Estimation of the properties of heterogeneous soils on a multiregional scale in Italy using near infrared spectroscopy and LOCAL calibration procedures

In 2001, Pioneer Hi Bred Italia srl and Di.Pro.Ve. collected 420 samples from arable soils all over Northern Italy in order to produce, through the NIR technology, rapid and cheap soil analysis for corn growers. The analyzed soil variables were: sand (SA) silt (SI), clay (CL), total lime (TL), organic carbon (OC), total nitrogen (TN). Air dried samples were analyzed twice according to the following reference methods: pipette for textural classes, gas-volumetric for TL and dry combustion for OC and TN, and scanned by a 5000 FOSS-NIRSystems spectrometer in the 1100-2500 nm range; spectra were collected on both 2 mm sieved samples and 0.5 mm additionally ground samples. The ranges of variables and the obtained standard deviations of reference methods (sd) were (g/kg): 61-929 (sd=21.4) for SA, 32-543 (sd=26.6) for SI, 31-680 (sd=11.3) for CL, 0-868 (sd=8.6) for TL, 3.6-114.5 (sd=0.64) for OC, 0.52-8.53 (sd=0.068) for TN. Because both the great heterogeneity of soil types and the unhomogeneous constitution of each sample, the best results were obtained using a LOCAL calibration procedure on 0.5mm samples spectra. A third of samples were randomly selected for constituting an independent validation set. Standard errors of prediction (SEPs) and validation r2 were: 74.7 (r2=0.793) for SA, 31.6 (r2=0.919) for CL, 3.1 (r2=0.945) for TL, 3.8 (r2=0.883) for OC and 0.29 (r2=0.884) for TN. No correlation was found for SI. In 2002 these calibrations were used for predicting a set of 7200 soils collected by Pioneer in italian corn areas: all samples with a distance of Mahalanobis above 3 from mean spectrum (about 10%) and 200 random samples were analyzed by wet chemistry and added to calibration set. Obtained final sets amounted to 1228, 1403 and 1389 samples respectively for SA and CL, TL, OC and TN. More robust and reliable LOCAL calibrations were calculated on these spectral databases and tested on 200 new validation samples; in order to provide a highest level of accuracy for OC and TN, samples were split in two concentration subgroups by a discriminant equation and validation samples were predicted with the appropriate subgroup database. SEPs obtained were: 68.9 for SA (r2=0.80), 36.3 for CL (r2=0.87), 0.94 for TL (r2=0.98), 4.05 for OC (r2=0.95) and 0.32 for TN (r2=0.93). Both the low number of outliers, less than 2%, altough analytical ranges were wider than in 2001, and the relatively low SEP values emphasize the usefulness of NIR estimation for routine, cheap analyses.