The applicability of surface nuclear magnetic resonance (NMR) in investigations of groundwater is often limited by high noise levels in many areas of interest. We have evaluated measurements from a high-noise-level area in Ristrup, Denmark, recorded with a Numis Poly instrument. Standard multichannel filtering techniques for noise reduction are inadequate for several data sets acquired in this area, and surface-NMR signals cannot be resolved from the acquired data. Based on a careful assessment of the frequency content of the data, we have determined how a model-based noise reduction approach can be generalized and used to subtract two harmonic noise components from the data. Reliable surface-NMR data can be extracted from the noise-reduced data. Moreover, we have determined the impact of the proposed processing approaches on our inversion results, and we have also developed an example in which the proposed methodology allowed us to reveal and avoid an otherwise overlooked contamination of the reference coil signals with a surface-NMR signal. No borehole data were available for the investigated sites, and the validity of the noise reduction approach was instead verified using a synthetic five-layer model embedded in noise-only records from Ristrup. Our results have determined that a careful processing of the recorded data made it possible to extract surface-NMR data in more places of interest.