Hyperspectral Unmixing

Some of files are present in Git LFS

How to Clone?

git-lfs url

What is Hyperspectral Unmixing?

The images captured by the HSI camera, will give us the image data in the form of a hyper-cube. In the hypercube, the pixels may not be a single pixel, it can be a combination of multiple spectra. Such pixels are known as mixed-pixels or hyper-pixels. We need to decompose these hyper-pixels into endmember signatures(pure pixel signatures) and corresponding abundances. The process of decomposition of hyper-pixels is known as Hyperspectral Unmixing. We have here used the Multiplicative Update NMF method for Unmixing.

Algorithm for Hyperspectral Unmixing using Non-negative Matrix Factorization using Multiplicative Update method

procedureHSU(A):
	Initialize W and H
	Define Convergence Tolerance
	StopIter ← maxIterations
	counter ← 0
	while !Converged do
		W_old ← W
		H_old ← H
		W ← W · (A x H_old.T)/(W_old x H_old x H_old.T)
		H ← H · (W_old.T x A)/(W_old.T x W_old x H)
		counter ← counter + 1
		Check Convergence
		if counter == StopIter then
			break
		end if
	end while
	return W,H
end procedure

How to run the code

This code will run for Indian Pines, SalinasA, and Pavia University hyperspectral datasets. Download the corrected version for Indian Pines and SalinasA. You can download them from this link: Hyperspectral Remote Sensing Scenes. After downloading them put it into the dataset folder.

The main file from where you can run then run the project is NMF_run.py. Depending on the dataset on which you want to run, uncomment the 3 lines of code in which the name of the desired dataset is written and keep the other two commented and then run the file. By default we have uncommented the Indian pines code.

As an output you will get two graphs and some console output. Console output will show the final unmixed abundance and spectral endmember matrices and a list of error and time taken per outer iteration. One grpah will show the error trend of NMF for the dataset and another will show the plot of unmixed spectral endmembers according to their calculated value for each spectral bands. (NMF will not give a 100% accurate graph of unmixed spectral endmembers)

NMF_tools.py contains the helper functions which are used in NMF_run.py. data folder will contain the list of abundance matrices, endmembers matrices, errors, and time taken all for each outer iteration so that you don't have to run the whole code to get the matrices again.