We analyze the acceleration records of a vertical array to study the linear and nonlinear behavior of the soft soil layers at LSST site, lotung, Taiwan. This array includes five triaxial accelerometers deployed at depths of 0, 6, 11, 17and 47m. During a 6-year operation from 1985 to 1990, 30 earthquakes (4.0 ≤ ML ≤ 6.5) triggered this array. The maximum PGA value is 258 gals recorded at the surface station. Spectral analyses show that the strong motion (PGA > 150 gals) causes the peaks of ratios to shift to lower predominant frequencies. The averaged spectral ratios of 15 well-recorded weak motion data (PGA < 100 gals) are selected as a reference; the shift of the maximum predominant frequency can reach 20%. Compared with the weak motions, the strong motions also decrease the amplification factor. The maximum reduction of the amplification can reach 50%. The results of waveform simulation show that the linear model based on the Haskell method can well predict the motions at various depths. However, this linear model does not work for the strong motion data. Here a nonlinear numerical scheme, such as DESRA-2, is required and can significantly improve the simulation results although the PGA value at the surface station is still underestimated. Overall, the nonlinear numerical calculation is feasible to predict the strong motions for a horizontal layered structure.