A method is presented for eliminating near-surface scattered noise from seismic data. Starting from an appropriately chosen background model, a surface-consistent scattering model is determined using linearized elastodynamic inverse scattering theory. This scattering model does not necessarily equal the actual scatterer distribution, but it enables one to calculate, approximately, the near-surface scattered part of the data. The method honors at least some of the complexity of the near-surface scattering process and can be applied in cases where traditional methods, like wavenumber-frequency filtering techniques and methods for static corrections, are ineffective. From a number of tests on synthetic data, we conclude that the method is rather robust; its main sensitivity is because of errors in the determination of the background Rayleigh-wave velocity.