Pedotransfer functions for estimating soil bulk density from existing soil survey reports in Brazil

Soil bulk density (Db) measurement is essential to estimate soil carbon reserves. However, field sampling, specially at depth, and direct measurement of Db is labor intensive, tedious and often impractical. Thus regression models or pedotransfer functions (PTFs) based on easily measured soil properties are an alternative for laborious Db measurements. A forward stepwise multiple regression routine was used to predict Db from 17 soil properties using a data set (data set 1) constructed from the Soil Archives of Embrapa Solos, Rio de Janeiro, Brazil. A first exploratory regression model using 1002 soil samples of data set 1 led to the development of a model, which indicated that total nitrogen (N), clay content (Clay) and sum of basic cations (SB) were the three strongest contributors to Db prediction (Adjusted-R2 = 0.71, Standard error of the estimate = 0.10). A simplified regression model was developed using easily-measured soil attributes, such as soil organic carbon (TOC), Clay and SB. This model described 66% of the variation of Db in 1396 soil samples distributed at all depths (Standard error of the estimate = 0.11). Clay content showed a good correlation with predicted Db (Beta value = − 0.58) followed by TOC (Beta value = − 0.51) and SB (Beta value = 0.21). Partitioning the data set 1 into groups by soil depth (0–30 and 30–100 cm) and soil order (Latossolos and Argissolos) did not improve the accuracy of regression equations. In addition, we tested the Db predictive potential of the proposed model and three existing models (two from Brazil and one from the US) on an independent soil data set (data set 2). A general overestimation of predicted Db by the US model, with mean predicted error (MPE) of 0.11 shows that published PTFs developed on different environments should be used with care. Existing Brazilian models developed for the Amazon biome, on the other hand, were found to produce a slight underestimation with MPE values ranging from − 0.03 to − 0.16. The proposed simple regression model including Clay, TOC and SB was observed to be the most accurate and least biased.