Hyperspectral Remote Sensing Notes

Hyperspectral Image Analysis

  • Covers spectral matching, pixel-based, and sub-pixel-based image classification.

Spectral Matching

  • Pure spectral signature of a material is called an endmember.
  • Endmembers are crucial for distinguishing earth materials.
  • Goal: compare endmembers with reference spectra to enable image classification at the pixel level.
  • Image endmembers are extracted using algorithms like PPI, FIPPI, and N-FINDR.
  • Reference spectra are measured manually or obtained from spectral libraries (e.g., USGS, ASTER).
  • Spectral matching algorithms assign pixels to classes based on spectral similarity.
  • Challenges: spectra disagreement due to different collecting conditions.

Pixel-Based Image Classification

  • Uses high-quality data for spectral sensing on a pixel-by-pixel basis.
  • Does not consider the mixed pixel effect.
  • Example: Spectral Angle Mapper (SAM).

Spectral Angle Mapper (SAM)

  • Assumes each pixel represents one ground cover material.
  • Measures spectral similarity by calculating the angle between image pixel spectra and reference spectra.
  • Smaller angles indicate closer matches.
  • A threshold can be defined to exclude pixels based on the angle.
  • Disadvantage: does not account for sub-pixel values; problematic for heterogeneous surfaces.

Sub-Pixel-Based Image Classification

  • Addresses heterogeneous surfaces and mixed pixels.
  • Spectral signature of a mixed pixel is a combination of endmember signatures.
  • Spectral Unmixing (SU): extracts spectral signatures and quantifies their spatial distribution.
  • Decomposes pixel spectrum into constituent spectra (endmembers) and abundances.
  • Involves endmember extraction and abundance map estimation.
  • Endmember extraction algorithms: PPI, FIPPI, N-FINDR.
  • Abundance maps represent endmember distribution.