With 298 heterogeneous soil samples from Yixing (Jiangsu Province), Zhongxiang and Honghu (Hubei Province), this study aimed to combine a successive projections algorithm (SPA) with a support vector machine regression (SVMR) model (SPA-SVMR model) to improve the estimation accuracy of soil organic carbon (SOC) contents using the laboratory-based visible and near-infrared (VIS/NIR, 350−2500 nm) spectroscopy of soils. The effects of eight spectra pre-processing methods, i.e., Log (1/R), Log (1/R) coupled with Savitzky-Golay (SG) smoothing (Log (1/R) + SG), first derivative with SG smoothing (FD), second derivative with SG smoothing (SD), SG, standard normal variate (SNV), mean center (MC) and multiplicative scatter correction (MSC), on SPA-based informative wavelength selection were explored. The SVMR model (i.e., SVMR without SPA) and SPA-PLSR model (i.e., SPA combined with partial least squares regression (PLSR)) were developed and compared with the SPA-SVMR model in order to evaluate the performance of SPA-SVMR. The results indicated that the variables selected by SPA and their distributions were strongly affected by different pre-processing methods, and SG was the optimal pre-processing method for SPA-SVMR model development; the SPA-SVMR model using SG pre-processing and 28 SPA-selected wavelengths obtained a better result (R2V = 0.73, RMSEV = 2.78 g·kg−1 and RPDV = 1.89) and outperformed the SVMR model (R2V = 0.72, RMSEV = 2.83 g·kg−1 and RPDV = 1.86) and the SPA-PLSR model (R2V = 0.62, RMSEV = 3.23 g·kg−1 and RPDV = 1.63). Most of the spectral bands used by the SPA-SVMR model over the near-infrared region were important wavelengths for SOC content estimation. This study demonstrated that the combination of SPA and SVMR is feasible and reliable for estimating SOC content from the VIS/NIR spectra of soils in regions with multiple soil and land-use types.