Cross-Scene Relationship Mining with Learning Graph Net for Hyperspectral Image Classification

• Published in: 2021 2nd International Conference on Artificial Intelligence and Computer Engineering (ICAICE) pp. 519 - 524
• Main Authors: Chen, Junbin, Ye, Minchao, Lu, Huijuan, Lei, Ling
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.11.2021
• Subjects: Artificial intelligence; Computer architecture; Convolutional neural networks; cross-domain variational autoencoder; cross-scene classification; heterogeneous transfer learning; Hyperspectral image; Hyperspectral imaging; Image classification; learning graph net; Manifolds; Transfer learning
• DOI: 10.1109/ICAICE54393.2021.00106

The problem of hyperspectral image (HSI) classification is usually accompanied by the problem of high dimension and few samples, that is, the high-dimensional-small-sample-size problem. In recent years, transfer learning has been widely used to solve this problem. In the cross-scene HSI classification, we consider a scene with a rich number of samples (called source scene) and a scene with a small number of samples (called target scene). The idea of transfer learning is to transfer the knowledge contained in the rich samples of source scene to target scene. Many HSI classification methods assume that two scenes come from the same feature space. However, the facts are often unsatisfactory, and the two scenes are likely to come from different feature spaces. In this case, we proposed a heterogeneous transfer learning method named cross-domain variational autoencoder (CDVAE), which achieved good results. But the imperfection is that CDVAE cannot use unlabeled samples on target scene to help classification. Therefore, on this basis, we have proposed a learning graph net (LGnet) of using convolutional neural networks (CNN) and graph to learn the relationship between cross-scene samples, so as to use the potential information of unlabeled samples. Then, a new method cross-domain variational autoencoder with learned graph (CDVAE-LG) was proposed by combining LGnet with CDVAE. The experimental results show that CDVAE-LG can effectively learn the information between cross-scene samples and help classification.