The exploration of urban underground spaces is of great signiffcance to urban planning, geological disaster prevention, resource exploration and environmental monitoring. However, due to the existing of severe interferences, conventional seismic methods cannot adapt to the complex urban environment well. Since adopting the single-node data acquisition method and taking the seismic ambient noise as the signal, the microtremor horizontal-to-vertical spectral ratio (HVSR) method can effectively avoid the strong interference problems caused by the complex urban environment, which could obtain information such as S-wave velocity and thickness of underground formations by fitting the microtremor HVSR curve. Nevertheless, HVSR curve inversion is a multi-parameter curve fftting process. And conventional inversion methods can easily converge to the local minimum, which will directly affect the reliability of the inversion results. Thus, the authors propose a HVSR inversion method based on the multimodal forest optimization algorithm, which uses the efffcient clustering technique and locates the global optimum quickly. Tests on synthetic data show that the inversion results of the proposed method are consistent with the forward model. Both the adaption and stability to the abnormal layer velocity model are demonstrated. The results of the real ffeld data are also veriffed by the drilling information.