Building a pedotransfer function for soil bulk density on regional dataset and testing its validity over a larger area

- Generalized boosted models algorithm was used to build pedotransfer functions for soil bulk density on Region Centre data.
- Distance matrices were used to determine the validity domain of pedotransfer functions.
- Validity domain successfully excluded dissimilar samples in mainland France.
- Additional sampling strategy improved the prediction ability of pedotransfer functions.

Though soil bulk density (BD) is important for crop growth and land management, information about BD is often missing in soil database due to the fact that the determination of BD is time consuming and labor intensive. In order to fill this gap, pedotransfer functions (PTFs) have been developed for predicting bulk density during last decades. To avoid non-valid extrapolation, the validity domain of PTFs should be investigated. In this study, a PTFs for bulk density was built using GBM model on data from Region Centre and we tested its validity on data from mainland France except Region Centre. Standardized Euclidean distance was applied to distinguish dissimilar soil samples from Region Centre data and other regions data. The established PTFs were not suitable for predicting these dissimilar samples. In order to make a balance between prediction accuracy and the number of samples can be predicted with our PTFs, different cutoff limits of Standardized Euclidean distance was compared. When optimized distance cutoff limit was at 97%, compared to result without excluding dissimilar samples, RMSPE value on test data decreased from 0.179 to 0.163 g cm− 3. An additional sampling strategy based on these dissimilar samples can improve the predictive ability of PTFs. We also suggest that the purposive sampling strategy or the systematic and dense sampling strategy will do large contribution in robust modelling of PTFs.