Mid-infrared spectral data holds immense potential for accurate, cost-effective, and rapid prediction of soil organic carbon (SOC) content. To enhance the universality of spectroscopic data estimation models, a spectroscopic feature enhancement strategy was employed and combined multiple machine learning methods by using the Stacking algorithm to construct a robust model for estimating SOC content. Various spectroscopic feature enhancement methods and their combinations were applied to enhance the features of mid-infrared soil spectra and select the optimal strategies. The Stacking algorithm was used in conjunction with multiple machine learning methods to build an ensemble model, aiming to improve the model’s generalization ability. The estimation performance of the ensemble model was compared with that of partial least squares regression (PLSR), gradient boosting trees (GBT), and 1-dimensional convolutional neural network (1D-CNN) models. The results demonstrated that the optimal spectral characteristics enhancement strategy can significantly improve the correlation between soil spectra and soil organic carbon content, and the optimal Pearson correlation coefficient reached -0.82. Compared with PLSR, GBT, and 1D-CNN models, the ensemble model exhibited higher estimation accuracy and robustness across various spectral datasets. In particular, under the spectral characteristic enhancement strategy of first derivative combined with multivariate scatter correction, the ensemble model demonstrated excellent estimation performance (R2=0.92, RMSE was 1.18g/kg, RPD was 3.52). The proposed method enabled timely and accurate estimation of SOC, which can provide a scientific basis for modern agricultural management.