Endmember Explorer

Overview

Endmember Explorer identifies representative spectral components in a hyperspectral image and estimates their spatial abundance.

An endmember is a spectral signature representing a relatively pure or characteristic component of the hyperspectral dataset.

The tool combines endmember-number estimation, dimensionality reduction, N-FINDR extraction, abundance estimation, spectral visualization, and abundance mapping.

The analysis consists of four principal steps:

1. Determine the number of endmembers.

2. Reduce the spectral dimensionality.

3. Extract representative endmember spectra using N-FINDR.

4. Estimate the abundance of each endmember throughout the image.

The resulting spectra and abundance maps can then be inspected and exported.

The number of endmembers determines how many spectral components IDCubePro attempts to extract from the dataset.

When Auto estimate is selected, IDCubePro estimates the number of spectrally distinct components automatically using the NWHFC/HFC approach.

This is recommended as the initial setting when the expected number of endmembers is not known.

The automatically estimated number should be treated as a data-driven estimate rather than an absolute biological or chemical truth.

The estimate can be affected by image noise, preprocessing, spectral redundancy, mixed pixels, and weak spectral components.

Uncheck Auto estimate to specify the desired number of endmembers manually.

Manual selection is useful when prior experimental knowledge suggests the expected number of components or when the automatic estimate is not appropriate for the dataset.

Using too few endmembers can merge distinct spectral populations.

Using too many endmembers can split similar populations or extract noise and minor spectral variations as separate components.

2. Extraction Method - N-findr

N-FINDR is used to extract the endmember spectra.

N-FINDR searches for spectra that form a simplex with large volume in reduced spectral space.

Conceptually, these spectra represent extreme spectral signatures within the dataset and are therefore candidates for relatively pure spectral components.

N-FINDR is widely used for hyperspectral endmember extraction but does not guarantee that every extracted component corresponds to a chemically or biologically pure material.

Before N-FINDR extraction, the hyperspectral data are represented in a lower-dimensional spectral space.

Two reduction methods are available:

Mnf - Minimum Noise Fraction

MNF is the recommended default for many hyperspectral datasets.

MNF attempts to organize spectral information according to signal-to-noise characteristics, helping separate informative spectral structure from noise-dominated components.

This can be particularly useful for hyperspectral images containing substantial detector or measurement noise.

Pca - Principal Component Analysis

PCA reduces dimensionality by identifying directions of maximum variance in the spectral dataset.

The first principal components contain the largest fractions of total spectral variance.

PCA is computationally efficient and useful when variance-based dimensionality reduction is appropriate.

Unlike MNF, standard PCA does not explicitly model the noise structure of the dataset.

For routine hyperspectral endmember extraction, MNF is generally a useful starting choice.

PCA can be used for comparison or when conventional variance-based dimensionality reduction is preferred.

Results may differ because the two methods emphasize different characteristics of the spectral dataset.

After the endmember spectra are extracted, IDCubePro estimates how strongly each endmember contributes to every image pixel.

The result is an abundance map for each endmember.

Fast LS performs rapid least-squares estimation of endmember contributions.

Negative abundance estimates are removed in the current implementation.

Fast LS is computationally efficient and is recommended for initial exploration and rapid visualization.

Fcls - Fully Constrained Least Squares

FCLS performs constrained spectral unmixing.

It is intended for analyses in which physically constrained abundance estimates are preferred.

FCLS generally requires more computation than Fast LS but can provide more interpretable abundance estimates for linear mixture models.

Use Fast LS for rapid exploratory analysis.

Use FCLS when constrained abundance estimation is important for the final interpretation.

It can also be useful to compare Fast LS and FCLS results to determine whether the spatial patterns are robust to the abundance-estimation method.

Click Run Analysis after selecting the desired settings.

1. Estimate or read the requested number of endmembers.

2. Perform the selected dimensionality reduction.

3. Extract endmember spectra using N-FINDR.

4. Calculate abundance maps.

5. Display the extracted spectra.

6. Generate abundance-map thumbnails.

7. Configure the RGB abundance composer.

The Endmember Spectra panel displays the spectral signature of each extracted endmember.

Each curve represents one endmember.

Inspect the spectral shape, peaks, valleys, slopes, and wavelength-dependent features when interpreting the extracted components.

Endmember spectra may subsequently be compared with known reference spectra or spectral libraries.

Each abundance map shows the spatial distribution of one extracted endmember.

Brighter or higher-valued regions indicate stronger estimated contribution of that endmember according to the selected abundance model.

Click an abundance-map thumbnail to display the corresponding map in the main Image / Abundance Map panel.

If more endmembers are present than can be displayed simultaneously, use the slider below the thumbnails to browse through the maps.

The RGB Abundance Composer combines three abundance maps into one false-color image.

Select an endmember for each channel:

Click Update RGB to generate the composite.

For example, assigning EM 1 to red, EM 2 to green, and EM 3 to blue allows the spatial distributions of three spectral components to be compared simultaneously.

Areas containing mixtures of the selected components appear as combinations of the corresponding colors.

Click Original to return the main display to the hyperspectral dataset preview.

The preview is provided for spatial orientation and comparison with the abundance maps.

Click Export Spectra after completing the analysis.

The extracted endmember spectra can be exported as:

Excel and CSV exports contain the wavelength column followed by one column for each extracted endmember.

The MATLAB export stores the wavelength vector and endmember spectral matrix.

An extracted endmember is a mathematical spectral component identified from the hyperspectral dataset.

An endmember should not automatically be interpreted as a specific chemical, material, tissue, cell type, or molecular species.

Identification should be supported by additional evidence such as known spectral references, spectral libraries, experimental controls, spatial context, or independent analytical measurements.

Hyperspectral pixels frequently contain signals from more than one component.

Abundance mapping attempts to estimate the relative contribution of the extracted endmembers within these mixed pixels.

The validity of this interpretation depends on the assumptions of the spectral mixing model and the quality of the extracted endmembers.

Endmember extraction can be strongly affected by preprocessing.

Depending on the dataset, useful preprocessing may include:

• Reference or reflectance correction
• Removal of noisy wavelength regions
• Removal of invalid or saturated pixels

Use preprocessing appropriate for the imaging system and experimental objective.

Recommended Starting Workflow

For a new hyperspectral dataset, a useful starting workflow is:

1. Load the hyperspectral dataset.

2. Apply appropriate calibration and preprocessing.

4. Keep Auto estimate (NWHFC) selected.

8. Inspect the extracted endmember spectra.

9. Inspect the individual abundance maps.

10. Use RGB Composer to compare major spatial components.

11. If needed, repeat with a manually selected endmember count.

12. Compare Fast LS with FCLS for the final analysis.

13. Export the endmember spectra.

Too many endmembers are detected The automatic estimate may be identifying noise or small spectral variations as separate components. Consider preprocessing the dataset or using a smaller manual endmember count.

Too few endmembers are detected Weak or highly similar spectral populations may not be separated by the automatic estimate. Try a larger manual endmember count and inspect the resulting spectra and maps.

Endmember spectra appear noisy Inspect the raw dataset and consider removing noisy spectral regions or applying appropriate spectral preprocessing before extraction.

Noise in the original spectra, inappropriate endmember selection, background pixels, or model mismatch can produce noisy abundance maps.

An endmember appears in unexpected regions Inspect its spectrum and compare it with the original image. The component may represent background, illumination variation, scattering, noise, or a spectral mixture rather than the expected target.

MNF and PCA give different endmembers This is expected. MNF and PCA construct different reduced spectral representations and can therefore lead N-FINDR to identify different extreme spectra.

Fast LS and FCLS give different abundance maps This is also expected because the two methods impose different constraints on the spectral unmixing solution.

Important

Endmember extraction and spectral unmixing are exploratory and quantitative analytical methods.

The quality of the result depends on image calibration, spectral quality, preprocessing, dimensionality reduction, the number of selected endmembers, and the validity of the spectral mixing model.

Extracted endmembers and abundance maps should therefore be interpreted in the context of the experiment and independently validated whenever possible.