Spiking of NIR regional models using samples from target sites: Effect of model size on prediction accuracy

Near infrared reflectance (NIR) spectroscopy has many advantages because it is a rapid and cost-effective technique. A needed steep is the development of soil spectral libraries and models (calibrations using multivariate techniques). The calibrations should contain the variability of the target site soils on which the models are to be used. Many times this premise is not easy to fulfil. A classical way to solve this problem is by the spiking of models with a few samples from the target site (local samples), and the subsequent recalibration of models. In this regional-scale study we hypothesized that small-sized models should integrate the spectral characteristics of local samples more easily than large-sized models and as consequence should produce more accurate predictions. With the aim to test this hypothesis different-sized models were constructed using different quantities of soil samples. Partial least squares (PLS) regression was used to construct the models which were relating NIR spectra to the Nitrogen Kjeldahl (NKj) contents in soil samples. Two strategies were used for the selection of samples in models: (i) strategy based on spectra characteristics (SC) and (ii) strategy based on NKj values (NV) of the samples. These different-sized models were sequentially spiked with local samples from target sites and recalibrated. The predictions accuracies obtained with the models before and after spiking were studied being the NKj the soil property selected. In general all predictions were very accurate including those obtained before the spiking of models. Predictions accuracy increased as consequence of spiking in three of the four target sites studied. A negative trend was observed between prediction accuracy and model size. The lower errors were obtained using small-sized models after spiking which were more accurate than local models too. It was noticeable the high accuracy obtained by local models which were constructed using only 20 local samples. Before spiking SC models were more accurate than NV models but scarce differences between both strategies were observed after spiking with 20 samples. The results suggested that small-sized models can be useful for local predictions after spiking and they were also emphasizing the relevant role of local samples in models. The results obtained could encourage the expansion of this technique because large data based seem not be needed. Therefore NIR users could primarily focus most of their efforts on obtaining highly accurate analytical values in a few set of samples. (C) 2010 Elsevier B.V. All rights reserved.