Soil Organic Matter (SOM) is a crucial soil parameter that affects the sowing rate. Real-time control of the sowing rate based on SOM information is the cutting-edge research area of planting technology, which can make full use of land resources to tap the yield potential, accurately and adequately adjust the number of seeds to maximize the return. This article focuses on the North China Plain, one of the main corn-producing areas, as the study area, and the sandy loam soil in this area has been collected by visible-near infrared (300~2 500 nm) spectra. Monte Carlo cross-validation is used to eliminate abnormal samples, and the Savitzky-Golay convolution smoothing method is used to smooth and denoise the spectral data. Respectively through Competitive adaptive reweighted sampling (CARS), Successive projections algorithm (SPA), Competitive adaptive reweighted sampling-Successive projections algorithm (CARS-SPA), Uninformative variables elimination (UVE) and Variable Combination population Analysis (VCPA), and other wavelength screening methods to extract effective variables. Combined with Partial least squares regression (PLSR), the SOM content prediction models of full wavelength and characteristic wavelength were established respectively. The results showed significant differences in the number of wavelengths and wavelength positions screened by different methods. The spectral features selected by the CARS and SPA algorithms are distributed in the spectral range, while the bands selected by UVE and VCPA were concentrated. Moreover, the characteristic variables could be further optimized based on the CARS-SPA method, and the characteristic wavelength was only 15% of the total wavelength. By comparing the modeling and prediction effects of different models, except for the UVE and VCPA algorithms, the models constructed by the other algorithms can all effectively predict the SOM content, and their RPD values were all greater than 2.0. The PLSR model based on CARS-SPA has the best performance. Its RP2 and RPD were 0.901 and 3.188 respectively, higher than other methods. It reduces the interference of invalid information on the prediction effect, but the computational efficiency of the model is significantly improved, which can realize the reliable prediction of SOM content in this area. This research can provide method references for rapid prediction of SOM content and instrument design.