We have measured and compared the impacts of hauler-based, clear-fell, forest harvesting on the performance of six prediction techniques representing quantitative soil-landscape modelling, geostatistical, and class-based approaches to the spatial prediction of soil properties. The prediction techniques included three class-based techniques (CB1-3), multi-linear regression (MLR), regression kriging (RK), and ordinary kriging (OK). Six soil properties (macroporosity, total C, Bray P, Bray K, Bray Mg, and pH) were measured at 208 sample points on a regular 16.7-m grid pattern within each of two sub-adjacent five-ha plots comprising mainly Ultisols. One plot was under first rotation mature trees (i.e. pre-harvested) and the other had been harvested and was under second rotation two year-old trees (post-harvested). A jack-knifing procedure was used for the validation of the prediction techniques in each plot. For each soil property, the techniques were compared in terms of their bias (mean error), precision (root-mean-square error), and relative prediction performance (mean and standard deviation of rank and goodness-of-prediction), both within and between plots. We found that most techniques gave less biased and less precise predictions for most soil properties after forest harvesting. Across all soil properties, the relative performance of some prediction techniques (CB1, RK, and OK) generally became poorer after forest harvesting whereas the relative performance of other techniques (CB2 and CB3) generally improved. On balance, the performance of MLR remained the same. The geostatistical OK technique was generally the best predictor of soil properties across both plots. However, the CB2 technique also performed reasonably well (especially after harvesting).