Deep canonical correlation analysis for hyperspectral image classification

Multi-view learning (MVL) is a technique which utilizes multiple views of data simultaneously during training to learn more expressive representations. Multi-view learning has been gaining a large amount of interest in various machine learning applications recently. In this paper, we focus on learning representations prior to classification using multi-view learning via deep canonical correlation analysis (DCCA) in hyperspectral image processing. We propose a classification framework including a proposed view generation approach. The motivation of our proposed view generation approach is to fuse spatial and spectral information. The performance of our proposed view generation approach is compared with the other view generation methods in the literature; namely the uniform band slicing and correlation-partition-based clustering. To evaluate the effectiveness of the proposed approach, we performed experiments on two commonly used hyperspectral image datasets. Experimental results based on two hyperspectral image datasets demonstrate that the proposed classification framework provides satisfactory classification performances.