Earthquakes in sea areas pose a serious threat to the development of China's marine resources and the economic construction of coastal areas. A crucial prerequisite for conducting relevant seismicityresearch is to compile an earthquake catalog for China's seas and adjacent regions. The commonly used magnitude scale in China is surface wave magnitude(M_S), however moment magnitude(M_W) has beencommonly used in latest seismic activity models andrecommended as a recognized magnitude standard worldwide. Therefore, it is necessary to fit the conversion relationship between surface wave magnitude of China and moment magnitude. Taking China's seas and neighboring regions as the study area, we collected the earthquake records from 1988 to 2020 with surface wave magnitudes from the China Earthquake Networks Center(CENC) and moment magnitudes from the Global Centroid Moment Tensor(GCMT). Based on the dataset, the year, depth, longitude, latitude, and M_S of CENC were extracted as impact factors, and the corresponding M_W of GCMT were used as markers. Then aBP neural network was trained to establish a moment magnitude conversion model. Meanwhile, least square regression and orthogonal regression were also applied to buildtraditionallinear models for comparisons. The results show that mean absolute error and root mean square error of least square regression and orthogonal regression are about 40% higher than that of BP neural network. Furthermore, the residual error of our proposed method has smaller absolute value and more concentrated distribution.