In hyperspectral remote sensing, spectra are increasingly analysed using methods developed for laboratory studies, such as derivative analysis. These techniques require smooth reflectance spectra. Therefore, there is a need for smoothing algorithms that fulfil the requirement of preserving local spectral features while simultaneously removing noise. Noise occurs in variable intensity and over different band widths. Several methods for smoothing a signal exist, including the widely used median and mean filters, the Savitzky–Golay filter applied to laboratory spectra, the cubic spline, and the recently developed transform-based thresholding using the wavelet transform. We compare all these methods using reflectance spectra of the canopy of salt marsh vegetation. The best trade-off between noise reduction and the preservation of spectral features was found to be the wavelet transform, specifically using a translation invariant de-noising based on the non-decimated or stationary wavelet transform.