Potentials and limitations of global versus local calibration models for soil organic carbon on large scale

The application of NIR in soil science is a rapidly growing field. However, soil NIR is with particular challenges due to the often relatively small concentration of the target substance (such as soil organic matter SOC) and the enormously divers matrix that is derived from very different minerals in different particle size classes. For large scale application of NIR to predict soil parameters there are two strategies to deal with this challenges: i) global calibrations that include the diversity of matrices in one model, ii) local calibrations that divide the global data set in certain classes according to predefined parameters such as soil particle size distribution. This work contributes to the further development of soil organic carbon (SOC) determination using near infrared spectroscopy (NIRS) and chemometrics analysis. The main objective of this study was to investigate the potential and limitations of a global versus a local calibration model for SOC on a regional scale for German soils. A total of 5361 soil samples (0-100 cm soil depth) were collected at 1011 sampling locations in grasslands and croplands from all over Germany covering a range of clay content from 10.2 to 71.3 % and a range of SOC content from
0 - 568.3 g C kg-1. They were scanned with a FT-NIR spectrometer in the laboratory. Partial least squares (PLS) regression is used to construct the SOC calibration models. The selection criteria for the classes for the local calibration were (i) soil particle size distribution (silt, sand, clay, and loam) and (ii) levels of SOC content. For global calibration all samples were included and resulted in a RMSE of 8.0 g C kg-1 and R2 of 0.93 and an RPD of 3.9. However, the results of all local calibrations achieved higher accuracies. For local calibrations with the particle size distributions as classification criteria four classes were calibrated separately: silt, sand, clay and loam (combined RMSE 5.3 g C kg-1), in two classes: loam/silt/clay and sand (RMSE 5.6 g C kg-1). Local calibration with soil carbon as classification criteria derived a combined RMSE of 3.4 g C kg-1 if the sample set was split into four classes (0-30, 31-70, 71-180, >181 g C kg-1). However if the sample set was split into five classes the combined RMSE of 4.1 g C kg-1 increased again likely due to sample numbers (4888 samples) dropping below 800 samples in a certain class. In general, the approach to obtain lower errors using local calibration could be interpreted as spectral similarity. Local calibrations were able to reduce the calibrations error as compared to a global calibration indicating this strategy is successful in coping with the large matrix variability of soils. Future developments shall allow classification of the sample not based on measured soil properties but on the spectral properties itself.