Semi-Supervised Hyperspectral Image Classification Based on Label Propagation via Selected Path

Most graph-based semi-supervised classification methods do not perform well in hyperspectral image classification tasks due to their high complexity and other limitations. This paper proposes a label propagation semi-supervised classification algorithm which uses a few selected important paths to considerably reduce computational costs. It establishes that the most important paths for label propagation are minimum cost paths. It proves that minimum cost paths exist in minimum cost trees (MCT), and proposes a method based on a variant minimum spanning tree (MST) combined with priority queue to construct MCTs. The algorithm propagates labels from unlabeled nodes to labeled ones, a unique way different from any other studies where propagation is in the opposite direction, which brings about several clear advantages. These include that only one propagation path is required for each unlabeled node, improving both timing and memory performance. It also helps to solve a problem posed by sparse graphs where some image pixels cannot be classified, a situation which is especially problematic in large-scale image classification. The proposed method has the advantages of linear computational complexity, is independent of data dimension, has fewer parameters and is insensitive to the values of parameters. Moreover, it does not need large numbers of labelled pixels nor complex training processes. Experiments on hyperspectral images have shown that, compared with several existing algorithms, the proposed method achieves better performance in less time. The paper addresses some fundamental issues regarding propagating labels in graph based semi-supervised classifications. Due to the simplicity and the fast speed of the algorithm, it is also suitable to be integrated into both state-of-the-art and future hyperspectral image classification frameworks which have a label propagation stage.