Soil organic matter (SOM) content is affected by a variety of factors that operate at different scales. The purpose of the present study was to investigate the multi-scale spatial relationships between SOM content and influencing factors, including elevation, slope, topographic wetness index, soil bulk density, sand content, silt content, clay content, and soil spectral components. Soil samples were collected from the Taiyuan basin, a typical area of the Chinese Loess Plateau, by establishing sampling transects at the upper, middle, and lower parts of the basin. At different locations in the basin, the multi-scale correlations of SOM content with the influencing factors were analyzed, and the SOM contents at the sampling scale were predicted using the multivariate empirical mode decomposition. The results showed that (1) the multivariate empirical mode decomposition method could separate the transects of SOM content into six, eight, and seven intrinsic mode functions for the upper, middle, and lower parts of the basin, respectively. A scale of 1000 m was the main representative scale for SOM content in the entire basin, and the number of representative scales increased along the river. (2) In contrast to the correlation at the sampling scale, the multi-scale spatial correlation between SOM content and the influencing factors revealed that the correlation of elevation and SOM content was dominant at larger scales. Meanwhile, the correlation of slope and SOM content was significant for the middle part of the basin, and the relationship between the topographic wetness index and SOM content was significant for both the middle and lower parts. However, the correlation between soil bulk density and SOM content was much more complex and differed at the various scales and locations; in the upper part of the basin, the relationship between the silt and SOM contents was more apparent than that of the sand and clay contents. In addition, spectral component 1 was significantly correlated with SOM content in the entire basin. (3) The multivariate empirical mode decomposition method was more accurate at predicting SOM content than the stepwise multiple linear regression. Therefore, taken together, the results of the present study provide a basis for soil digital mapping, dimension design of farmland, and SOM content prediction on the Chinese Loess Plateau.