3.4.1 Soil pH prediction
The soil pH was most accurately predicted by MARS
(R2c=0.90; MBE=-0.003;
RMSEc=0.34; RPDc=2.98) and SVR
(R2p=0.66; MBEp=0.04;
RMSEp=0.51; RPDp=1.62) during
calibration and validation, respectively (Figure 4a). The calibration
accuracy of the SVR (R2c=0.89;
MBEc=-0.008; RMSEc=0.35;
RPDc=2.89) was comparable to that of MARS. The
validation prediction accuracy of SVR was classified as acceptable. The
overall rank based on the model evaluation parameters indicated SVR
(1.5) as the best performing model to predict soil pH. Though the
calibration ranking of MARS was 1, the rank during validation was 4.25
indicated inefficiency of the MARS model. The cross-validation accuracy
to predict soil pH using the PLSR of reflectance in mid-infrared (MIR)
and combined VIS-NIR-MIR was recorded by Viscarra Rossel et al. ,
(2006) as R2adj 0.75 and 0.33,
respectively. They recorded the best prediction using MIR and observed
RMSE of 0.10 unit. The prediction accuracies reported in the present
study for pH were less accurate than literatures as R2of 0.74 using PLSR (Reeves & McCarty, 2001), 0.73 using PLSR (Reeveset al. , 1999), 0.70 using MARS (Shepherd & Walsh, 2002), 0.70
using PCR (Islam et al. , 2003), 0.56 using PCR (Sun et
al. , 2003). In the present study, the poor predictions for the soil pH
might be attributable to the lower variability in the full (CV=16.75%),
calibration (CV=18.46%) and validation (CV=15.44%) dataset (Table 1).
Most of the samples of the study had acidic soil reaction.