LAUSR

Abstract Spectral variable selection is widely accepted as an important step in the quantitative analysis of visible and near-infrared (Vis–NIR) reflectance spectroscopy (350–2500 nm), as it tends to have parsimonious data representation and can result in multivariate models with greater predictive ability. In this study, a total of 123 intact soil cores (8.4 cm internal diameter and 40 cm long) were collected from paddy fields in Yujiang, China. The Vis–NIR spectra, measured in the laboratory for flat, horizontal surfaces of soil core sections at two depths (5 cm and 10 cm), were used to estimate the rice root density by support vector machine regression (SVMR). In addition to full-spectrum SVMR (FS-SVMR), four spectral feature selection techniques (competitive adaptive reweighted sampling, CARS; genetic algorithm, GA; successive projections algorithm, SPA; and uninformative variable elimination, UVE) were applied with SVMR (i.e., CARS-SVMR, GA-SVMR, SPA-SVMR, and UVE-SVMR) to determine the technique with the most accurate predictions. The coefficient of determination ( R 2 ), root mean square error (RMSE), and residual prediction deviation (RPD) were used to evaluate the accuracy of the optimized calibration models. Based on the independent validation data set ( n  = 73), the order of the prediction accuracy was CARS-SVMR ( R 2 P  = 0.92; RMSE P  = 2.56; RPD P  = 3.42) > GA-SVMR ( R 2 P  = 0.91; RMSE P  = 2.92; RPD P  = 2.99) > FS-SVMR ( R 2 P  = 0.88; RMSE P  = 4.28; RPD P  = 2.83) > SPA-SVMR ( R 2 P  = 0.87; RMSE P  = 4.43; RPD P  = 2.55) > UVE-SVMR ( R 2 P  = 0.86; RMSE P  = 4.49; RPD P  = 2.51). The proposed method of CARS-SVMR had a relatively strong ability for spectral variable selection, with the fewest variables ( n  = 60) and the shortest computation time (0.12 min), while retaining excellent prediction accuracy. In conclusion, SVMR combined with CARS has great potential to accurately determine the rice root density of intact soil cores of paddy fields using Vis–NIR spectroscopy.