Compressed spatial–spectral feature representation for hyperspectral ground classification

Abstract

The difficulty of classification tasks in hyperspectral imagery (HSI) strongly depends on the representation of spectral or spatial information. Vast amounts of approaches have been proposed to deal with spectral and spatial feature extraction, respectively. However, most of the methods neglect the inherent relationships between them. Inspired by the extreme learning machine (ELM) theory, the authors propose a new fusion-ELM framework for multiple sources representation learning and fusion. The resultant features are utilised to deal with HSI classification. With the multiple network channels and aggregation layers, the presented scheme could achieve spatial and spectral feature representations of inputs, respectively, and obtain optimal joint feature. Experimental results show that their fusion-model leads to decent improvements in classification accuracy over spectral-only, spatial–spectral-joint model and deep learning framework on two hyperspectral benchmarks.

Keywords

• hyperspectral imaging
• feature extraction
• geophysical image processing
• pattern classification
• image representation
• learning (artificial intelligence)
• image classification
• multiple sources representation learning
• resultant features
• hsi classification
• multiple network channels
• optimal joint feature
• fusion-model
• classification accuracy
• spatial–spectral-joint model
• deep learning framework
• hyperspectral benchmarks
• compressed spatial–spectral feature representation
• hyperspectral ground classification
• classification tasks
• hyperspectral imagery
• spectral information
• spatial information
• spectral feature extraction
• spatial feature extraction
• methods neglect
• inherent relationships
• extreme learning machine
• fusion-elm framework