3.4.9 Soil Ca prediction
Similar to the results of Na prediction, the MARS model had the highest
prediction accuracy for Ca using the calibration dataset
(R2c=0.81; MBEc=-0.09;
RMSEc=1.60; RPDc=1.78). It was then
followed by SVR (R2c=0.77;
MBEc=-0.06; RMSEc=1.49;
RPDc=1.91) which was noticeable here that the rank for
calibration using SVR (1.25) was better than MARS (1.75) and their
corresponding predictions were considered as excellent. The PLSR
provided the highest accuracy using the validation dataset (Figure 4i)
with a prediction accuracy of
R2p=0.43; MBEp=-0.11;
RMSEp=2.27; RPDp=1.20 and rank=2.63.
But, this validation prediction was non-reliable. Malmir et al.(2019) also reported excellent prediction using PLSR model of
reflectance data (400-1000 nm) with
R2cv=0.81 and
RMSEcv=260.79 (for grounded sample) and
R2cv=0.81 and
RMSEcv=260.97 (for sieved sample). With the use of the
airborne hyperspectral imaging (350-2400 nm) and VIS-NIR spectroscopy
(400-2500nm), the soil Ca concentration could be predicted with
acceptable R2=0.69-0.80 (DemattĂȘ et al. , 2016;
Hively et al. , 2011). Janik et al. (1998) and Cozzolino &
Moron (2003) recorded a very good Ca prediction accuracy
(R2=0.89 and 0.90, respectively) using PLSR modeling
of VIS-NIR and MIR spectral data, respectively.