Semisupervised Dual-Dictionary Learning for Heterogeneous Transfer Learning on Cross-Scene Hyperspectral Images

Lack of labeled training samples is a big challenge for hyperspectral image (HSI) classification. In recent years, cross-scene classification has become a new research topic. In cross-scene classification, two closely related HSI scenes are considered, one contains adequate labeled samples, namely source scene, while the other one contains only a few labeled samples, namely target scene. The goal of cross-scene classification is utilizing the labeled samples in source scene to benefit the classification in target scene. In most cases, different HSIs are imaged by different sensors, leading to different feature dimensions (numbers of bands) in different scenes. In this situation, heterogeneous transfer learning is demanded. In this article, we propose a heterogeneous transfer learning algorithm namely semisupervised dual-dictionary nonnegative matrix factorization (SS-DDNMF). SS-DDNMF consists of two contributions. 1) Dual-dictionary nonnegative matrix factorization (DDNMF): DDNMF trains two dictionaries for source and target scenes, respectively, aiming at projecting the source and target features to a shared low-dimensional subspace, eliminating the difference between feature spaces. In DDNMF, within-scene and cross-scene graphs are built to maintain the similarities between pixels. 2) Semisupervised learning for target scene: as the limited number of labeled pixels in target scene will affect the graph building of DDNMF, semisupervised learning is adopted in target scene. In details, superpixel segmentation is adopted to generate pseudolabels for some unlabeled pixels, thus more “labeled” pixels can be considered for building better graphs. The effectiveness of SS-DDNMF is verified by experiments on cross-scene HSIs.