Estimating soil total nitrogen contents from visible/near-infrared reflectance based on successive projections algorithm

The Visible/near-infrared reflectance (Vis/NIR) spectroscopy combined with multiple regression methods have been employed to assess various soil properties quantitatively. However, useless or irrelevant information in some bands worsen the performances of the calibrated models. Successive projections algorithm (SPA) is an effective technique for selecting informative variables, and its application potential in estimating soil contents using visible/near-infrared reflectance (Vis/NIR) spectroscopy has not been explored. This paper aimed to explore the effect of SPA combined with support vector machine regression (SVMR) on soil total nitrogen (TN) contents estimation using Vis/NIR spectroscopy. For this purpose, SPA-SVMR was compared with full spectrum based SVMR (FS-SVMR) for soil total nitrogen (TN) estimation, respectively. The results showed that the SPA-SVMR model (R²_v = 0.58, RMSEV = 0.364) performed better than the FS-SVMR model (R²_v = 0.49, RMSEV = 0.394). We concluded that the SPA could eliminate the uninformative variables and improve the performance of SVMR on estimating TN contents. In further study, using the SPA-SVMR method to estimate other soil properties is necessary to test this finding