Visible–near-infrared diffuse reflectance spectroscopy (vis–NIR DRS) has been widely used to predict soil organic carbon (SOC) in the laboratory. Predictions made directly from soil spectra measured in situ under field conditions, however, remain challenging. This study addresses the issue of incorporating in-situ reflectance spectra efficiently into calibration data when a few field measurements only are available. We applied the synthetic minority oversampling technique (SMOTE) to generate new data with in-situ reflectance spectra from soil profiles. Subsequently, we combined existing spectral libraries with these new synthetic data to predict SOC by partial least squares regression (PLSR). We found that models with added synthetic spectra always outperformed models based on the spectral libraries alone and in most cases also those with added in-situ spectra only. We used the models to predict the distribution of SOC in soil profiles under five different land uses at Mount Kilimanjaro (Tanzania). Based on our results, we propose a framework for predicting SOC with a limited number of in-situ soil spectra. This framework could effectively reduce the costs of developing in-situ models for SOC at the local scale.