Domain adaptation and label noise tolerant training for the classification of remote sensing data in topographic applications
Franz Rottensteiner(Leibniz Universität Hannover, Germany)

The classification of images and other remote sensing data is a fundamental task to derive semantic information about the objects in the depicted scene automatically. Current research focuses on statistical approaches, in which the knowledge about the objects is given implicitly in the form of training samples that are used to train a classifier. These approaches can be easily adapted to new domains by defining a new representative set of training data. However, this flexibility comes at a cost: the need to generate training data. This presentation will discuss two different strategies for reducing the requirements w.r.t. the availability of training data for supervised learning in topographic applications. The first strategy is Domain Adaptation, a specific setting of transfer learning (Pan & Yang, 2010). Here we assume a large set of potential training data to be available from earlier projects. However, in these training datasets, the features may follow a different distribution than those in the new image which needs to be classified. On the one hand, we may apply source selection techniques to find out which existing datasets are most similar to the new images and can, thus, be used to train the new classifier (Vogt et al., 2018). On the other hand, we can adapt a classifier trained on the existing data by instance transfer, i.e. by successively incorporating samples from the new image that received their class labels (“semi-labels”) from the current version of the classifier (Paul et al., 2016; Vogt et al., 2018). Results show that both source selection and instance transfer can be quite efficient in cases where the domains are not too different. However, it turns out to be difficult to predict whether this prerequisite is fulfilled or not (Paul et al., 2018). The second strategy for reducing the requirements w.r.t. manually generated training samples is to use an existing topographic map of the area covered by the image to be classified for training. In this case, training samples are abundant, but some of them may be wrong due to temporal changes, and the training procedure has to cope with these wrong class labels (label noise; Frénay & Verleysen, 2014). We adapt the method for label noise tolerant logistic regression (Bootkrajang & Kabán, 2012) to take into account that in topographic application, label noise typically occurs in larger spatial clusters corresponding, for instance, to newly developed settlement areas. In addition, the parameters of the noise model that are estimated in the training process are used to provide transition probabilities between individual time steps of multitemporal images. Our experiments show that this leads to high classification accuracies without any new manually labelled training data. Problems mainly occur if newly constructed objects have a different appearance than those of the same type that do already exist in the map (Maas et al., 2016; 2018).

REFERENCES
  • Bootkrajang, J., Kabán, A., 2012. Label-noise robust logistic regression and its applications. Joint European Conf. on Machine Learning and Knowledge Discovery in Databases, pp. 143–158.
  • Frénay, B., Verleysen, M., 2014. Classification in the presence of label noise: a survey. IEEE Transactions on Neural Networks and Learning Systems 25(5):845–869.
  • Maas, A., Rottensteiner, F., Heipke, C., 2016. Using label noise robust logistic regression for automated updating of topographic geospatial databases. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences III-7, pp. 133–140.
  • Maas, A., Rottensteiner, F., Alobeid, A., Heipke, C., 2018. Multitemporal classification under label noise based on outdated maps. Photogrammetric Engineering and Remote Sensing 84(5): 263-277.
  • Pan, S. J., Yang, Q., 2010: A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering 22(10): 1345–1359.
  • Paul, A., Rottensteiner, F., Heipke, C., 2016: Iterative re-weighted instance transfer for domain adaptation. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences III-3, 339-346.
  • Paul, A., Vogt, K., Ostermann, J., Rottensteiner, F., Heipke, C., 2018: Unsupervised source selection for domain adaptation. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences XLII-2, 845-852.
  • Vogt, K., Paul, A., Ostermann, J., Rottensteiner, F., Heipke, C., 2018: Unsupervised source selection for domain adaptation. Photogrammetric Engineering and Remote Sensing 84(5): 249-261.