data quality estimate of hyperspectral data
A python module for hyperspectral image processing
Status: Beta
Brought to you by:
tboggs
Menu
▾
▴
I guess the amount of eigenvalues after PCA says something about data quality? Is there another quality estimate you could foresee? Can I easily quantify noise level? And what about the image covariance matrix?
I wouldn't say that the number of eigenvalues (for a specified fraction of image variance) is an indicator of data quality. Rather, it is an indicator of how much correlation/redundancy there is in the spectral bands for the specific image being considered. Noise does affect principal components (PCs) but even if there were no noise at all, for realistic scenes you would still expect a large fraction of the image variance to be captures by a relatively small number of PCs.
It is also important to recognize that by throwing out PCs you are losing information. For example, if you are looking for anomalous pixels in an image, they are likely to be associated with smaller PCs. So you should always consider why you are discarding PCs (i.e., for dimensionality reduction, visualization, etc.).
It is best to have actually noise measurements from the sensor (ideally near the time of image collection) but there are also ways to empirically estimate and remove the noise. Minimum Noise Fraction (MNF) is one such technique but it is not yet implemented in SPy.
I do not understand your question about the image covariance matrix. The covariance is used in PCA (the PC transformation diagonalizes the covariance matrix). MNF also uses image and estimated noise covariance for dimensionality reduction.
Okay, thanks for the explanation. I heard about MNF yes. I'm just looking around for a method I could use to classify the data quality but I see it is much more complicated than I thought (as always ;-)