The ProbeField approach for estimating SOC by field vis-NIR spectra corrected for moisture effects

The vis-NIR spectra acquired in the field are not collected under controlled conditions as in the laboratory, and they are affected by some disturbing factors such as variable light conditions, surface roughness and soil moisture. Consequently, the soil properties estimation models, calibrated using soil spectral libraries (SSL) consisting of dry and sieved samples, most of the time fail if applied to field spectra. In the framework of EJP SOIL ProbeField project, we collected field soil spectra in 85 locations (5 countries) and for each sampling location, the spectra were acquired directly on the ground surface (NO), and then after flattening it (FL). All the spectra were harmonized according to an internal soil standard (ISS). LUCAS topsoil, GEO-CRADLE and ProbeField SSLs were used to train machine learning for soil organic carbon (SOC) prediction, these models were applied to all the field spectra (NO and FL) both applying and not applying the external parameter orthogonalization (EPO) approach to remove the effect of soil moisture. <br/><br/>We observed a significant increase in SOC prediction accuracy using EPO for all the SSLs, and the best statistics were obtained using the GEO-CRADLE SSL for the FL spectra (RMSE=1.09%; RPD=1.96; R2=0.74). However, in general, the comparison between NO and FL in terms of accuracy did not show significant differences. The two large SSLs, GEOCRADLE and LUCAS, provided the best performances as compared to the ProbeField SSL. This is probably due to the higher soil types and SOC variability of the larger SSLs. These results clearly indicate the need to remove soil moisture effect from field spectra to obtain reasonable accurate estimates of SOC content exploiting large SSLs, and how the EPO can be successfully used for this purpose.