Ground roll noise on land data sets overwhelms the desired reflection seismic signal unless special steps are taken in data acquisition and processing to control it. This is usually done in the field by the design of group arrays for data acquisition. On the other hand, if multicomponent data are acquired, it is possible to remove ground roll during processing using polarization analysis. Even though this processing is computation-intensive, the potential exists for obtaining results similar to conventional data acquisition, but with deployment of fewer sensors in the field with minimal group array effects. It also has potential for deriving new information.We describe a two-dimensional polarization-filter analysis for use with vertical and in-line sensors. A time-domain spectral matrix technique is developed since the recorded seismic signal is the superposition of multiple signals in the time domain, each with different frequency content and time-varying polarization. This technique is implemented by decomposing the signal into individual frequency components using narrow band-pass filters and defining the polarization characteristic using sliding time windows. We show that both incoherent noise and specific linearly polarized constituents can be successfully filtered.