The ANN models predicted impossible (< 0) and highly implausible C/N ratios (< 15) for a large number of pixels in the cores amended with green manure (Table 2 and S2), with up to 47 % of pixels containing predictions outside of the calibration range and deemed unreliable. For the alkyl ratio, the number of pixels predicted outside the calibration range using the ANN ensemble was very similar to that of C/N and corresponded to largely the same pixels. We thus considered the ANNs unreliable and used the weighted averaged models from the RF and the median predictions from the random ensemble ANNs for predictive purposes of the two types of OM amendment and for further interpretation. The model averaging resulted in the reduction in the proportion of pixels predicted out of the calibration range (20 % for green manure and < 0.2 % for straw).