Combining multivariate method and spectral variable selection for soil total nitrogen estimation by Vis–NIR spectroscopy

Abstract

ABSTRACT Successful estimation of soil total nitrogen (TN) content by visible and near-infrared (Vis-NIR) reflectance spectroscopy depends on selecting appropriate variable selection techniques and multivariate methods for regression analysis. This study aimed to explore the potential of combining multivariate method and spectral variable selection for soil TN estimation by Vis-NIR spectroscopy. 95 soil samples were collected from Jiangsu Province, China, and their TN contents, and reflectance spectra were measured. Four multivariate methods (extreme learning machine, ELM; backpropagation neural network, BPNN; support vector machine regression, SVMR; partial least squares regression, PLSR) combined with three variable selection techniques (competitive adaptive reweighted sampling, CARS; genetic algorithm, GA; successive projections algorithm, SPA) were used for model calibration. Results showed that the ELM model outperformed the BPNN, SVM, and PLSR models. The CARS was superior to GA and SPA techniques in selecting effective variables. The best estimation accuracy (R2 = 0.79) was obtained by the ELM-CARS model. Furthermore, the output of the ELM-CARS model presented the highest similarity to the standard soil TN fertility grades, with a correct classification rate of 82.9%. In conclusion, ELM combined with CARS has great potential to estimate the soil TN content and assess the TN fertility levels using Vis-NIR spectroscopy.