Using the strong-constraint Regional Ocean Modeling System Four-Dimensional Variational Data Assimilation system (ROMS-4DVAR), this study assimilates AVISO 0.25° × 0.25° gridded sea surface height, Odysst L3 multi-satellite merged 0.1° × 0.1° sea surface temperature, and temperature–salinity profiles from the UK Met Office EN4 dataset. The results show that the root-mean-square error (RMSE) and bias of sea surface height decrease from 0.10 m (0.03 m) to 0.03 m (0.01 m), and those of sea surface temperature from 0.70°C (0.12°C) to 0.32°C (0.03°C). The RMSE and bias of subsurface temperature and salinity also decline markedly over the South China Sea region. The assimilation effectively suppresses the spurious Kuroshio intrusion loop in the Luzon Strait: the RMSE of sea level anomaly drops from 7.5 cm to 3.7 cm and the correlation coefficient rises from 64.8% to 87.8%, approaching the quality of reanalysis products. Comparisons with independent BGC-Argo temperature–salinity profiles in the northern and central South China Sea reveal that the analysis field yields substantial improvements in the northern basin, primarily because the spurious Kuroshio loop is constrained. In the central basin, where the free-running model already performs reasonably well and is less affected by the loop, the analysis field nonetheless provides further gains and overall outperforms the CMEMS reanalysis. With respect to dynamical processes, the analysis field accurately reproduces the deepening and shoaling of the mixed layer and thermocline, whereas the reanalysis performs poorly in this regard. Although velocity observations are not assimilated, subsurface currents are markedly improved: for zonal velocity, the bias declines from −7.53 cm/s to −3.39 cm/s, the RMSE from 8.56 cm/s to 3.78 cm/s, and the correlation coefficient from −0.67 to 0.74; for meridional velocity, the bias declines from 3.92 cm/s to 2.09 cm/s, the RMSE from 5.51 cm/s to 3.89 cm/s, and the correlation coefficient from −0.15 to 0.48. Water mass properties are also brought closer to observations. These results demonstrate that ROMS-4DVAR can effectively correct large-scale circulation biases in the South China Sea and substantially improve the representation of mesoscale processes and their subsurface structure.