Remote sensing retrieval has been widely used for dynamic monitoring of the physical and chemical properties of regional soil, but there are few studies on the areas with low organic carbon content and uneven underlying surface which have unremarkable soil spectral characteristics. The cinnamon soil belt in Loess Plateau has multiplicity topography, widely distributed hills and low organic matter content. Large areas soil degradation caused by mining activities has resulted in the fact that soil spectroscopy characteristics are strongly disturbed, which has some inhibiting effect on the remote sensing retrieval accuracy of soil organic carbon content at the regional scale. Based on cinnamon soil belt with typical coal mining subsidence area in Shanxi Province as an example, this research used the surface reflectance and outdoor sample data from the field of coal mining subsidence area to retrieve soil organic carbon content. Conducting comparative experiments on the atmospheric correction methods of the Landsat8 OLI image in the study area by the FLAASH model and the 6SV model combined with high spatial and temporal resolution aided meteorological data to analyze the effect on soil spectral curve and organic carbon content in the mining subsidence area of the cinnamon soil belt and recognize sensitive bands. Multiple linear regression(MLR), BP neural network(BP) and partial leas squares regression(PLSR) model were established to retrieve soil organic carbon content by using the original spectral reflectance R and mathematical transformation forms such as R, log (1/R) and R′. The results showed that the atmospheric correction effect of the 6SV model was better than that of the FLAASH model which could effectively eliminate the interference of atmosphere and topography to reflectance. The reflectance of visible light decreased and the near-infrared rose obviously. The soil reflectance spectra of different organic matter content was clear. The bands of 640~670, 850~880, 1 570~1 600, 2 110~2 290 nm were highly indicative of soil organic carbon content. Compared with multiple linear regression (Coefficient of determination R2 was 0.765) and BP neural network (R2 was 0.767), the partial least-squares regression model had the highest retrieval accuracy (R2 was 0.778). It was found that the 6SV atmospheric correction model and partial least squares regression modeling combined with aided meteorological data which had high spatial and temporal resolution could significantly improve the retrieval accuracy of soil organic carbon in the mining subsidence area of the cinnamon belt. The soil organic carbon content in the study area from 2013 to 2015 was retrieved based on this model. Results showed that the soil organic carbon content in the middle of the study area was higher than that in both sides, and the soil organic carbon content was restored by reclamation. The results can be used to reveal the spatial-temporal distribution of soil organic carbon in the mining subsidence area of the cinnamon belt in the Loess Plateau, providing theoretical and technical support for improving regional soil spectral analysis, land reclamation evaluation, establishment of carbon flux observation network in mining subsidence area of the cinnamon belt and estimation of soil carbon pool, which provides the basis for the ecological sustainable development of the cinnamon belt in the regional and global scales.