This dataset contains pixel-wise soil organic carbon (OC) predictions for agricultural fields in the Belgium, produced using a multi-source data fusion approach. Predictions are provided as GeoTIFF files, one per field per model configuration. Models were trained on field-sampled OC reference measurements combined with three remote sensing sources: (1) VTT close-range hyperspectral imagery (6 bands, calibrated to reflectance), (2) Kuva Space L2A airborne hyperspectral imagery (22 bands), and (3) Sentinel-2 L2A multispectral imagery (10 bands: B2–B8A, B11, B12). A bare-soil filter (NDVI < 0.35, taking the minimum NDVI across all satellite sources) was applied before training to reduce vegetation contamination. Four feature configurations were evaluated — VTT only, VTT + Kuva, VTT + Sentinel-2, and VTT + Kuva + Sentinel-2 (full fusion) — each trained with both Random Forest and XGBoost regressors under 5-fold cross-validation. The prediction maps in this dataset were produced with the full-fusion (VTT + Kuva + Sentinel-2) configuration.