Abstract: To achieve high-accuracy prediction of the spatial distribution of underwater noise during the operational phase of offshore wind farms, this study proposes an environment-driven ED-XGBoost machine learning model. Based on field measurement data from two 6-MW turbines in an offshore wind farm located in Fujian Province, an input feature vector was constructed, comprising distance, bearing angle, depth, wind speed, wind direction, water temperature, and sound speed profile. The performance of linear regression (LR), Gaussian process regression (GPR), and XGBoost models was compared using 10-fold cross-validation. Results show that the ED-XGBoost model achieved the best performance, with a root mean square error (RMSE) of 1.79 dB, a mean absolute error (MAE) of 1.45 dB, and a coefficient of determination (R²) of 0.98. Further investigation reveals that wind speed contributes most to underwater noise levels, followed by depth and bearing angle; the captured nonlinear response relationships are consistent with the physical mechanisms of acoustic propagation. The proposed method requires no complex physical parameters, offers high computational efficiency and strong physical interpretability, and provides a reliable tool for rapid forecasting of underwater noise from offshore wind farms and for assessing their marine ecological impacts.