Soil organic carbon (SOC) spatial variability and temporal dynamics are strongly affected by natural and anthropogenic processes occurring at the landscape scale. Spatiotemporal modeling is increasingly used to understand these dynamics, and the initialization of the SOC stocks prior to modeling is a key point. This study aims at modeling SOC distribution at high resolution, for an area of 10 km² in a complex agricultural landscape (NW France). 200 points were selected using conditioned Latin hypercube sampling in order to cover the whole range of ancillary variables (elevation, Modified Compound Topographic Index, K emissions and land use). This sampling strategy enables to select a limited number of sampling sites covering the study site heterogeneity. We used a data mining tool, Cubist, to build a rulebased predictive model and predict SOC at 8 different depths up to 105 cm. Predictive environmental data consisted in the data used in the conditioned Latin hypercube sampling, to which were added topographic attributes derivated from the DEM and geological variables. Two independent datasets were used to evaluate our predictive maps at the landscape scale and at hedgerow proximity. The respective RMSE for these datasets were 7.47 and 4.77 g/kg. The best prediction levels were obtained for the 15-to-60-cm soil layers. The final maps show that the carbon stocks in the soil below 30 cm were not negligible. The SOC 3D map we obtained will be used as soil data input in a soil evolution model, coupling SOC dynamics and soil erosion fluxes.