To address the issue of degraded predictive model performance caused by high-dimensional breach features and strong correlations among features, a peak discharge prediction model for earth-rock dam breaches was developed by integrating the Lasso algorithm with the XGBoost model. This model uses the Spearman correlation coefficient method to analyze the correlations between features, utilizes the Lasso algorithm for further feature selection, and obtains an optimal feature subset by eliminating redundant features. The feature subset is input into the XGBoost model to predict the peak breach discharge. Comparative results with support vector regression and ridge regression machine learning models show that the proposed model exhibits strong nonlinear information mining capability, effectively reduces the dimensionality of high-dimensional features, and improves the prediction accuracy while reducing model complexity.