Local Structure Graph Discriminant Embedding for Hyperspectral Image Classification

Zehua Zou,Fulin Luo,Jiamin Liu,Guangyao Shi,Yufei Liu
DOI: https://doi.org/10.1109/igarss47720.2021.9554688
2021-01-01
Abstract:Graph learning is an effective technique to reduce the dimensionality of hyperspectral image (HSI) and improve classification result. However, the previous graph methods don't consider the local structure of each pixel in HSI. HSI has a complex non-linear structure, and the local structure can be regarded as a linear distribution. Therefore, we propose a local structure graph discriminant embedding (LSGDE) method to better reveal the intrinsic properties of HSI. This method constructs an intraclass and an interclass structure graphs to compact the intraclass samples and separate the interclass samples. Meanwhile, an interclass weighted scatter based on probability distribution is designed to enhance the discriminative ability of different classes. Then, a projection matrix can be obtained to map high-dimensional data into a low-dimensional space. Experiments on the HoustonU data set show that LSGDE can achieve better performance than the related DR methods.
What problem does this paper attempt to address?