Soil Spectral Libraries facilitate agricultural production taking into account the principles of a low-input sustainable
agriculture and provide more valuable knowledge to environmental policy makers, enabling improved decision making
and effective management of natural resources in the region. In this paper, a comparison in the predictive performance of
two state of the art algorithms, one linear (Partial Least Squares Regression) and one non-linear (Cubist), employed in
soil spectroscopy is conducted. The comparison was carried out in a regional Soil Spectral Library developed in the
Eastern Macedonia and Thrace region of Northern Greece, comprised of roughly 450 Entisol soil samples from soil
horizons A (0-30 cm) and B (30-60 cm). The soil spectra were acquired in the visible – Near Infrared Red region (vis-
NIR, 350nm-2500nm) using a standard protocol in the laboratory. Three soil properties, which are essential for
agriculture, were analyzed and taken into account for the comparison. These were the Organic Matter, the Clay content
and the concentration of nitrate-N. Additionally, three different spectral pre-processing techniques were utilized, namely
the continuum removal, the absorbance transformation, and the first derivative. Following the removal of outliers using
the Mahalanobis distance in the first 5 principal components of the spectra (accounting for ~99.8% of the variance), a
five-fold cross-validation experiment was considered for all 12 datasets. Statistical comparisons were conducted on the
results, which indicate that the Cubist algorithm outperforms PLSR, while the most informative transformation is the
first derivative.
|