3.4.7 Soil K prediction
The prediction of the K during calibration was the best when SVR was used (R2c=0.95; MBEc=-27.59; RMSEc=246.34; RPDc=3.35) and it was considered as an excellent prediction (rank 1.00). It was followed by MARS which also generated excellent predictions (R2c=0.88; MBEc=-43.62; RMSEc=346.90; RPDc=2.38) and had a rank of 2.00. For validation, the highest prediction accuracy (R2p = 0.46; MBEp = 35.46; RMSEp=604.31; RPDp=1.31) was achieved using the PLSR, but was considered as non-reliable (Figure 4g). The validation and an overall ranks of PLSR were 1.25 and 2.25, respectively. Xu et al. (2018) reported better and improved prediction for soil total potassium (TK) with R2=0.55-0.77 using SVMR and BBPNN models over the PLSR model by Cozzolino et al. , (2013) and Schirrmann et al. , (2013). Better performance of SVMR and BBPNN models for predicting P and K might be attributed to non-linear behavior of the soil variables with spectral reflectance data which was better captured by SVMR and BBPNN compared to the best performing model (PLSR) identified in our study. Poor prediction of soil P and K using the VIS-NIR region have been reported for lab and field-based spectroscopy in numerous studies (He et al. , 2007; Kuang & Mouazen, 2011; Malmir et al. , 2019; Viscarra Rossel et al. , 2006). Wenjun et al. (2014) recorded poor predictions for soil P and K using the lab and field-based VIS-NIR spectroscopy. Malmir et al. (2019) found the inability of PLSR model to predict soil P and K using VIS-NIR (400-1000 nm) spectroscopy. Based on the overall ranking, the PLSR model for predicting soil P and K found to be the better compared to others, which could be due to the high variability of P and K content and broader spectral range (350-2500 nm).