This paper describes a technique for noise reduction in synthetic aperture radar interferometry. The noisy interferogram is decomposed using undecimated wavelet transform and the coefficients are weighted. A novel method for computing the weights for each subband, based on an estimate of the relative noise content in them, is presented with a median filter used as the noise estimator. The proposed technique is not optimised for any specific signal or noise models. Results show that this technique provides an improvement of around 15% over the conventional boxcar filter in terms of estimated height error of a digital elevation model constructed from the filtered interferogram.