Multidimensional deconvolution (MDD) by point-spread-function removes the blurring effects and the spread distortions typically generated in the signal representation by seismic interferometry (SI). Under suitable conditions, the MDD inversion of SI signals reconstructed by the Kirchhoff-Helmholtz integral of crosscorrelation type is a valuable and robust technique to recover the Green’s function of the subsurface. A basic requirement for the effective MDD application to SI data is to know the separated incoming and outcoming wavefields at the receivers illuminated by the real sources. We extended the MDD concept to the virtual reflector (VR) approach by crossconvolution and compared equivalent results obtained with approximations for the propagated wavefields. Examples were discussed with Arctic data of an on-ice shallow-water seismic experiment affected by strong and dispersive flexural ice wave noise. The target was to improve the signal-to-noise ratio by redatuming the sources at the sea bottom. First, we processed the raw input signals recorded by sea-bottom receivers to obtain an approximation of the incoming wavefield from on-ice sources. Then, we processed the data of the whole 2D seismic line in two different ways: applying MDD by crosscorrelation and by crossconvolution to, backward, SI and, forward, VR interferometric results, respectively. The analysis of the signals redatumed to the sea bottom showed that the flexural ice noise was significantly mitigated with respect to the conventional interferometry approach, with improved reflections for seismic investigation purposes. The agreement of the phase in the SI and VR results after MDD inversion confirmed that the approach was robust and enabled us to enhance the signals combining the stacked sections obtained by SI MDD and VR MDD. The resulting MDD SI stacked sections showed an improvement in signal quality especially at low frequencies with respect to the stacked section obtained by conventional processing of the original data.