Fast and accurate magnitude estimation is essential for earthquake early warning system(EEWs). The magnitude estimation methods based on a single characteristic parameter of the initial wave is widely used in EEWs. However, the empirical formula of magnitude established by a single characteristic parameter cannot sufficiently utilize information related to the magnitude in the first arrival wave, which greatly limits the effectiveness of magnitude estimation. this study has proposed a new approach (GPR-M) based on Gaussian process regression to estimate magnitude that includes 10 characteristic parameters in the time domain, frequency domain, and time-frequency domain of initial wave. GPR-M is trained and tested using a large number of surface strong earthquake records in Japan, and compared with the maximum predominant period τpmax
method and the peak displacement Pd
method. The results show that the accuracy of GPR-M is significantly better than τpmax
in estimating magnitude with and without hypocentral distance in both cases. In addition, the generalization ability test of GPR-M trained by Japanese data using Chile's surface strong motion records shows that GPR-M has better generalization ability than τpmax
. GPR-M can effectively improve the accuracy of magnitude estimation for EEWs without being affected by regional differences..