The importance of zeros in digital soil mapping II: A case study of depth-to-bedrock mapping in New Brunswick, Canada

The reliance on legacy data for digital soil mapping (DSM) applications leads to unintended data distributions that researchers have little control over. One such case is the abundance of excess zeros which may be present for some soil properties, causing zero-inflation (ZI). If ignored, ZI can cause underfitting of extremes in a dataset, violation of statistical assumptions, and other modelling challenges. In this study, a two-part modelling framework was utilized to address ZI in a depth-to-bedrock mapping case study. Several machine learning algorithms were tested within the two-part modeling framework, with subsampling of the classification algorithms to determine soil absence or presence. The final predicted map was then compared to maps generated from models without accounting for ZI. Finally, to test the effects of training and validating with numerous zeros, the external validation data were subsampled to reflect the known landscape with respect to existing bedrock outcrop maps. Accuracies of the ZI models were highest using an over-sampled subset for training, with the best performing ZI model including an over-sampled random forest classification and a random forest regression component. ZI models consistently outperformed models that ignored ZI, seeing a modest increase of 0.05 concordance among the strongest models. Once the validation data were sampled to better reflect reality, concordance dropped from 0.55 to 0.31. Ultimately, while accounting for ZI consistently outperformed ignoring it, extreme ZI still caused overly optimistic validation results.