Abstract: The engineering characteristics of marine site are the important basis for the design, construction, operation and maintenance of major marine projects and the research focus of geotechnical earthquake engineering. The strong ground motion records of 6 seabed stations in Sagami Bay, Japan from 2000 to 2018 are collected, after screening and pre-processing, the horizontal-to-vertical spectral ratio (HVSR) method is used to calculate the Fourier amplitude spectrum ratio of each record. The optimal site characteristics of Sagami Bay are obtained through the dynamic clustering learning combined with the optimal classification. The results show that the dynamic clustering learning method has strong performance and robustness for time series data, and can meet the local and global resolutions. The site of Sagami Bay can be macroscopically divided into submarine canyon area and basin area. The dominant period of the canyon area is significantly smaller than that of the basin area, and the high frequency attenuation is relatively slow. The engineering characteristics of marine site are given by the topographic slope of the local site. The relevant research results may provide reference for simulation of sea ground motion and construction of marine projects.