Self-taught Learning für die Klassifikation von Landbedeckungen großer Gebiete unter Verwendung von multispektralen Fernerkundungsdaten
Zusammenfassung der Projektergebnisse
The goal of this project was the development and evaluation of a self-taught learning framework for landcover classification, using remote sensing image data. Self-taught learning is an increasingly used paradigm, which however, is mostly by applied to low-dimensional (i.e. gray-valued or RGB) image data. The approach enables the use of labeled pixels (i.e., with reference information) and unlabeled pixels from arbitrary scenes and different acquisitions dates. Therefore, this approach is advantageous to semi-supervised frameworks, since the unlabeled data can contain unknown and irrelevant classes. Moreover, the classes need not to be explicitly modeled. Generally, our investigations showed that the self-taught learning framework is able to exploit the information from unlabeled data and it can learn a representation of the data which makes the classification task more easy and applicable to large-scale learning. We developed an efficient strategy to select the most relevant unlabeled samples to be used in this framework. Furthermore, we have done considerable research in the area of sparse representation-based models, which is the most commonly used approach to self-taught learning. This comprises a strategy to integrate prior information into this framework, where the prior information of our interest was spatial information and geometric information. We conducted experiments with multi-spectral and hyperspectral data which were acquired over large areas with satellite sensors and, in order to demonstrate its applicability to related applications. In summary, self-taught learned turned out to be a versatile tool which can be combined with further strategies such as structured learning, and which outperforms classical and state-of-the-art classification techniques such as supervised learning.
Projektbezogene Publikationen (Auswahl)
- (2014). Shapelet-based sparse image representation for landcover classification of hyperspectral data. IAPR Workshop on Pattern Recognition in Remote Sensing (PRRS)
Roscher, R. & Waske, B.
(Siehe online unter https://doi.org/10.1109/PRRS.2014.6914277) - (2015). Landcover classification with selftaught learning on archetypal dictionaries. IEEE International Geoscience and Remote Sensing Symposium (IGARSS), 2358-2361 [IEEE GRSS Symposium Award for the best paper]
Roscher, R., Römer, C., Waske, B., & Plümer, L.
(Siehe online unter https://doi.org/10.1109/IGARSS.2015.7326282) - (2016). Discriminative archetypal self-taught learning for multispectral landcover classification. IAPR Workshop on Pattern Recognition in Remote Sensing (PRRS)
Roscher, R., Wenzel, S., & Waske, B.
(Siehe online unter https://doi.org/10.1109/PRRS.2016.7867022) - (2016). Shapelet-based sparse representation for landcover classification of hyperspectral images. IEEE Transactions on Geoscience and Remote Sensing, 54(3), 1623-1634
Roscher, R., & Waske, B.
(Siehe online unter https://doi.org/10.1109/TGRS.2015.2484619)
