|
Hi Mike and Joe
Firstly, sorry for such a long posting – I won’t do it again, promise.
Mike, thanks for the overview of the processing procedure in your last post – this has been a big help. I should make it clear that I am under no allusion that I can waltz in and provide any real improvements to your already excellent results. You guys really know your stuff and I have a lot to learn. I doubt now that the more conventional techniques used in remote sensing for atmospheric correction are transferable to planetary RS as they are not designed to cope with multiple frames of the same wavelength band.
These techniques I refer to are either absolute models (that use non-visible wavelengths to measure atmospheric absorption) or relative models (that use ground-based target spectral signatures to correct the effect of the atmosphere ‘relative’ to image derived signatures).
There is one method though that I alluded to in my last post that may be used on DSOs and not dissimilar to the correction technique you use at the pre-processing stage – ie. altering the gain of the sensor for each wavelength - called Dark Object haze reduction (some times called the histogram method of haze reduction) and is described by Pat Chavez (1988) – I can email via PM the paper if anyone is interested.
Basically, It requires the analyst to look in dark areas of an image where there should be no reflectance (earth: deep clear water or shadowed areas; Jupiter image: the dark space to the side of the disk). While not strictly true, these pixels should be zero in value, but of course they often have values significantly >0 due to atmospheric scattering and residual noise of the sensor. These minimum values (different for each band due to different amounts of scattering) can be used as a Starting Haze Value (SHV) – the amount that should be subtracted from each image band respectively to reduce haze. However, Chavez describes how only one SHV is typically used from one wavelength band (e.g. the blue band, as it scatters light the most) and is then used along with a simple scattering model to predict the SHV in each of the other bands. For example:
Simple scattering model:
Very Clear = λ-4
Clear = λ-2
Moderate = λ-1
Hazy = λ-0.7
Very Hazy = λ-0.5
So given the approx. centre wavelength for RGB with a Very Clear sky, then:
B = 0.485um x -4 = 18.07
G = 0.560um x -4 = 10.17
R = 0.660um x -4 = 5.27
Now calculate a multiplication factor to predict haze:
18.07/18.07 = 1
10.17 / 18.07 = 0.56
5.27 / 18.07 = 0.29
So finally, if the SHV in the blue band is say 40 (for 8 bit data) then the predicted SHV for all other bands would be:
Band 1 (blue): 1 x 40 = 40
Band 2 (green): 0.56 x 40 = 22.4
Band 3 (red): 0.29 x 40 = 11.6
Notably, these SHVs should be determined from the raw radiance data (measured in physical units - watts) and not the quantized 8, 10… or 32bit image data. Most RS data like that from Landsat’s ETM+ sensor, is supplied as 8 bit data. All users are supplied with the transfer gain and offset values for each wavelength band – each band will record a slightly different gain – ie. The range of the incoming signal (radiance) relative to the range of the signal output (the image quantization level), which effects contrast. The offset is the recording of energy when there is no energy present, which effects brightness and is equivalent to the SHV of each band.
So after this very long winded explanation, you guys are already altering the contrast and brightness during both the pre and post processing stage – ie. you are basically doing the same thing. The only difference is, in RS they might be applied a little more objectively rather than subjectively (primarily for consistency between image dates for change detection work). Clearly, the objective approach does not make for better visual reconstruction.
I am going to stop over theorizing now and learn more about what you guys do and just start having a go – Bird’s July 25 Jupiter image is telling me just that. Thanks for listening.
Nix.
|