 0.854 to 0.967, indicating a substantial increase in the model's predictive accuracy. Additionally, the root mean square error decreases from 0.296 to 0.123, signifying a considerable reduction in prediction error. These results highlight the effectiveness of combining feature selection and data augmentation to enhance the predictive power of the model. The feature selection process, guided by SHAP, identifies several key predictors that play a dominant role in determining population relocation demand. Among the most influential features are the maximum 3-day cumulative rainfall (MCR) and the maximum cumulative rainfall over the 15 days prior to the event (MRPE). These features are the most important in predicting the relocation number during flood events. Conclusions: The proposed relocation number prediction framework, integrating feature selection through SHAP and data augmentation techniques, is a highly effective tool for forecasting the relocation number during flood disasters. The XGBoost model, after optimization through Bayesian hyperparameter tuning and data augmentation, demonstrates significantly improved prediction accuracy and robustness. This approach can be instrumental in supporting disaster management teams with more reliable forecasts, allowing for better planning and more timely deployment of resources. Improving the model's ability to generalize to unseen data ensures accurate predictions even in regions with limited historical data. Thus, this study provides a valuable decision-making support tool for emergency response teams, helping to streamline resource allocation and evacuation planning during flood disasters and thereby minimizing the impact of the disaster on human lives and infrastructure.