Optimizing Pedotransfer Functions for Estimating Soil Bulk Density Using Boosted Regression Trees

Abstract

Pedotransfer functions (PTFs) are used to estimate certain soil properties that are difficult and costly to measure from others more easily available. Bulk density is one important soil property. Although not requiring complex analysis, its measurement remains time consuming and is lacking in many soil surveys. For several decades, PTFs have been developed for predicting soil bulk density. Most of these PTFs are suited only for specific agro‐pedo‐climatic conditions, however, and can be applied only within a limited geographic area. In this study, we derived and experimented with two new PTFs based on a multiple additive regression trees (MART) method, and assessed their performance compared with existing PTFs when applied to a country‐level soil database, the Réseau de Mesures de la Qualité des Sols (RMQS) survey network. This database was designed to include the major soil types and land uses in France. The first proposed PTF (Model m) involves only three predictors typically found in the existing PTFs for bulk density (C content and texture) and the second one (Model M) includes eight easily accessible quantitative and qualitative predictors (e.g., soil taxon). Both models significantly outperformed existing PTFs. Without arbitrarily partitioning the data set before fitting the model, the m and M MART models yielded R2 values of 0.83 and 0.94, respectively. The predictive quality on independent data, assessed using cross‐validation, was also improved compared with published PTFs, with R2 reaching 0.62 and 0.66 and root mean square prediction errors of 0.123 and 0.117 Mg m−3 for the two MART models.