Skip to main content

Hyperspectral Image Classification and Dimensionality reduction: An Orthogonal Subspace Projection Approach

Citations: 1,471 | Paper Published: 01 Jul 1994 | Institutions: University of Maryland, Baltimore | Authors: J.C. Harsanyi, Chein-I Chang

Abstract in 2 Lines

The paper describes a technique for processing hyperspectral imagery that achieves two goals: detecting and classifying materials in each pixel, while reducing data volume. The approach involves projecting pixels onto an orthogonal subspace to suppress undesired spectral signatures, then maximizing the signal-to-noise ratio of the signature of interest.

Contributions​

The contribution of this paper is a technique for processing hyperspectral imagery that simultaneously reduces data dimensionality and produces new images highlighting the presence of each signature of interest. The approach involves projecting pixels onto an orthogonal subspace to suppress undesired spectral signatures, then maximizing the signal-to-noise ratio of the signature(s) being analyzed. This method can be applied to both spectrally pure and mixed pixels in order to classify materials within each pixel while reducing overall data volume.

Practical Implications​

  • The technique described in the paper can be used to process hyperspectral imagery, which has applications in fields such as remote sensing and medical imaging.
  • By reducing data volume while maintaining critical information, the approach allows for more efficient processing and analysis by human experts.
  • The ability to simultaneously detect and classify materials within each pixel makes it easier to identify specific features or anomalies within an image.
  • This method is applicable not only for spectrally pure pixels but also mixed ones.

Methods​

Methodology

The method used in this paper involves projecting each pixel vector onto a subspace that is orthogonal to undesired spectral signatures. This process nulls the interfering signatures and maximizes signal-to-noise ratio for the signature of interest, resulting in a single component image representing classification for that specific material. The approach can be extended to multiple k-signatures of interest simultaneously while reducing data dimensionality.

Results​

Performance Analysis

The paper describes the application of the orthogonal subspace projection (OSP) technique to a subsection of an AVIRIS scene covering Lunar Crater Volcanic Field in Northern Nye County, NV. The authors state that extensive field work has been done in this area and previous spectral mixture analysis models have been applied to this dataset as well. They also mention that atmospheric water bands and low SNR bands were removed from the data before applying their OSP method which reduced image cube dimensions from 224 to 158 bands.

Limitations​

The paper does not explicitly mention any limitations of the proposed technique. However, it is important to note that while this method can reduce data dimensionality and classify materials within each pixel efficiently, there may be other factors such as computational complexity or accuracy trade-offs that need to be considered when applying this approach in practice. Additionally, further research could explore how well this technique performs on different types of hyperspectral imagery datasets with varying levels of noise or interference from environmental factors.

Table of Contents