We introduce the notion of naive evidential classifier. This classifier, which has a structure mirroring the naive Bayes classifier, is based on the Transferable Belief Model and use mass assignments as its uncertainty model. This new method is based achieves more robust inferences, mainly by explicitly modeling imprecision when data are in little amount or are imprecise. After introducing the model and its inference process based on Smet's generalized Bayes theorem (GBT), we specify some possible methods to learn its parameters, based on the Imprecise Dirichlet Model (IDM) or on predictive belief functions. Some experimental results on an agronomic application are then given and evaluated.