While popular solutions exist for land cover mapping, they become intractable when in a large-scale context (e.g. VHR mapping at the European scale). In this paper, we consider a popular classification scheme, namely combination of Differential Attribute Profiles and Random Forest. We then introduce new developments and optimizations to make it: i) computationally efficient; ii) memory efficient ; iii) accurate at a very large scale; and given its efficiency, iv) able to cope with strong differences in the observed landscapes through fast retraining. We illustrate the relevance of our proposal by reporting computing time obtained on a VHR image.