The need for unfiltered luminance channels
Talking about exposure times, I argue that a ratio for LRGB of 1:1:1:1 should be achieved. If the Luminance channel is much deeper, then it will contain faint levels for which no colour information is available anymore. Hence, these parts will appear with little or no colour, i.e. grey.
Before I proceed I will define the following terms:
Squared brackets [ ] denote the luminance channel.
Given that a luminance filter with UV- and IR-cutoff covers about the same bandpass as the RGB filters together, the RGB filters together should record about as much light as the L filter alone, provided the exposure time ratio is about 1:1:1:1.
In that case, i.e. an approximately equal amount of time is spent for each filter, the following questions arise:
To find the answers to these 3 questions, I selected exposures from a deeper data set taken under excellent conditions with my 12.5" f/5.1 Newtonian and a ST-10XME. The RGB filters (transmission) are standard SBIG filters, the L filter (transmission) is from Astronomik and its bandbass is from 390 nm to 670 nm.
| Filter | Exposure time [s] |
| L | 12x600 |
| R | 6x1200 |
| G | 6x1200 |
| B | 6x1200 |
The images in each filter were registered and resampled to the same astrometric grid using a Lanczos3 kernel, and stacked using a weighted mean coaddition. The coadded images were then normalised to an exposure time of 1000 seconds for better comparison.
Synthetic luminance channel
I created a straight RGB = R + G + B (without any weighting) to obtain a synthetic luminance channel. This is the closest match to a real luminance exposure. If the stacked transmission curves of RGB yielded the transmission curve of L, then the amount of sky noise would be identical in both the RGB and the L.
Differences occur because the combined transmission of RGB is a bit less than that of L, and because there are 18 readouts for RGB as compared to 12 for L. However, the exposures are reasonably well sky dominated so that the readout noise can be neglected.
In the same fashion, I created a deep luminance image by adding up L, R, G and B without weighting.
In order to directly compare the noise between [RGB], [L] and [LRGB], these three images were rescaled such that objects therein have the same brightness as in the [L] image. The properties and results are then as follows (noise measured in an area free of any objects from some 10000 pixels):
| Filter | Noise level | Noise reduction by a factor of |
Effective exposure time increased by a factor of |
| R | 6.1 | -- | -- |
| G | 6.2 | -- | -- |
| B | 4.5 | -- | -- |
| [RGB] | 14.1 (not rescaled: 11.7) | 1.00 | 1.00 |
| [L] | 10.2 (not rescaled: 10.2) | 1.38 wrt. [RGB] | 1.91 wrt. [RGB] |
| [LRGB] | 7.6 (not rescaled: 15.7) | 1.86 wrt. [RGB] | 3.44 wrt. [RGB] |
Impact on the luminance images
The black&white images below illustrate the increase of S/N, going from [RGB] (left), to a normal luminance exposure [L] (middle), to the stack of both, [LRGB] (right). They are "as is", i.e. unmanipulated.
The images are noise normalised, i.e. the noise amplitude in the sky background is the same.
Thus the increase in image depth can be seen as a brightening of the faintest objects from the left to the right.
Some random background galaxies
A cluster of galaxies at intermediate redshift (z = 0.3...0.4)
The lobe of a galaxy to illustrate the impact on diffuse objects
The [L] image offers a small but significant improvement over the synthetic [RGB] image.
The composed [LRGB] image is deeper than the [L].
Impact on colour images
The following images illustrate the overall improvement. The luminance layered colour images in the rightmost three columns were created using the Lab technique.
Contrary to the examples above, these images are flux normalised so that objects have the same brightness in all exposures, while the noise is getting smaller from the left to the right.
Particularily noteworthy - and most surprising - is the big improvement when going from a straight RGB to a [RGB]-RGB. That is, the quality of the image can be improved significantly when combining the R, G and B channels and using the sum as a fake luminance image.
The [L]-RGB is also clearly better than the [RGB]-RGB. There are two reasons this. First, the L filter provides independent information, hence there is no 'self-enhancement' like in the [RGB]-RGB. Second, the [L] image is a bit deeper than the [RGB].
Lastly, the [LRGB]-RGB is the best, as all available information is combined.
No manipulative processing (background smoothing or sharpening or the like) has been done to these images.
The RGB to the left and the three different luminance-layered RGBs were created exactly as described here, that is creating the RGB, converting it to Lab format, replacing the lightness channel with the luminance image, and converting everything back to RGB format.
Unfortunately, I do not know how exactly the noise propagates in these colour images during this conversion process as the internal Photoshop maths and details of their RGB and Lab formats are unknown to me. Hence I can not quantify the noise, but I think the images speak a very clear language themselves.
Summary
