On Hyperspectral Classification in the Compressed Domain
This enables real-time on-site processing for hyperspectral imaging applications, though it is incremental as it builds on existing compressive sensing architectures.
The paper tackles hyperspectral pixel classification directly in the compressed domain without full data reconstruction, showing that using distinct measurement matrices for different pixels improves classifier accuracy and consistency.
In this paper, we study the problem of hyperspectral pixel classification based on the recently proposed architectures for compressive whisk-broom hyperspectral imagers without the need to reconstruct the complete data cube. A clear advantage of classification in the compressed domain is its suitability for real-time on-site processing of the sensed data. Moreover, it is assumed that the training process also takes place in the compressed domain, thus, isolating the classification unit from the recovery unit at the receiver's side. We show that, perhaps surprisingly, using distinct measurement matrices for different pixels results in more accuracy of the learned classifier and consistent classification performance, supporting the role of information diversity in learning.