Abstract: Ocean sound speed profiles are key environmental parameters for bathymetry. To address the high cost and limited spatiotemporal coverage of in situ sound speed observations, this study proposes an ocean sound speed profile reconstruction method based on Empirical Orthogonal Function (EOF) decomposition and a deep learning Transformer model. Sea surface sound speed, sea surface temperature, sea surface salinity, geographic coordinates (latitude and longitude), and date are taken as input variables. A model specifically designed for sound speed profile reconstruction is employed to estimate the EOF coefficients, which are subsequently combined with the historical mean profile to reconstruct the full-depth sound speed profile. Experiments using Argo observational data indicate that, in both the Northwest Pacific and South Atlantic, the proposed method outperforms mean profiles, WOA23, XGBoost, BP-NN and MLP model reconstructions in terms of maximum sound speed difference, mean error, and root mean square error. The method is capable of stably capturing complex sound speed variations and achieving sound speed profile reconstruction. This approach provides an effective technical solution for regional sound speed modeling and sound speed acquisition in bathymetry under conditions lacking in situ observations.