Local Discriminative Embedding Broad Learning System With Graph Convolutional for Hyperspectral Image Classification

Abstract

Broad learning system has attracted increasing attention in the hyperspectral image (HSI) classification, due to its universal approximation capability and high efficiency. However, two main challenges remained. One is that the mapping from the original BLS input to the mapped feature (MF) is linear, which is difficult to fully represent the complex spatial-spectral features of HSI. The other is that BLS fails to explore the local geometric structure relationship between samples within HSI. To overcome the limitations mentioned above, we propose a local discriminative embedding broad learning system with graph convolutional (GDEBLS). To address the first challenge, GDEBLS utilizes the graph convolution operation to aggregate the node information in the adjacent graph to learn the context and obtain the rich nonlinear spatial-spectral features in HSI. To deal with the second challenge, our method utilizes a neighborhood selection approach based on manifold structure to calculate the true distances between samples in the manifold space, overcoming the limitations of Euclidean distance measurement. Next, We introduce local manifold structure and discriminative information into BLS. The experimental results show that the proposed method significantly surpasses other state-of-the-art methods.