Due to the highly integrated configuration of underwater electro-hydraulic actuators (EHAs), their modal frequencies are densely distributed, making the dominant modes difficult to identify accurately using traditional approaches. To address this issue, a modal identification method integrating weighted frequency-domain subspace identification with parameter optimization was proposed. First, a frequency-domain subspace model was developed based on the frequency response functions (FRFs). To alleviate inconsistencies among different measurement locations, the FRF data were preprocessed using the H1 estimator combined with a frequency-domain weighting strategy. Subsequently, nonlinear parameter optimization was introduced to further improve identification stability. Experimental modal analyses were conducted on an underwater EHA prototype. The optimized model achieved an average fitting accuracy of 89.30% relative to the measured data. Comparisons of the identified modes under oil-filled and oil-free conditions showed that internal oil filling shifted the low-order natural frequencies of the underwater EHA to lower values, while the average damping ratio increased by 28.24%. Both land-based and underwater vibration experiments revealed that the peak frequencies in the vibration spectra generally corresponded to the identified natural frequencies. Among the extracted modes, the third- to sixth-order modes made the most significant contribution to the vibration response. In addition, the underwater environment exhibited a vibration suppression effect. Overall, the proposed approach effectively extracts dominant modes from complex underwater EHA structures and provides useful support for subsequent dynamic analysis and vibration suppression design.