Mapping of soil parent material in Brittany: learning with points as training data for regional extrapolation
Résumé
A machine learning system (MART) was used to predict soil parent material at regional scale. We tested the use of punctual soil observations as training data, instead of soil maps used in previous studies, which enables a more even distribution of the training data over the study area (Brittany, NW France). This region shows mainly metamorphic, igneous and sedimentary substrates. However, superficial deposits (Aeolian loam, colluvium and alluvial material), represent very often the actual soil parent material and are very poorly delimitated by existing geological maps. To calibrate a predictive model, MART requires (i) training data, which correspond to the variable to predict, and (ii) environmental predictors. Model predictions were compared (i) to punctual data not used to calibrate the model (crossvalidation), (ii) to the whole punctual dataset, and (iii) to existing soil maps (external validation). The three kinds of validation showed an accuracy of parent material prediction of 56%, 80% and 54%, respectively. Aeolian loam was among the three best predicted substrates. Poor prediction was associated with uncommon materials and areas with high geological complexity, where existing maps used for external validation were particularly imprecise. Parent material prediction can be use to predict other soil properties, like soil waterlogging