Deep self-taught learning in remote sensing

Deep self-taught learning (DSTL) is a deep learning-based machine learning paradigm which originates from self-taught learning (STL, Raina et al. It is a valuable procedure to a combined exploitation of unlabeled and labeled data for classification, without the constraint that both need to contain the same classes or follow the same data distribution. Therefore, it can utilize unlabeled data from further scenes and acquisition times. The core of self-taught learning is a representation learning method such as sparse representation which learns a rich data representation which can be used, for example, as input into a classification algorithm. This chapter introduces the basic concept of STL and DSTL utilizing sparse representation and illustrates how it can be used for the classification of remote sensing images. In contrast to similar deep learning methods like autoencoders, DSTL has a higher interpretability, because the representation consists of an easily comprehensible (non)-linear combination of real data samples. This chapter explains connections to other procedures by drawing parallels and analyzing advantages and disadvantages.