Accurate baseflow separation is essential for understanding groundwater contributions, managing water resources, and improving hydrological modeling. In this study, we propose a novel machine learning-based baseflow separation method, HyGBR-Baseflow (Hybrid Gradient Boosting Regressor-Based Baseflow) separation, developed and applied across major Indian river basins over the period 1961–2021. We used a Gradient Boosting Regressor trained on a hybrid target dataset comprising: (i) high-confidence reference baseflow points identified during the non-monsoon season and (ii) a subset of Eckhardt-filter-based baseflow during the monsoon season. The hybrid labeling strategy addresses the scarcity of reliable baseflow reference data during monsoon periods. We used a comprehensive set of predictors, including catchment-averaged precipitation, temperature, model-derived baseflow, surface runoff, soil moisture, and outputs from existing graphical and digital filter-based separation methods. We evaluate model performance using NSE, R², KGE, and the refined Index of Agreement (d_r). We show that HyGBR-Baseflow separation outperforms traditional baseflow separation techniques across diverse hydrological regimes and seasons, demonstrating robustness and hydrological realism. The seasonally adaptive design of the method enhances its applicability in season-driven catchments and data-scarce settings. Our results provide a reliable tool for long-term subsurface flow analysis, essential for supporting sustainable water management, drought assessments, and hydrological modeling.