The study of land use at large scale has recently been approached as the description of a territory in terms of Local Climate Zones (LCZ). These are subdivisions of a landscape into a set of categories which are of uniform coverage, represent similar land cover characteristics and patterns and are of several hundreds of meters in surface. Specifically, we map the region of NorthRhine-Westfalia (Germany), a region of approx. 34'000 km2, in terms of LCZ, with a statistical model trained using multimodal geospatial data (remote sensing, land cover, 3D models and volunteered geographic information) and constrained with both spatial interaction and scaling behavior logics. The proposed model improves LCZ classification accuracy, in particular for scarce classes, with an increase of up to 4 κ points and 5% in average accuracy as compared to conventional pixelwise classification.