Sustainable agriculture practices are often hampered by the prohibitive costs associated with the generation of fine-resolution soil maps. Recently, several papers have been published highlighting how visible and near infrared (vis–NIR) reflectance spectroscopy may offer an alternative to address this problem by increasing the density of soil sampling and by reducing the number of conventional laboratory analyses needed. However, for farm-scale soil mapping, previous studies rarely focused on sample optimization for the calibration of vis–NIR models or on robust modelling of the spatial variation of soil properties predicted by vis–NIR spectroscopy. In the present study, we used soil vis–NIR spectroscopy models optimized in terms of both number of calibration samples and accuracy for high-resolution robust farm-scale soil mapping and addressed some of the most common pitfalls identified in previous research. We collected 910 samples from 458 locations at two depths (A, 0–0.20 m; B, 0.80–1.0 m) in the state of São Paulo, Brazil. All soil samples were analysed by conventional methods and scanned in the vis–NIR spectral range. With the vis–NIR spectra only, we inferred statistically the optimal set size and the best samples with which to calibrate vis–NIR models. The calibrated vis–NIR models were validated and used to predict soil properties for the rest of the samples. The prediction error of the spectroscopic model was propagated through the spatial analysis, in which robust block kriging was used to predict particle-size fractions and exchangeable calcium content for each depth. The results indicated that statistical selection of the calibration samples based on vis–NIR spectra considerably decreased the need for conventional chemical analysis for a given level of mapping accuracy. The methods tested in this research were developed and implemented using open-source software. All codes and data are provided for reproducible research purposes. Highlights: Vis–NIR spectroscopy enables an increase in sampling density with little additional cost. Guided selection of vis–NIR calibration samples reduced the need for conventional soil analysis. Error of spectroscopic model prediction was propagated by spatial analysis. Maps from the vis–NIR augmented dataset were almost as accurate as those from conventional soil analysis.