GRAPH CONVOLUTIONAL NEURAL NETWORKS FOR HYPERSPECTRAL DATA CLASSIFICATION

摘要

Graph based manifold learning and embedding techniques have been very successful at representing high dimensional hyperspectral data in lower dimensions for visualization and classification. Graph based convolutional neural networks (GCNs) have been recently developed for applications on high-dimensional irregular domains represented by graphs, such as citation networks. In this paper, we demonstrate a framework that can leverage GCNs to effectively represent data residing on smooth manifolds, such as reflectance spectra of hyperspectral image pixels. We also propose a robust spatial-spectral semi-supervised adjacency matrix that learns the underlying manifold structure of the data using a limited amount of labeled spectra and a large amount of unlabeled spectra. Classification performance with a benchmark hyperspectral image analysis dataset is also provided that demonstrates the efficacy of this approach.