Low-Rank Representation Based Domain Adaptation for Classification of Remote Sensing Images

Abstract

A low-rank representation (LRR) based domain adaptation method is proposed for classification of remote sensing images. LRR achieves domain adaptation by constraining one domain can be well reconstructed by the other domain. In this paper, source data are transformed to target domain so that the transformed source domain data can be linearly reconstructed by the data of target domain. The domain distribution difference can be reduced by constraining the reconstruction matrix to be low rank. Further, we introduced a per-class maximum mean discrepancy (MMD) strategy to obtain an improved cross-domain alignment performance. The experimental results using hyperspectral remote sensing images demonstrated the effectiveness of the proposed method.