We investigate the relationship between landscape heterogeneity and the spatial distribution of small mammals in two areas of Western Sichuan, China. Given a large diversity of species trapped within a large number of habitats, we first classified small mammal assemblages and then modelled the habitat of each in the space of quantitative environmental descriptors. Our original two step "classify then model" procedure is appropriate for the frequently encountered study scenario: trapping data collected in remote areas with sampling guided by expert field knowledge. In the classification step, we defined assemblages by grouping sites of similar species composition and relative densities using an expert-class-merging procedure which reduced redundancy in the habitat factor used within a multinomial logistic regression predicting species trapping probabilities. Assemblages were thus defined as mixtures of small mammal frequency distributions in discrete groups of sampled sites. In the modelling step, assemblages' habitats and environments of the two sampled areas were discriminated in the space of remotely sensed environmental descriptors. First, we compared the discrimination of assemblage/study areas by linear and non-linear forms of discriminant analysis (linear discriminant analysis versus mixture discriminant analysis) and of multiple regression (generalized linear models versus multiple adaptive regression splines). The "best" predictive modelling technique was then used to quantify the contribution of each environmental variable in discriminations of assemblages and areas. Mixtures of Gaussians provided a more efficient model of assemblage coverage in environmental space than a single Gaussian cluster model. However, non-linearity in assemblage response to environmental gradients was consistently predicted with lower deviance and misclassification error by multiple adaptive regression splines. The two study areas were mainly discriminated along vegetation indices. However, although the normalized difference vegetation index (NDVI) could discriminate forested from non-forested habitats, its power to discriminate assemblages in Maerkang, where a greater diversity of forest habitat was observed, was seen to be limited, and in this case NDVI was outperformed by the enhanced vegetation index (EVI). Our analyses highlight previously unobserved differences between the environments and small mammal communities of two fringe areas of the Tibetan plateau and suggests that a biogeographical approach is required to elucidate ecological processes in small mammal communities and to reduce extrapolation uncertainty in distribution mapping