Infrared with the M system

I've been processing super-color infrared images using the standard technique of swapping red and blue channels in post.

I was uncomfortable using the process from a more 'purist' point of view... It seemed like a gimmick invented to make the photos 'pretty' --- for me this kind of turns the corner from 'photography' into 'art.' It didn't seem scientifically logical.

As I worked with the images --- I was increasingly puzzled by the behavior of images with different filters... Knowing a little about sensor technology and having studied physics and light, I expected that as the IR-spectrum sensor recorded light deeper into the infrared, the image data would be coming almost entirely from the red-sensitive pixels on the sensor, and that green and blue channels would record almost nothing.

That isn't what happens at all. Using a 'super color filter' such as a 590 nm filter or the Hoya 25A red filter, the images out of the camera, after white balancing, do have red skies, but blue foliage. Where's the blue coming from if this is supposedly infrared? Also the blue and green histograms of the image straight out of the camera show a LOT of data, which also seemed counterintuitive.

Foliage is very dark, almost black, at the shorter IR wavelengths such as 650 nm, and bright at the deeper (longer wavelength) IR 800 nm. Shouldn't it appear red or orange in an infrared image?

Also, a clear 'blue' sky gets darker and darker as you go deeper into the infrared (less and less 'red' light). If color infrared is truly just 'sliding' the whole spectrum into the infrared --- re-white-balancing, as it were - then shouldn't the sky should 'stay' bluish with the color shift?

I believe the answer to this is that there's a lot more going on that meets the eye (pun intended).

The answer seems to lie with the color curves for digital camera sensors, of which there are several examples on the pages below:

https://www.researchgate.net/figure/Spectral-sensitivity-of-unfiltered-Canon-350D-sensor_fig1_229046541

https://maxmax.com/faq/camera-tech/spectral-response/canon-40d-study

Those curves are both for older Canon camera sensors, however I think the sensor technology hasn't changed much, and modern sensors likely have similar curves, for all camera brands.

The curves show that around 650 nm in the near-infrared, the green and blue sensitivity curves hit a 'minimum' where they are recording very little light. This makes sense as it's the 'red' and infrared part of the spectrum.

In the visible spectrum, the order of the peaks is blue, green, red, as our eyes expect and the sensor records. Color image processing shows the short wavelengths as more 'blue' and the longer wavelengths as more 'red' with a full spectrum in between.

However ---and this is key --- the green pixel sensitivity curve rises again in the deeper infrared, and has another peak around 750 nm. The blue pixel sensitivity curve also rises again and has a peak around 800 nm!

This means the green and blue pixels start functioning again in the deep infrared - perhaps taking on new identities, and doing things we might not expect!

From the spectral curves in the links above we can see In the infrared spectrum between 550 nm and 1000 nm the order of the peaks is red, green, blue! This is the OPPOSITE of the order of the peaks in visible light!

So, in the deep infrared the 'red' pixels are recording more of the shorter wavelengths than the green pixels, and the blue pixels are recording more of the longer wavelengths that the green pixels (once we normalize the color levels by 'white balancing')! In other words, the red pixels have become the infrared 'blue' and the blue pixels have become the 'infrared' red.

It would appear that channel swapping red and blue for infrared makes logical sense from a physics and 'perceptual' point of view. That's probably why red-blue color swapping results in a more 'natural-looking' image to our eyes.

The channel swap results in an image that's correctly color-shifted temperature-wise into the infrared part of the spectrum. The channel swap is needed because the relative sensitivity of the blue, green, and red pixels reverses in the IR spectrum between 550 and 1000 nm.