SPICE-HSI 1.4

GatorSense Sparsity Promoting Iterated Constrained Endmembers Toolkit - Python Implementation

SPICE

Sparsity Promoting Iterated Constrained Endmembers

---

NOTE: If the SPICE Algorithm is used in any publication or presentation, the following reference must be cited:

Zare, A.; Gader, P.; , "Sparsity Promoting Iterated Constrained Endmember Detection in Hyperspectral Imagery,"" IEEE Geoscience and Remote Sensing Letters, vol.4, no.3, pp.446-450, July 2007.

NOTE: If the code is used anywhere or in any presentation or publication, include the following reference: Caleb Robey, Taylor Glenn, Alina Zare, & Paul Gader. (2018, October 24). GatorSense/SPICE_py v1.0 (Version v1.0). Zenodo. http://doi.org/10.5281/zenodo.1470878

---

The SPICE Algorithm in Python is run using the function:

from SPICE import *

endmembers, P = SPICE(inputData, parameters)

If you would like to use the default parameters (described below), use the command:

parameters = SPICEParameters()

The inputData input is a DxM matrix of M input data points with D dimensions. Each of the M pixels has D spectral bands. Each pixel is a column vector.

This form can be achieved from a three-dimensional hyperspectral numpy array using the following commands:

import numpy as np

inputData = np.reshape(inputData, (inputData.shape[0]*inputData.shape[1], inputData.shape[2]))

The parameters input is a struct with the following fields:

parameters.u :  This is the regularization parameter that trades off between the RSS and SPT terms.
parameters.gamma : Gamma constant for the SPT term, controls the degree of sparsity desired
parameters.changeThresh : Stopping Criteria, Set this to the desired change threshold for the objective function
parameters.M : Number of Initial Endmembers
parameters.iterationCap : Maximum Number of Iterations
parameters.endmemberPruneThreshold : This is the pruning threshold for endmembers
parameters.produceDisplay : Set this to 1 if progress display is desired, 0 otherwise
parameters.initEM = None : By setting this to None, the algorithm randomly selects initial endmembers from the input data. You can also provide initial endmembers by inputting a matrix of endmembers.  Every column is one endmember.  The number of endmembers should match parameters.M.
parameters.qp_solver = 'cvxopt' : This can be 'cvxopt' or 'QPP'. cvxopt is slower, but may be used on matrices where QPP hits errors
parameters.prescale = True : Set this to True to normalize input data between 0 and 1

The parameters structure can be generated using the SPICEParameters.m function.
unmix2, which is imported with from SPICE import *, is a required helper function which unmixes the data points given the endmembers.

Note: Often the parameters must be adjusted for a particular data set. Generally, u is set to between 0.001 and 0.1 depending on noise levels in the data. gamma is generally set to a value between 1 and 10 depending on the data set. We have also found that SPICE has improved performance if the data has been normalized between 0 and 1 before running SPICE (e.g. Subtracting the minimum and then dividing by the max OR normalizing each spectrum by its L2 norm).

Running the Demo

This repository includes sample data in the form of a pickle file called "hsi_data.pkl". This contains a hyperspectral data cube, which can be analyzed by the SPICE algorithm.

To run the algorithm, use the command:

python spice_py_demo.py

The algorithm should run for no more than 40 iterations (typically much less) and will detect 4 or 5 endmembers, depending on the randomized endmember initialization parameter. After the algorithm is finished, you will be prompted to choose whether you would like to graph the output. Choose yes (Y) and a figure will appear with the proportions of each endmember in the context of the original image. Expand this window for a cleaner view of the plots. After closing this figure window, a plot of the wavelength and reflectance of each endmember will appear.

Requirements

It is recommended that you use a python virtual environment for this project using the following commands from the SPICE_py directory:

• pip install virtualenv (If you don't have the package installed already)
• python3 -m venv spice_env (Linux/Mac OSX)
• source ./spice_env/bin/activate (Linux/Mac OSX)
• python -m venv spice_env (Windows)
• spice_env\Scripts\activate.bat (Windows)

This program uses the python packages in the requirements.txt file. Those can be installed using the command:

pip install -r requirements.txt

This must also be done from the SPICE_py directory. If you run into issues, particularly on Windows, you may want to consider using conda forge and conda environments for the install.

Note: This code uses qpsolvers (see the QPP.py file) by Stephane Caron [e-mail]

Questions

If you have any questions, please contact:

Alina Zare
Electrical and Computer Engineering
University of Florida
azare [at] ufl.edu

This product is Copyright (c) 2018 All rights reserved.