The spectral assessment of soil properties is handicapped by the fact that spectral predictive mechanisms often vary from one population to another. In a landscape approach, heterogeneous conditions with a wide variety of combinations of spectrally active factors have to be considered. Heterogeneity, however, is one main reason for poor predictions from spectroscopic data, as an optimal calibration needs limited but sufficient set heterogeneity. For our study, the investigated plots were located in an area that covered about 600 km²; geologic conditions and sampled soil types were highly variable. In total, 172 soil samples were taken from the top horizon of agricultural fields, afterwards analysed in the laboratory for total organic carbon (OC) and black carbon (BC) and additionally measured with a full range ASD FieldSpec-instrument. The heterogeneity of the sample set was reflected by both the analysed soil parameters and the measured soil spectra. As a consequence, one “global” calibration model (with PLSR) provided only moderate results for the studied soil variables. In the following we focused on two issues, which were i) to replace the global calibration by local calibration procedures, and ii) to study the effect of spectral variable selection for calibration success. For the CARS selection procedure (“competitive adaptive reweighted sampling”), the results demonstrated that more accurate estimates can be obtained using selected variables instead of the full spectrum.