Jump to content
UltravioletPhotography

Petroglyphs in Desert Varnish


Andrea B.

Recommended Posts

Valley of Fire State Park, Overton, Nevada, USA

25 February 2016

Petroglyphs on Red Sandstone

 

Comment:

Dark desert varnish consists of a thin layer of clay with black maganese oxide deposited on the desert rocks by bacteria. In the Infrared photograph the desert varnish is strikingly dark against the IR-reflective unvarnished areas. In UV, however, there is much less contrast between the desert varnish and the sandstone.

 

The petroglyphs in Valley of Fire were carved by the ancient Anasazi, a group of prehistoric Native Americans, now known as the Ancestral Puebloans.

 

See also: http://www.ultraviol...-in-the-desert/

 

Equipment [Nikon D600-broadband + Coastal Optics 60/4.0 UV-Vis-IR Apo Macro]

 

Visible [f/11 for 1/250" @ ISO-100 with Baader UV/IR-Cut in Sunlight]

petroglyphs_visible_20160225valleyOfFireStateParkNV_45200pn.jpg

 

Ultraviolet [f/11 for 2.5" @ ISO-100 with BaaderU in Sunlight]

petroglyphs_uvBaader_20160225valleyOfFireStateParkNV_45214pn.jpg

 

Infrared [f/11 for 1/30" @ ISO-100 with B+W 093 in Sunlight]

petroglyphs_093IR_20160225valleyOfFireStateParkNV_45228pn.jpg

Link to comment

Ooh. When you get that kind of response, the Independent Component Analysis lets you extract the signal from the background pretty well. Here are the results for this case (using R,G,B, UV, and IR as the 5 channels). ICA produces the same number of output channels as inputs, but I've thrown away the irrelevant ones (for example corresponding to alignment errors). The remaining two are the rock background and the petroglyphs. (Andrea, apologies for deleting the watermark, but the ICA routine gets screwed up by them, so I had to crop it. Nobody will mistake these for your original images, however!)

 

post-94-0-89477600-1484632215.jpg

 

...and the rock without the petroglyphs!

post-94-0-82986600-1484632209.jpg

Link to comment
I'd like to learn more about ICA as well (and also PCA). Generally speaking, I'm able to produce practically identical results using just contrast adjustments on the source images. Could you share what info ICA provides? Its killing me that I'm missing something useful here! I understand the ICA theory in principal, but (regarding image/photo applications) I haven't seen a good case yet that differs significantly from basic adjustments made in an image editing program (e.g., contrast, channel toning, etc).
Link to comment
Mark, ICA literally is just adding and subtracting channels, but it does it in a way that optimizes a certain function of the histogram of each of the combinations of the channels for maximum statistical independence from the other combinations. But at the end of the day it is still just adding and subtracting channels. So of course you could add and subtract the channels by hand to produce the same result! But it would not necessarily by optimal (unless you have a very good eye). What you can't do is get there by contrast adjustment alone — because that operation doesn't combine different channels.
Link to comment
I think this could be very useful in a calibrated workspace. That is, in a controlled lighting situation where each component/contribution to an image can be a known, then that could be used to calibrate ICA extraction of wanted, or unwanted, data. This could be used, for example, as a convenient method for dealing with IR contamination in a light source where filtration is not a viable option (e.g., fluorescent UV bulbs). Because IR and UV data may in fact compete in parts of an image this might not be a perfect solution, but it could certainly help. I wonder if I could use this to extract UVIVF from lint in so many of my UVIVF images. Oh, wouldn't that be nice!
Link to comment

OlDoinyo, PCA is an eigenvector decomposition of the correlation matrix. ICA starts by doing PCA as a first step, and then it looks for another orthogonal matrix that makes each component of the transformed space as non-gaussian as possible. For the case where you start with an RGB color space, each component of the result will be a weighted sum of R, G and B. The histogram of that weighted sum will maximize some measure of the non-gaussianity (for example, the kurtosis). The resulting orthogonal matrix will not usually be the matrix of eigenvectors, however.

 

The intuition is that the central limit theorem says that if you mix a bunch of signals together, the histogram of the resulting mixture gets more gaussian. So if you want to separate the signals from each other, then you want to maximize NON-gaussianity.

 

I promise I'll write all this up for you guys soon. In the mean time, this tutorial is quite well-written.

Link to comment

Please sign in to comment

You will be able to leave a comment after signing in



Sign In Now
×
×
  • Create New...