This paper addresses domain adaptation for the pixel-wise classification of\nremotely sensed data using deep neural networks (DNN) as a strategy to reduce\nthe requirements of DNN with respect to the availability of training data. We\nfocus on the setting in which labelled data are only available in a source\ndomain DS, but not in a target domain DT. Our method is based on adversarial\ntraining of an appearance adaptation network (AAN) that transforms images from\nDS such that they look like images from DT. Together with the original label\nmaps from DS, the transformed images are used to adapt a DNN to DT. We propose\na joint training strategy of the AAN and the classifier, which constrains the\nAAN to transform the images such that they are correctly classified. In this\nway, objects of a certain class are changed such that they resemble objects of\nthe same class in DT. To further improve the adaptation performance, we propose\na new regularization loss for the discriminator network used in domain\nadversarial training. We also address the problem of finding the optimal values\nof the trained network parameters, proposing an unsupervised entropy based\nparameter selection criterion which compensates for the fact that there is no\nvalidation set in DT that could be monitored. As a minor contribution, we\npresent a new weighting strategy for the cross-entropy loss, addressing the\nproblem of imbalanced class distributions. Our method is evaluated in 42\nadaptation scenarios using datasets from 7 cities, all consisting of\nhigh-resolution digital orthophotos and height data. It achieves a positive\ntransfer in all cases, and on average it improves the performance in the target\ndomain by 4.3% in overall accuracy. In adaptation scenarios between datasets\nfrom the ISPRS semantic labelling benchmark our method outperforms those from\nrecent publications by 10-20% with respect to the mean intersection over union.\n