3.4.2 Salinity (EC) prediction
The soil salinity is usually measured in terms of
EC1:2.5 and EC of saturation paste extract
(ECe). In the present investigation,
EC1:2.5 has been considered as a measure of soil
salinity. The results of the prediction of EC1:2.5 were
similar to that of soil pH1:2.5. The MARS model
performed the best for calibration
(R2c=0.95, MBEc=-0.05,
RMSEc=1.45 and RPDc = 3.68) and it was
classified as excellent which was followed by the SVR
(R2c=0.92; MBEc=-0.03;
RMSEc=1.92; RPDc=2.76). The prediction
accuracy of the SVR for validation was found excellent
(R2p=0.80; MBEp=0.30;
RMSEp=2.64; RPDp=2.06) (Figure 4b). The
SVR had the best validation and an overall rank of 1.00 and 1.38,
respectively. This was followed by MARS which had a prediction accuracy
of R2p=0.66;
MBEp=0.86; RMSEp=3.88;
RPDp=1.40). In the present study, non-linear models
predicted the soil salinity (EC1:2.5) better than the
linear models. These results are in line with the findings of Bilgiliet al. , (2011), Farifteh et al. , (2007), Nawar et
al. , (2014), Sidike et al. , (2014). Nawar et al. , (2014)
reported that the non-linear multivariate technique (MARS)
(R2=0.73; RMSE=6.53; RPD=1.96) is more suitable to map
the soil salinity than the linear model (PLSR)
(R2=0.70; RMSE=6.95; RPD=1.82) and the performance of
the MARS model was improved using the continuum-removed spectral data in
400-2500 nm wavelength range. On the contrary to these results,
Improvement in prediction using the non-linear model might be due to
their capability to fit the complex and non-linear relationships
(Friedman, 1991; Nawar et al. , 2014; Volkan Bilgili et
al. , 2010). The studies have demonstrated that the high soil salinity
implies a non-linear relationship between the measured salinity and the
spectral reflectance (Farifteh et al. , 2007; Sidike et
al. , 2014; Weng et al. , 2008). Farifteh et al. , (2007)
considered the PLSR as advantageous model over ANN as the prediction
accuracies were similar and low time requirement for the establishment
of the model and reproducibility of the final model. But, in the present
investigation, the predictions using the non-linear models (MARS and
SVR) were comparatively better than those by linear.