Crosswell and VSP seismic survey data are frequently contaminated by coherent noise trains in the form of direct tube waves and tube-wave-to-body-wave conversions. The problem was especially acute in a recent high resolution underground seismic experiment that was conducted in a hard rock environment. Conventional processing, such as fk filtering, was ineffective in suppressing such noise events.A procedure, based on singular value decomposition (SVD) of a seismic section, was employed for cancelling coherent tube-wave related noise. An eigen-analysis splits the section into linearly independent eigenimages. Each eigenimage is a characteristic content of the seismic record. Its energy contribution to the total energy of the seismic section is represented by the magnitude of its corresponding eigenvalue. The leading eigenimages, obtained after windowing and flattening, represent high amplitude and correlatable energies such as tube wave noise. The trailing eigenimages represent low amplitude and least correlatable energy such as random background noise. Useful records can be extracted by subtracting those eigenimages dominated by the noise from the total seismic gather.The optimal removal of the tube waves requires the amplitude balancing of traces and the precise alignment of the tube wave event across the traces before eigen-analysis. The procedure is illustrated through application to reversed VSP data acquired in a metalliferous mine. Superior results are obtained in comparison to that of a conventional f-k filter.