Measuring soil organic carbon (SOC) inexpensively and accurately is crucial for soil health monitoring and agricultural decarbonization. Hyperspectral imaging is commonly evaluated as an inexpensive alternative to dry combustion for SOC measurement, but existing end-to-end approaches trained to predict SOC content from spectral data frequently fail to generalize when applied outside of their ground-truth geographic sampling distributions. Using stratified data from the USDA Rapid Carbon Assessment (RaCA), we demonstrate a method to improve model generalization out-of-distribution by training SOC regression alongside models that reconstruct input spectra. Because hyperspectra can be collected from remote platforms such as drones and satellites, this approach raises the possibility of using large hyperspectral Earth observation datasets to transfer SOC inference models to remote geographies where geographically-dense ground-truth data collection may be expensive or impossible. By replacing the decoder with a simple physics-informed model, we also learn an interpretable spectral signature of SOC, confirming its dark hue and expected reflectance troughs. Finally, we show that catastrophic generalization failures can be better addressed with these architectures by fine-tuning on large quantities of hyperspectral data.