The agro-pastoral ecotone is a transitional zone between farming areas and grassland pastoral areas. Accurate estimation and monitoring of soil organic matter (SOM) has important significance for carbon pool estimation and agricultural production. Taking the typical agro-pastoral ecotone in northeast China as the study area, Landsat 8 OLI and ALOS 12.5m DEM as the data sources, the input variables included band reflectivity, reflectivity logarithm, brightness index and terrain factors. The multi-spectral inversion model of SOM in the agro-pastoral ecotone was constructed by using multiple linear stepwise regression (MLSR) model, random forest (RF) model and BP neural network (BPNN) model, respectively. The results showed that: 1) According to the order of importance, the logarithm of band 4, band 5, band 6 and brightness index of Landsat 8 OLI were selected as input variables, and the accuracies of RF and BPNN models were better than that of MLSR model. 2) After adding elevation (E) and slope of aspect (SOA), the prediction accuracies of the three models all improved, and the accuracy of BPNN model improved most, with R² increased by 0.22 and RMSE decreased by 0.40 g/kg. The optimal inversion accuracies of the three models from high to low was: BPNN model (R²=0.82, RMSE=1.4 g/kg) > RF model (R²=0.71, RMSE=1.9 g/kg) > MLSR model (R²=0.66, RMSE=8.8 g/kg). The research can provide methodological support for the study of SOM spatial and temporal changes in agro-pastoral ecotone.