Re-using models trained on a specific image acquisition to classify landcover in another image is no easy task. Illumination effects, specific angular configurations, abrupt and simple seasonal changes make that the spectra observed, even though representing the same kind of surface, drift in a way that prevents a non-adapted model to perform well. In this paper we propose a relative normalization technique to perform domain adaptation, i.e. to make the data distribution in the images more similar before classification. We study optimal transport as a way to match the image-specific distributions and propose two regularization schemes, one unsupervised and one semi-supervised, to obtain more robust and semantic matchings. Code is available at http://remi.flamary.com/soft/soft-transp.html. Experiments on a challenging triplet of WorldView2 images, comparing three neighborhoods of the city of Zurich at different time instants, confirm the effectiveness of the proposed method that can perform adaptation in these non-coregistered and very different urban case studies.