Self-taught learning for land cover mapping of large areas, using multispectral remote sensing data

地球观测数据在支持决策支持系统和监测若干多边环境条约的遵守情况方面发挥着重要作用。遥感数据的土地覆被图是这方面最常用的产品，制定可行和准确的分类策略是一个正在进行的研究领域。特别是对较大区域的分类往往具有挑战性，例如，由于缺乏足够数量的训练和验证数据。本研究项目旨在为遥感数据的土地覆盖分类开发一个自学学习框架。该方法可以使用来自任意场景和不同获取日期的标记像素（即带有参考信息）和未标记像素。与半监督框架相比，未标记的数据可以包含未知和不相关的类。此外，不需要显式地对类进行建模。开发的框架将用于对来自不同研究地点的多光谱遥感数据进行分类，例如，以(i)农田、（ii）森林和（iii）城市土地利用为特征的研究地点。将评估自学框架的性能，并将其与其他方法在准确性和计算复杂性方面进行比较。

Earth Observation data play a major role in supporting decision-support systems and monitoring compliance of several multilateral environmental treaties. Land cover maps of remote sensing data are the most commonly used product in this context and the development of feasible and accurate classification strategies is an ongoing research field. Particularly the classification of larger areas is often challenging, e.g., due to the lack of adequate amount of training and validation data. This research project aims on the development of a Self-taught Learning framework for the land cover classification of remote sensing data. The approach enables the use of labeled pixels (i.e., with reference information) and unlabeled pixels from arbitrary scenes and different acquisitions dates. In contrast to semi-supervised frameworks, the unlabeled data can contain unknown and irrelevant classes. Moreover, the classes need not to be explicitly modeled. The developed framework will be used for classifying multispectral remote sensing data from different study sites, e.g., which are characterized by (i) cropland, (ii) forests and (iii) urban land use. The performance of the Self-taught learning framework will be assessed and compared to other methods in term of accuracy and computational complexity.