3.2 Spectral preprocessing influence
Spectra obtained from Vis-NIR and MIR spectrometers were subject to
preprocessing, such as absorbance, first order derivative, second order
derivative, multiple scatter correction and standard normal variate.
Spectra pre-treatment is a mathematical manipulation that enhances the
spectral information and eliminates the physical effect of light
scattering, which can be due to particles of different sizes and shapes
of samples (Minasny and McBratney, 2008) and is thus the most important
step before any chemometric modeling. Different pre-processing
transformations have been applied in numerous studies to transform soil
spectral data, remove noise, accentuate features, and prepare them for
chemometric modelling. However, the first derivative, second derivative,
SNV and MSC manipulation did not greatly enhances some of the spectral
features compared to reflectance. Moreover reflectance (unprocessed
spectra) presented the best performance as compared to other
preprocessing methods, irrespective of the models used (PLSR, RF, SVR or
MARS) (Table 2 & 3) and was thus considered to be the most robust
spectral preprocessing method based on its predictive performance for
EC. Some earlier results (Moros et al. 2009) also suggest that
calibration models in which spectra were not preprocessed are more
sensitive to changes compared to models for which preprocessing was
applied and Nawar et al. (2016) re-confirmed it, and used no
preprocessing for prediction. Reflectance has also been successfully
used in other studies, to estimate soil properties (Viscarra Rossel et
al. 2006, Nawar et al. 2016). Vibhute et al., (2018) reported
electrical conductivity to be better calibrated
(R2 = 0.80 and RMSE = 2.07) before pre-treatments than
after pretreatment of spectra and Nocita et al. (2014) applied
continuous removal reflectance to predict the soil properties by diffuse
reflectance spectroscopy from soil samples throughout the European
Union. The present study demonstrates that reflectance (unprocessed
spectra) (Fig 3 a & b) is better than any preprocessing tool for
prediction of EC regardless of the method applied and demonstrates its
suitability for prediction of EC, both in the Vis-NIR and MIR spectral
regions.