Démarche statistique pour la sélection des indicateurs par Random Forests pour la surveillance de la qualité des sols
Résumé
The volume of data, and the large number of biological variables to be tested (one hundred), require analytical techniques, such as Random Forests, which can overcome the problem of multi-colinearity for the selection of indicators, sensitive to various factors. Random Forests methodology is appropriate for the selection of the most discriminant variables. So, we searched for the best way to select them, by bringing together all biological variables, representing the Microflora and Fauna. This approach focuses on impact indicators from the Bio2 program, indicators of flora and indicators of accumulation (snails) were not included. This work has been implemented on the three factors of discrimination : land use, metallic contamination levels and organic contamination levels. We grouped the most discriminating variables from each RF analysis. Linear discriminant analysis was then implemented for each factor, in order to develop a predictive model.
Origine : Fichiers éditeurs autorisés sur une archive ouverte