Pedotransfer functions (PTFs) are used to estimate soil properties that are difficult and costly to measure, from others properties that are available. MART, namely Multiple Additive Regression Trees belongs to the boosted regression trees (BRT) family. It has been applied in various scientific fields such as remote sensing, ecology and prediction of species distribution, medicine and chemometrics and only very recently to soil science. The MART method, which includes the use of stochastic gradient boosting, is known for having a set of interesting properties, although as for other techniques such as neural networks, attention must be paid to overfitting behavior. It can work with either qualitative or quantitative predictive variables, can handle missing data, correlated predictive variables and is robust to the presence of outliers within the dataset and to the use of irrelevant predictor variables. It comes with different output for interpreting the results and assessing the validity of the fit. Here, we present development of PTFs using MART for diverse soil science application as estimation of missing values of bulk density of French metropolitan soils, prediction of soil carbon stocks in Guadeloupe (French Caribbean Island) and development of correspondence function between different methods of heavy metals analysis (aqua regia and total analysis, i.e. inductively coupled plasma mass spectroscopy after dissolution with hydrofluoric and perchloric acids). MART proved to be a versatile and convenient tool for building such functions without much a priori knowledge about the relationships between response variable and predictors. MART was able to grasp the full dataset diversity when fitting PTFs as challenging as PTF for bulk density. (Texte intégral)