Comparative Evaluation of Tree-Based Ensemble Learning for Correcting Residuals in SWOT Altimetry-Derived Marine Gravity Anomalies

The quality of the SWOT-derived marine gravity field and its subsequent refinement are investigated in the Western Pacific (120–140°E, 0–30°N). Inverting 41 cycles of the SWOT L3_LR_SSH v2.0.1 product yields a gravity field with RMSEs of 2.476 and 4.321 mGal against independent JAMSTEC and NCEI shipborne gravity, respectively, and the accuracy gain converges as the number of SWOT altimetry repeat cycles increases, showing diminishing marginal returns. To further refine this field, three tree-based ensemble learning methods—Random Forest (RF), histogram-based gradient boosting (HGB) and LightGBM (LGBM)—are trained on 67 multi-source features with the SWOT-minus-shipborne residual as the target, and validated under two spatially isolated strategies: a 2° × 2° spatial block hold-out with a 10 km buffer and a cruise-wise hold-out over seven complete cruises, in which training samples within 10 km of any test point are removed. All three methods reduce the error. On the block hold-out, LGBM performs best (RMSE 2.398 mGal, a 3.6% improvement over the 2.488 mGal baseline), followed by HGB (2.402 mGal, 3.5%) and RF (2.422 mGal, 2.7%); on the cruise hold-out the gains are markedly larger, with HGB best (RMSE 1.703 mGal, a 8.6% improvement over the 1.863 mGal baseline), then LGBM (1.714 mGal, 8.0%) and RF (1.794 mGal, 3.7%). A consistent reversal is found between the two assessment levels: RF leads on the in-training out-of-fold RMSE but is weakest on unseen cruises, indicating that RF is competitive within the training distribution yet weaker than boosting methods when extrapolating to unsurveyed areas. Power spectral density analysis shows that the improvements concentrate at wavelengths of 8–45 km, whereas wavenumber below 0.0221 km−1 are slightly degraded.