Building predictive maps of assemblage habitat, a widely used method in conservation and landscape management, is based on fitting habitat model on a training area, corresponding to a limited region in the environmental space. Model predictive performances need to be robustly evaluated on test data set. This is often realized at the landscape level, i.e taking as test data set one part of the original sample or a resampling one. Predicting assemblage occurrences at a regional level requires a step further in the modelling validation stage: testing model extrapolation performances. We estimated and compared the predictive performances of two scales of predictive mapping of small mammals assemblages in a remote area of Sichuan province. Small mammal assemblages were defined in two distant areas and differed between both areas. Their habitats were modelled, predicted and mapped using ETM bands at two different spatial scales: local (in each area) versus regional (including both areas). While locally trained models provided large predictive errors on independent data sets, the regionally trained model more accurately predicted assemblage occurrences and could be considered at this state of the research as an appropriate method to map assemblage regionally.