To quantify long-term seabed morphology in the complex, non-stationary environment of the Nakdong River Estuary and to support on-site decision-making, we built and validated an XGBoost-based bathymetry prediction model using grid-based depth time series compiled from regular surveys conducted between 2013 and 2022 (17 campaigns for Area A and 16 for Area B). The inputs were designed as a minimal, core feature set comprising the most recent depth, first- and second-order differences, seasonal-cycle encoding, neighborhood statistics (mean and variance), and terrain metrics (slope and curvature). To assess temporal generalization, we combined forward-chaining (past-to-future) block time-series cross-validation with spatial block validation. The evaluation showed that Area A (shallow, highly variable) achieved R² = 0.903, RMSE = 35.2 cm, MAE = 23.5 cm, MAPE ≈ 9.40%, nRMSE ≈ 0.141, and KGE ≈ 0.883 (reference depth 2.5 m), while Area B (deeper, more stable) achieved R² = 0.960, RMSE = 38.3 cm, MAE = 26.5 cm, MAPE ≈ 5.30%, nRMSE ≈ 0.077, and KGE ≈ 0.940 (reference depth 5.0 m), confirming long-term consistency and reproducibility. Predicted bathymetry maps for 2023∼2030 reproduce, in Area A, the spatial migration and amplitude associated with seasonality and sandbar reorganization —while tending to smooth abrupt short-term changes along steep slopes and bar boundaries—and, in Area B, maintain the continuity and stability of the central channel. Major limitations include (1) constraints on learning short-period events due to uneven survey intervals, (2) the exclusion of exogenous physical drivers such as rainfall, discharge, tides, and sediment supply, and (3) attenuation of sharp changes inherent to the split-and-averaging nature of tree boosting. Nevertheless, the model provides a practical baseline for channel maintenance, planning the timing and scale of dredging, early identification of hazardous zones, and resurvey prioritization; future integration of exogenous variables, observation/data fusion, and physics–data hybridization is expected to improve both predictive skill and interpretability.