MATLAB Hyperspectral Toolbox

January 16, 2025 · View on GitHub

A comprehensive MATLAB toolbox for hyperspectral image processing and analysis, providing state-of-the-art exploitation algorithms for research and educational purposes.

Overview

This toolbox was originally developed to support research for my Master's thesis "An Evaluation of Three Endmember Extraction Algorithms: ATGP, ICA-EEA, and VCA" under the advisorship of Dr. Tim Kane at Penn State's Remote Sensing and Space Systems Lab. The work has since evolved into a broader collection of hyperspectral analysis tools that continues to serve the remote sensing research community.

Key Capabilities & Examples

Noise Handling and Signal Quality

50% reflectance endmembers with various levels of noise added

Demonstration of how different SNR levels affect spectral signatures, from 10dB (high noise) to 60dB (low noise). This understanding is crucial for processing real sensor data.

Hyperspectral Data Visualization

Decorrelation Stretched Image of Moffett Field, CA Dataset

Decorrelation stretched image computed from bands 199, 126 and 51 (RGB) showing the Moffett Field, CA test site

Material Abundance Mapping

Abundance map of Sagebrush

Abundance map showing the distribution of Sagebrush across the Moffett Field, CA region

Abundance map of Lichen

Abundance map showing the distribution of Lichen across the Moffett Field, CA region

Abundance map of Oak Tree

Abundance map showing the distribution of Oak Trees across the Moffett Field, CA region

Abundance map of Antigorite

Abundance map showing the distribution of Antigorite across the Moffett Field, CA region Progressive examples of material abundance mapping showing different material concentrations overlaid on Google Earth imagery of Moffett Field, CA. These maps demonstrate the toolbox's ability to identify and quantify material distributions across a scene.

Features

The toolbox includes implementations of various hyperspectral exploitation algorithms:

Core Capabilities

  • Target Detection Algorithms
    • RX Detector
    • Matched Filter
    • ACE (Adaptive Cosine Estimator)
    • Various hybrid detectors
  • Material Abundance Mapping
    • FCLS (Fully Constrained Least Squares)
    • UCLS (Unconstrained Least Squares)
    • NNLS (Non-Negative Least Squares)
  • Spectral Unmixing
    • VCA (Vertex Component Analysis)
    • ICA-EEA (Independent Component Analysis - Endmember Extraction Algorithm)
    • ATGP (Automated Target Generation Process)
  • Data Processing
    • MNF (Minimum Noise Fraction)
    • PCA (Principal Component Analysis)
    • Spectral Angle Mapper (SAM)
    • Spectral Information Divergence (SID)
  • Visualization Tools
    • 2D/3D data conversion
    • Colormap generation
    • Advanced plotting functions

File Support

  • AVIRIS (.rfl) reading
  • ASD FieldSpec file reading
  • ENVI signature import
  • SPECPR file format support
  • Various hyperspectral data format conversions

Dependencies

  • MATLAB (version requirements TBD)
  • FastICA Toolbox - Required for certain unmixing functions
    • Add to MATLAB's path using addpath('path_to_fastica')

Installation

  1. Clone the repository:
git clone https://github.com/isaacgerg/matlabHyperspectralToolbox.git
  1. Add the toolbox to your MATLAB path:
addpath('path_to_toolbox')
  1. Install FastICA if needed for unmixing functionality

Usage

The toolbox includes several demo scripts to help you get started:

  • hyperDemo.m - General toolbox functionality
  • hyperDemo_detectors.m - Target detection algorithms
  • hyperDemo_RIT_data.m - Working with RIT dataset
  • hyperDemo_ASD_reader.m - Reading ASD FieldSpec data

Function Documentation

Target Detection Algorithms

hyperAce.m (Adaptive Cosine/Coherence Estimator)

Implements the ACE detector which normalizes the matched filter by both the background clutter energy and target energy:

ACE(x) = (s^T Σ^(-1) x)^2 / ((s^T Σ^(-1) s)(x^T Σ^(-1) x))

where s is the target signature, x is the test pixel, and Σ is the background covariance matrix.

hyperAmsd.m (Adaptive Matched Subspace Detector)

Implements a GLRT detector for signals in subspace interference:

AMSD = (x^T P_B⊥ x - x^T P_(B,S)⊥ x) / (x^T P_(B,S)⊥ x)

where P_B⊥ is the projection onto the orthogonal complement of the background subspace.

hyperAtgp.m (Automated Target Generation Process)

Implements an orthogonal projection-based endmember extraction:

  1. Finds pixel with largest magnitude
  2. Projects data onto space orthogonal to found pixel
  3. Repeats until desired number of endmembers found

hyperGlrt.m (Generalized Likelihood Ratio Test)

Implements the GLRT detector:

GLRT(x) = (s^T Σ^(-1) x)^2 / (s^T Σ^(-1) s)

Similar to ACE but without normalization by the pixel energy.

Dimensionality Reduction & Whitening

hyperMnf.m (Minimum Noise Fraction)

Implements noise-adjusted principal components:

  1. Estimates noise covariance Σn
  2. Whitens data using noise covariance
  3. Performs PCA on whitened data
MNF = eig(Σn^(-1/2) Σ Σn^(-1/2))

hyperNapc.m (Noise-Adjusted Principal Components)

Similar to MNF but uses a different noise estimation approach.

hyperWhiten.m

Implements data whitening:

x_white = Σ^(-1/2) x

where Σ^(-1/2) is computed via eigendecomposition.

Spectral Unmixing

hyperFcls.m (Fully Constrained Least Squares)

Solves the constrained optimization problem:

min ||Ax - b||^2 subject to sum(x) = 1 and x ≥ 0

where A contains endmember signatures and x contains abundances.

Note: This implementation contains a correction to the original formulation presented in "Fully Constrained Least-Squares Based Linear Unmixing" (Heinz, Chang, and Althouse, IEEE 1999).

hyperNnls.m (Non-Negative Least Squares)

Solves:

min ||Ax - b||^2 subject to x ≥ 0

hyperUcls.m (Unconstrained Least Squares)

Solves the basic least squares problem:

min ||Ax - b||^2

hyperVca.m (Vertex Component Analysis)

Implements VCA endmember extraction by:

  1. Projecting data onto random vector
  2. Finding extreme projections
  3. Iterating with orthogonal projections

Similarity Measures

hyperCorr.m

Computes Pearson correlation coefficient:

ρ = cov(x,y) / (σx σy)

hyperNormXCorr.m (Normalized Cross Correlation)

Computes normalized cross correlation:

NCC = (x^T y) / (||x|| ||y||)

hyperSam.m (Spectral Angle Mapper)

Computes spectral angle:

θ = arccos((x^T y) / (||x|| ||y||))

hyperSid.m (Spectral Information Divergence)

Computes information theoretic measure of spectral similarity using relative entropy.

Statistical Detection

hyperHud.m (Hybrid Unstructured Detector)

Combines structured (matched filter) and unstructured (RX) detectors:

HUD = αRX + (1-α)MF

hyperRxDetector.m (Reed-Xiaoli Detector)

Implements anomaly detection:

RX(x) = (x-μ)^T Σ^(-1) (x-μ)

where μ is the mean and Σ is the covariance.

Preprocessing

hyperConvexHullRemoval.m

Removes pixels inside the convex hull of selected vertices.

hyperDestreak.m

Removes vertical striping artifacts in push-broom sensors.

hyperNormalize.m

Normalizes spectra to unit length:

x_norm = x / ||x||

hyperResample.m

Resamples spectra to new wavelength positions using interpolation.

Independent Component Analysis

hyperIcaEea.m (ICA Endmember Extraction)

Uses FastICA to find statistically independent endmembers:

  1. Reduces dimensionality via PCA
  2. Applies ICA to find independent components
  3. Projects back to original space

hyperIcaComponentScores.m

Computes abundance maps from ICA components.

File I/O

hyperReadAsd.m

Reads ASD FieldSpec spectrometer files.

hyperReadAvirisRfl.m

Reads AVIRIS reflectance data.

hyperReadSpecpr.m

Reads SPECPR format spectral library files.

hyperGetEnviSignature.m

Reads spectral signatures from ENVI spectral libraries.

Visualization

hyperConvert2Colormap.m

Converts hyperspectral image to RGB using dimensionality reduction.

hyperConvert2d.m / hyperConvert3d.m

Converts between 2D (samples × bands) and 3D (rows × cols × bands) formats.

hyperImagesc.m / hyperImshow.m

Enhanced visualization of hyperspectral data cubes.

hyperMax2d.m

Finds local maxima in 2D arrays.

Performance Evaluation

hyperPpi.m (Pixel Purity Index)

Computes measure of pixel spectral purity through repeated projections.

hyperRoc.m (Receiver Operating Characteristic)

Computes ROC curves for detector performance evaluation.

Thanks to the permissive license of this toolbox, several derivative works have emerged:

PySPTools

Much of this codebase has been ported to Python in the PySPTools project, providing these algorithms to the Python community.

HyperSpectral Toolbox by David Kun

David Kun's fork of this toolbox maintains the MATLAB implementation while adding several new features.

Planned Features

Future development priorities include:

  • Joint Affine Matched Filter
  • Enhanced Matched Filter with signature statistics
  • RAF-SAM (Improved Spectral Angle Mapper)
  • ELM (Empirical Line Method) for radiance-to-reflectance conversion
  • Advanced covariance matrix inversion methods
  • Quadratic Detector
  • Additional unmixing algorithms (SMACC, AMEE, NFINDR)
  • Antonia Plaza's FastPPI
  • Joshua Broaderwater's hybrid detectors

Citation

If you use this toolbox in your research, please cite:

@Misc{matlab_hsi_toolbox,
    author    = {Isaac Gerg},
    title     = {Open Source MATLAB Hyperspectral Toolbox},
    howpublished = {\url{https://github.com/isaacgerg/matlabHyperspectralToolbox}},
    year      = {2006--2022}
}

Contributing

Derivative works must include the original citation. New contributors should add their names to the author line while maintaining the original author information.

Contact

Email: isaac.gerg@gergltd.com