Portable mid-infrared spectroscopy (pMIRS) combined with machine learning was used to predict selected parameters for soil organic carbon (SOC) storage. In particular, SOC, soil inorganic C (SIC), hot-water extractable C (hwC), clay and sand content were predicted for ten vineyards with varying geopedological settings. As a pre-test, spectra were collected from sieved and pressed tablets with 30 and 90 kPa respectively and compared to powdery samples in order to optimise sample preparation. Further, spectra from 30 kPa tablets were used to calibrate prediction models for a sample set (n = 540) of 10 vineyards with pronounced geopedological variation using Support Vector Machines (SVM). The calibrated SVM models performed well with R2 = 0.81–0.98 and RPIQ = 5.20–13.0 for all investigated parameters. Third, two years after the calibration samples, follow-up samples were collected from four of the vineyards. While the models performed excellent for hwC (R2 = 0.93), prediction accuracy for SOC was lower. Segmentation of the total dataset into SIC-free and SIC-containing samples resulted in better predictions of SOC of the first sampling period. For the prediction of the follow-up sampling dates, model performance could not be maintained. We conclude that pMIRS-SVM calibrations are suited for the prediction of parameters related to soil C storage under varying geopedological conditions and may provide potential for future C monitoring. Extending the database with additional samples from geopedological scenarios not included in this dataset may strengthen model robustness and help to evaluate effects of SIC content on model performance.