Soil organic matter (SOM) affects circulation and stability of the soil C pool, and plays a key role in global C change and local ecological balance. Timely and accurate acquisition of large‐scale farmland soil organic carbon (SOC) information is of great significance to the study of SOC reserves and distribution, farmland soil ecological security, and scientific crop management. In this study, we have evaluated the potentiality of visible‐near‐infrared spectroscopy to along with machine learning algorithms to predict SOM content. At present, soil characterization is mainly focused on hyperspectral data measured by remote‐sensing spectral index and spectrometer. This article aims to establish a statistical machine‐learning model to study the efficiency of spectral data to comprehensively evaluate SOM. The results show that after transforming the different forms of the original spectrum and combining with successive projection algorithm (SPA), the characteristic wavelength of organic matter at 350–2,500 nm can be effectively extracted. A combination of spectroscopy and chemometrics techniques can be used as a practical, rapid, low‐cost, and quantitative method to evaluate SOM. Based on SPA, partial least squares regression, support vector machine, random forest, a SOM prediction model with optimal characteristic wavelength is constructed. The best spectral processing method is continual removal combined with SPA, the validation coefficient of determination, root mean square error prediction, and ratio of performance deviation values are 0.82, 1.12, and 2.87, respectively.