Learning By Evolving Nonlinear Diffusion for Active Learning on Hyperspectral Images

A method for active learning of hyperspectral images (HSI) is proposed, which leverages time-evolving diffusion processes on graphs to determine query points and propagate their labels to the full data set. Initially, a diffusion process is defined on a graph in which each HSI pixel is a node, with edge connection strength scaling with the distance between pixels in the spectral domain. At each stage of the iterative sampling process, queried labels are used to update the underlying diffusion matrix by weakening edges between points in different classes, thereby infusing the intrinsic representation encoded in the diffusion matrix with revealed label information. The proposed method, Learning by Evolving Nonlinear Diffusion (LEND), combines robust performance with the mathematical tractability of diffusion geometry, leading to superior labeling accuracy with fewer labeled samples compared to a baseline in which the underlying matrix is not iteratively updated. Experiments on the Salinas A HSI demonstrate the effectiveness and efficiency of LEND.