The objectives of this work were to compare the performance of two variable selection methods, namely, Monte Carlo–uninformative variable elimination (MC-UVE) and successive projections algorithm combined with partial least squares regression (PLSR), multiple linear regression (MLR), and least squares–support vector machine (LS-SVM) for the measurement of soil nitrogen (N), organic carbon (OC), available phosphorous (P), and available potassium (K) using near-infrared (NIR) spectroscopy, and provide interpretation for the effective variables. A total of 198 soil samples collected from nine towns in Wencheng County, People’s Republic of China, were scanned by a Fourier-type NIR spectrometer. The entire data set was split randomly into calibration set (62%), validation set (19%), and prediction set (19%). The calibration set was used to build calibration models with a number of spectral variables optimized by the validation set. The prediction set was used for independent prediction of the established models. The proposed combination of MC-UVE-LS-SVM achieved the optimal prediction performance compared with PLSR, LS-SVM, successive projections algorithm–MLR, and MC-UVE-PLSR. The values of coefficient of determination (R2) and residual prediction deviation were 0.88, 0.89, 0.59, and 0.75, and 2.9, 3.1, 1.5, and 2.0 for N, OC, P and K, respectively. The results showed that MC-UVE was able to select important variables from the NIR spectra, and LS-SVM was preferable to linear PLSR and MLR for the prediction of soil properties in this work. Analysis of effective variables indicated that: some of the effective variables of N and OC were directly associated with their functional groups, whereas some impacted on the prediction accuracy by measuring soil moisture content. Soil P and K were found to be measurable with different levels of agreement, which was partly attributed to the covariation of moisture content and illite in soil.