The awareness that noise exposure is critical for human health is growing around the globe, and land-use regressions (LURs) are becoming a popular tool for producing noise exposure maps. One important factor for noise emissions is road traffic. The propagation in this regard is determined by the spatial layout of road infrastructure and the surrounding environment, respectively. LURs use geostatistical models and allow to extrapolate microphone measurements. In this study, we investigated whether models are prone to sampling artifacts. We used yearly averaged Lden simulations, compliant to the European noise directive 2002/49/EG, as input for 2000 virtual field campaigns. We permuted different sampling schemes (random, systematic, stratified) and sizes (n = 50, 100, 200, 500 to 1000) 100 times. The overall model performances varied substantially between 0.61 – 0.95 for R², 1.94 – 7.46 dB(A) for mean absolute error and 2.47 – 10.03 dB(A) for root mean squared error. Comparing the eventual model terms using variance analyses (ANOVA), we found significant differences between the sampling schemes for traffic information and land cover (e.g. vegetated surfaces) features. Simultaneously, less than half of the LURs’ weights differed significantly depending on the sampling size. Thus, our experiments give an in-depth view on the mechanics of LUR and their sensitivity with respect to sampled training data