Aiming at the problems that the frequency band division is not accurate due to the noise effect on the Fourier spectrum when the empirical wavelet transform (EWT) has been used for identifying the modal parameters of noisy signals, an improved empirical wavelet transform (IEWT) based procedure is proposed in this study to identify the modal parameters of civil engineering structures under ambient excitations. In the proposed IEWT-based procedure, the autoregressive (AR) power spectrum based on the Burg algorithm has been adopted to substitute the Fourier spectrum for the frequency band division, the random decrement technique (RDT) and the Hilbert transform (HT) based mono-component modal parameter identification method are all included. A typical simulated signal, the ASCE Benchmark structure under the white noise excitations and the Ting Kau Bridge under the typhoon excitations have been used to validate the correctness, effectiveness and applicability of the proposed procedure. The identified results show that (1) the proposed AR power spectrum based frequency band division technique can accurately estimate the frequency band boundaries and isolate noise, (2) the proposed IEWT-based identification procedure can effectively identify the low-order natural frequencies and damping ratios of the real civil engineering structures under ambient excitations, (3) the identified results are more accurate than the wavelet transform (WT) based method.