Domain Adaptation for Robust Classification of Disparate Hyperspectral Images

A major difficulty of remote sensing image classification comes from the lack of quality training data. In this paper, we propose a domain adaptation algorithm to alleviate this problem by leveraging labeled data from an auxiliary data source. The proposed approach aims to align the class distributions of two domains while enhancing their discriminability. Separate transformations are used to project data from the source and target domains into a latent space where the ratio of within-class distance to between-class distance is minimized. A probabilistic framework is presented such that the transformation for each domain is iteratively optimized through an alternating maximizing algorithm. The proposed method is optimized with a distance-based objective function that ensures preservation of stochastic neighborhood and discriminative information of both source and target data in the latent space. The proposed method can be used in original or transformed spaces, and to illustrate this, we provide results of this method in a space generated by the wavelet scattering transform, which imparts further robustness. Additionally, a challenging hyperspectral dataset is introduced for evaluating domain adaptation algorithms. We investigate the proposed method under three domain adaptation scenarios, and experimental results with real-word hyperspectral datasets demonstrate the efficacy of the proposed method compared to conventional domain adaptation approaches.