Near-infrared spectroscopy for within-field soil characterization: small local calibrations compared with national libraries spiked with local samples

The advantage of using near‐infrared spectroscopy to increase sample point density in soil mapping on farms relies on the number of conventional laboratory analyses for the calibrations being kept to a minimum. This study compared the performance of small farm‐scale calibrations (25 samples) with a larger national soil library (396 samples) and tested whether a site‐specific sample set selected from the national library, consisting of the 50 samples that were spectrally most similar to those of the local sites, could increase performance. In addition, the national library and selected subsets were augmented (‘spiked’) with up to 25 local calibration samples to test whether that had any additional effect on prediction errors and bias. Calibrations were made for predicting within‐field variation in clay, silt, sand, soil organic carbon (SOC), pH and phosphorus. Selecting a subset of samples from the national library did not improve the results compared with using the entire national library. However, spiking both libraries with local samples reduced the root mean squared error of prediction (RMSEP) considerably, mainly through a decrease in bias, and often resulted in comparable results to the local calibrations. There was a tendency for better clay and SOC predictions from spiking a reduced national library compared with spiking the entire national library, sometimes even resulting in better predictions than using the local calibrations. However, using local calibrations seems to be the best alternative for predicting soil properties at the farm or field scale, even with as few as 25 samples.