Detection of microseismic events from continuously recorded data is important for reservoir characterization. Using spectral matrix analysis of 3D particle motion in the time-frequency domain, we can continuously estimate the polarization of waveforms. Polarization analysis provides linearity and ellipticity of the particle motions. The time variation of polarization features is subject to a statistical test to detect P and S-wave arrivals. Polarization features are evaluated in the frequency band having highest power, which can reduce the effect of noise, allowing us to detect low signal to noise ratio events. As this method is based on the detection of coherent polarized waves, the detectability does not depend on the similarity of waveforms. We test our method on field microseismic data, showing the feasibility to detect buried events.