Abstract: Peak ground-motion acceleration (PGA) is a fundamental parameter in seismic design codes and in earthquake early warning systems. We collected the seismic data recorded by 40 strong-motion stations of the KiK-net seismic station array in Japan, and studied the statistic distribution of PGA amplification factor f_PGA. It was demonstrated that the f_PGA under a given seismic intensity input was basically log-normally distributed with its mean and standard deviation depending on the site conditions. Any individual site characteristic parameter, such as V_s30, V_s20 or soil thickness D, was poorly correlated with the statistic parameters, i.e., the mean and standard deviation, while a satisfactory correlation was obtained with respect to linear combinations of V_s30, V_s20 and D. By the regression of the data, the statistic parameters of f_PGA were calculated according to a linear combination of site characteristic parameters to build the probability density function of the log-normal distribution model of f_PGA. Following the f_PGA probability model, the ground surface PGA corrected by specific site characteristic parameters could be predicted under different probability levels, and testified by seismic data. The probabilistic predictions of PGA meet the demands of risk analysis in engineering practice, paving the way for site-corrected surface PGA prediction in earthquake early warning and fast seismic intensity estimation.