Hyperspectral image classification with limited number of labeled pixels is a challenging task. In this paper, we propose a bilayer graph-based learning fr a mework to address this problem. For graph-based classification, how to establish the