3.4.8 Soil Na prediction
The MARS model gave the highest prediction accuracy (R2c=0.87; MBEc=-0.98; RMSEc=10.65; RPDc=1.79) during calibration to predict the Na and the predictions were acceptable. The second-best model was SVR with a prediction accuracy of R2c=0.74; MBEc=-0.66; RMSEc=15.67; RPDc=1.21. The Na prediction using validation dataset had non-reliable prediction accuracy (R2p=0.55; MBEp=4.81; RMSEp=20.63; RPDp=0.82) with the SVR model. The prediction accuracy was the highest with SVR compared to other models. The best overall rank of 2.38 was recorded by SVR (Figure 4h). Malmir et al. (2019) reported acceptable Na prediction using PLSR and soil spectral data from 400 to 1000 nm. A very good prediction of Na extracted from soil saturation (n=402, R2=0.88, RMSE=2.45, RPD=2.89, excellent) using laboratory-measured soil spectral reflectance (350-2500 nm) was reported by Das et al. (2015). Islam et al. (2003) reported poor prediction for exchangeable Na using PCR modeling of spectral data in the range of 250-2500 nm. Similarly, Chang et al. (2001) reported inaccuracy in predicting Mehlich-3 extractable exchangeable Na using NIR-PCR model (R2<0.50) for a spectral range of 400-2498 nm. Poor prediction (R2=0.33) of exchangeable Na using the MIR spectroscopy was observed by Janik et al. (1998). Better performance of the non-linear model could be due to the fact that the relationship between spectral data and soil characteristic is rarely linear in nature. Variability in prediction accuracy of Na using reflectance spectroscopy could be due to variations in modeling, extraction or spectral data collection methods (field, processed or dried, intact core soil samples).