Polarization filtering methods based on a covariance matrix play an important role in the processing of multicomponent seismograms due to their explicit physical meaning, ease of implementation, and high efficiency. Conventional polarization filtering methods that are realized in a time domain have major limitations in resolving seismic signals in which waveforms or frequencies overlap. Time-frequency analysis methods are especially suitable for resolving separate seismic signals that overlap in time but have different spectra for instantaneous signal analysis. These methods can describe frequency components of a signal that change over time. Owing to the advantages of the time-frequency analysis method, it can be used in polarization analysis. This study presents a polarization filtering method based on the generalized S-transform to suppress surface waves in a time-frequency domain. On one hand, we remold the window function of the S-transform and improve the frequency resolution of seismic signals by increasing regulatory factors to create a nonlinear change in the window function with the signal frequency. On the other, we structure the cross-energy matrix in the time-frequency domain using the generalized S-transform, compute instantaneous polarization attributes by eigenanalysis, and design a filtering algorithm in the time-frequency domain to achieve polarization filtering of multicomponent seismic signals. The specialties of this method are that the length of the time window of the covariance matrix is determined by the instantaneous frequency of the multicomponent seismic data and it can adapt to the dominant period of the desired signal. Moreover, it calculates polarization parameters at each time-frequency point and no longer needs to perform interpolation. It is particularly accurate in processing signals with overlapping waveforms or frequencies in the time or frequency domain. The results of processing data from models and real three-component seismograms show that this method has very high clarity, high resolution, and practicability in the data analysis and processing of seismograms. This representation enables the detection of dispersion in polarization attributes, which can be further exploited to infer some physical characteristics of the medium under investigation. Moreover, this representation offers the ability to distinguish between attributes that belong to different coherent events that may overlap in time but with different frequency contents separated by time-dependent frequency cutoffs. Identifying and separating different wave types are made possible by designing filters that operate in the time-frequency domain. Attributes such as azimuth, dip, and signed ellipticity can also be used to improve the filtering algorithms.