Domain adaptation based on deep denoising auto-encoders for classification of remote sensing images

Abstract

This paper investigates the effectiveness of deep learning (DL) for domain adaptation (DA) problems in the classification of remote sensing images to generate land-cover maps. To this end, we introduce two different DL architectures: 1) single-stage domain adaptation (SS-DA) architecture; and 2) hierarchal domain adaptation (H-DA) architecture. Both architectures require that a reliable training set is available only for one of the images (i.e., the source domain) from a previous analysis, whereas it is not for another image to be classified (i.e., the target domain). To classify the target domain image, the proposed architectures aim to learn a shared feature representation that is invariant across the source and target domains in a completely unsupervised fashion. To this end, both architectures are defined based on the stacked denoising auto-encoders (SDAEs) due to their high capability to define high-level feature representations. The SS-DA architecture leads to a common feature space by: 1) initially unifying the samples in source and target domains; and 2) then feeding them simultaneously into the SDAE. To further increase the robustness of the shared representations, the H-DA employs: 1) two SDAEs for learning independently the high level representations of source and target domains; and 2) a consensus SDAE to learn the domain invariant high-level features. After obtaining the domain invariant features through proposed architectures, the classifier is trained by the domain invariant labeled samples of the source domain, and then the domain invariant samples of the target domain are classified to generate the related classification map. Experimental results obtained for the classification of very high resolution images confirm the effectiveness of the proposed DL architectures.