Séance Séminaire

Séminaire de Probabilités et Statistique

Une extension de CART aux données spatiales

Most of statistical learning techniques, such as Classification And Regression Trees (CART), were elaborated for independent samples to compute classification rules. This assumption is very practical to compute the empirical risk and asymptotic properties of estimators. In many environmental or ecological applications, data are from samples of some regionalized variables and the implicit assumption of independence is not acceptable. the sampling scheme is very irregularly located and sometimes strongly clustered. CART algorithm was adapted to the case of spatially dependent samples. We propose two approaches: the first one take into account irregularity of the sampling while the second one used spatial estimates involved in the construction of the discriminant rule. These methods were compared on a simulation study. The necessity of taking into account pattern of the sampling design was clearly illustrated on paleoecological data. This work is a result of a collaboration with L. Bel, D. Allard, J-M. Dubois and R. Cheddadi.