The repetitive observations of satellites provide rich multi-temporal information for coastal remote sensing, making it possible to improve the accuracy of bathymetric inversion through multi-temporal satellite data. This study takes Culebra, Puerto Rico, as the study area and attempts multi-temporal bathymetric inversion using 193 Sentinel-2 images and eight tracks of ICESat-2 ATL03 data. Two widely used machine-learning models, CatBoost and Random Forest (RF), were employed to construct bathymetric inversion models, and the Fusion followed by Inversion (FI) strategy and inversion followed by Fusion (IF) strategy were also compared. The results show that the R2 of inversion based on multi-temporal observations exceeds 97.47%, with RMSE lower than 1.00 m, and MAE lower than 0.54 m, making the results more accurate than most single-phase results. The FI strategy yields better results than the IF strategy, with an RMSE of 0.81 m. Both CatBoost and RF models exhibit comparable robustness under the tested scenarios, with CatBoost showing minor advantages in specific cases, achieving an average RMSE of 0.88 m. Furthermore, multi-temporal observations effectively mitigate environmental interference, such as clouds and waves, enhancing the reliability of bathymetric inversion. The findings highlight the potential of combining the FI strategy with advanced machine-learning models to achieve more reliable bathymetric inversion results.