Building a spectral soil library in a soil routine analysis laboratory to determine soil organic carbon using compact near-infrared spectrophotometers: Performance of global and local models.

Compact near-infrared (NIR) spectrophotometers are low-cost instruments enabling rapid, cheap, non-destructive, and environmentally friendly soil organic carbon (SOC) determination. However, the steps necessary to implement these instruments in routine soil analysis laboratories still need to be studied. The main objectives of this study are to assess the performance of spectral libraries built in a commercial soil analysis laboratory using compact NIR spectrophotometers aiming for SOC determination, compare the accuracy of global and local models, and evaluate strategies to implement local calibration. More than 9000 soil samples were included in this study. The independent validation was performed using 1000 randomly selected samples. This procedure was performed three times. Different modeling strategies were tested: conventional partial least square regression PLSR, global-PLSR (G-PLSR), local-PLSR (L-PLSR) based on the selection of n nearest neighbors samples (L-PLSRn), and local L-PLSR based on imposing Mahalanobis distance limit to select samples (L-PLSRdist). NeoSpectra compact spectrophotometer shows the best performance (root mean square error of the prediction, RMSEP, and R of 9.2 g kg and 0.6, respectively, for the G-PLSR, and of 6.5 g kg and 0.7, for L-PLSRn, were found). Models based on the NanoNIR spectrophotometer presented slightly lower accuracy (RMSEP and R values of, respectively, 11.4 g kg and 0.3, for the G-PLSR models, and of 8.6 g kg 0.3 for L-PLSRn models). Models built using G-PLSR performed worse than models built using local modeling strategies. Improvement in SOC determination of up to 42 % over the G-PLSR is found when L-PLSRn is used. Furthermore, the feasibility of improving the performance of local models by imposing Mahalanobis limit distances to select the calibration samples is proven. The results demonstrate that compact NIR spectrophotometers can be a low-cost alternative to partially replace the wet oxidation method on SOC measurement in routine soil analysis laboratories. For this purpose, building its own spectral library is currently the best strategy.