3.4.5 Soil N prediction
For N, excellent calibration prediction accuracy of R2c=0.87; MBEc=0.10; RMSEc=35.16; RPDc=2.49 was achieved by using the MARS model. The validation prediction accuracy was non-reliable with R2p=0.58; MBEp=-8.80; RMSEp=62.08; RPDp=1.34. Among all the tested models, the overall ranking of MARS (2.38) and PCR (2.38) was found the best. However, the validation prediction rank of PCR (1.25) was best to predict the N. Similar to our findings Xu et al. (2018) reported better prediction accuracy for total nitrogen (TN) using PCR as one of the calibration models with R2p=0.78-0.86 and RPDp=2.13-2.69. Yet in other studies, support vector machine regression (SVMR) and back propagation neural network (BPNN) showed better performance with R2p=0.69-0.88 predicting TN (Aliah Baharom et al. , 2015; Cozzolino et al. , 2013; Kodaira & Shibusawa, 2013; Kusumo et al. , 2008; Wenjun et al. , 2014). Bands around 1100, 1600, 1700-1800, 2000 and 2000-2400 nm have been identified as being important for SOC and TN (Stenberg et al. , 2010). Martin et al. (2002) found a high correlation (r=0.96) for the NIR predicted soil C and N. According to Williams & Norris (2001) prediction of N could be due to the known nitrogen specific absorption bands such as covalent bonds with H, C or O.