Taka brought up a good point in a response to an earlier posting of mine.
His question deals with the use of "exact" Rrs. (The NOMAD subset for this working group contains both the 'exact' and - for lack of better terminology - 'inexact' Rrs)
Here are some points to consider:
1) We currently produce "exact" nLw (to the best of our ability - and yes, using the Morel approach).
These "exact" nLw's are used in the derivation of Rrs, which is input to the various IOP models.
2) Our implementation of the Morel f/Q uses tables originally developed for the MERIS band set
So, I'm opening this Taka's question up to the group.
What is the most appropriate Rrs to use?
On a similar track:
How do we handle the fact that a number of the IOP inversions use the
Gordon (1988) formulation, which is not valid for view angles less than 20 degrees?
This is not a huge deal for SeaWiFS (which tilts 20 degress, so rarely views less than this), but MODIS and
MERIS are both nadir pointing and much of their swath will be belew 20 degrees sensor view angle.
My feeling is that we should use "exact" Rrs as best we know how to produce it. This also includes removal of solar zenith dependence on the above surface to below surface relationship [url="http://oceancolor.gsfc.nasa.gov/forum/oceancolor/topic_show.pl?tid=2657#fp"](as noted by Hubert)[/url], which I believe has already been removed via use of Morel's gothic_R.
One point to consider is that if we postpone the full normalization so that it can be done within the model inversion (as in the PML approach), we can not use such methods with confidence on Level-3 temporal composites. The path geometry associated with the "inexact" Rrs must also be provided to the model. The binned Rrs retrievals, however, may come from any part of the swath, and the mean path geometries that would be associated with the binned average are not necessarily representative (e.g., mean relative azimuth is effectively meaningless).
After posting my own message earlier regarding Rrs_ex, I found that Rrs_ex is a bit of headache for PML model... As you know, PML model actually finds not only IOPs but also f/Q (although implicitly), thanks to an iterative procedure. Thus, PML model already takes the angular distribution into account, and that was one of the novelty of PML model. LUT of f/Q used for the iterative procedure is generated by ourselves, and therefore different from Morel et al 2002. This means that we can't simply feed Rrs_ex into PML model, even if Rrs_ex is given to us.
As for use of Level-3 composite in PML approach (or in a model which derives f/Q by itself), composite images of Rrs or nLw can't be applied directly, as you pointed out.
There are perhaps two possibilities to ease the problem:
1) to use a default value of view angle (eg 45 degrees). Indeed this is how we have been doing by now when we use L3-composite images (note: sun angle is obtained from another LUT).
This is based on the following argument:
We can arugue that the ***underwater*** upward radiances originating from the nadir to approx. 32 degrees of nadir angle are relatively uniform, especially on the anti-sun side (i.e. the upward radiances scattered back towards the sun, which is better to use for remote sensing in practice; on the other side of azimuth, sun glint issue comes into play). The ***underwater*** upward radiances in that range correspond to the ***above-water*** upward radiances at 0 to 45 degree of ***zenith angle***, due to refraction at air-water boundary.
I know that the argument is not strictly true, but at least it is an approximation. It is clearly not a perfect solution, especially when sensor viewing angle is larger than 45 degrees. I am not familier with how radiance received at sensor angle > 45 deg. is appropriate for remote sensing, so the choice of 45 degrees was only an initial attempt to reduce the error for the case of sensor angle > 45 deg. and threfore may have to be reviewed.
2)to esimate IOPs at early stage of satellite processing, and generate composites of IOPs images, rather than to generate Rrs composite images first and then feed the Rrs composites into the IOP model.
I don't know if this is easy to do or not, but at least it is a possibility, I think.
You should be aware that the calibration INCLUDES the use of the Morel f/Q
If you choose to turn off this part of the normalization (by using brdf_opt=3)
you will need to used a different set of vicarious gain coefficients.
This is NOT an insignificant effect. Below is a comparison of SeaWiFS
vicarious gain coefficiets with and without the Morel f/Q applied:
f/Q ON: gain=[1.0361,1.0124,0.991,0.9973,0.9987,0.9743,0.9716,1.00]
f/Q OFF: gain=[1.0286,1.0044,0.983,0.9917,0.9955,0.9738,0.9716,1.00]
% difference as 100*(ON - OFF)/OFF
Notice the differences are spectral. Now, IF you do account for f/Q within
the inversion algorithm, the differences will be smaller than those I've listed
above, BUT the result will still be inconsistent. The degree of that inconsistency
will be dependent upon the level of inconsistency between the Morel f/Q
and you internal derivation as would be seen at our calibration site - MOBY.
Just so ya know...
Regarding your last point:
"2)to esimate IOPs at early stage of satellite processing, and generate composites of IOPs images, rather than to generate Rrs composite images first and then feed the Rrs composites into the IOP model."
"I don't know if this is easy to do or not, but at least it is a possibility, I think."
Retrieval of IOPs at Level-2, which are than composited to Level-3, is NASA's current approach. One of the goals of the workshop, however, is to assess differences between this approach and the alternative, wherein Level-3 Rrs composites are used as input to derive Level-3 IOP products. The need to evaluate the difference between these approaches arises because many researchers are using SeaWiFS, MODIS, or MERIS L3 composite nLw products as input to IOP models. We recognize that it might, in fact, be better to reduce noise in the Rrs through temporal/spatial averaging before application to model inversion, where the noise may get magnified in the derived products. The first step is to assess whether the two approaches yield significant differences, all things being equal. I am building the code to effectively allow Level-3 binned products to be input to the msl12 algorithms, to answer this first-order question.
However, if the consensus among workshop participants is that accurate IOP retrieval can best be achieved by incorporating sub-surface inhomogeneity corrections into the model iteration, I believe this negates the use of Level-3 composite Rrs as input to the model inversion, since composited radiant path geometries (especially relative azimuth) are not meaningful.
It would also suggest that the model inversion can provide a better correction to fully normalized nLw or Rrs, relative to our existing f/Q corrections based on Morel's tables. If that's really the case, we would want to use the IOP model results to produce the best-possible Rrs for all other derived products. I would like to avoid the introduction of inconsistencies across the product suite.
Many thanks for the information. It was very helpful indeed.
I see your point regarding the comparison between IOPs derived from L2 and L3. Very good strategy.
As for f/Q correction, I need to think a bit more...
By the way, when you use your f/Q correction for the satellite radiance, do you need Chla value for the correction?
Is any document available regarding your f/Q correction procedure? (I know that you used Morel et al 2002, but I would like
to know many other things, too).
nLw_ex = nLw * B0 / B
B = (f/Q) * Gothic_R for observed radiant path geometry (solz, senz, relaz)
B0 = B for solz=0, senz=0, relaz=0
and Gothic_R accounts for fresnel reflection & refraction at the interface. The msl12 implementation is updated from Morel 2002 to include Gordon 2005 and Wang 2006 changes to Gothic_R (references below).
The correction is actually derived through iteration, as we need nLw to compute chlorophyll (i.e., nLw -> chl -> nLw_ex -> chl' -> nLw_ex' -> chl'' -> nLw_ex'').
We are currently providing to all the IOP models Rrs_ex = nLw_ex / <F0>. We save the B0/B term internally, so it is possible to provide to the model(s) Rrs(=Rrs_ex*B/B0). That seems to be what the PML code is designed to expect. Alternatively, the PML code could use Rrs_ex and solz=0, senz=0, relaz=0 as the path geometry, since that is the appropriate geometry (theoretically) for Rrs_ex.
I would stress, though, that this workshop is not intended to be an algorithm shoot-out. It is precisely these kinds of differences that we wish to identify and explore. I think the important question here is whether an IOP model inversion can provide a better B0/B than we can currently produce externally. In principle, probably yes, in practice, ...?
The relevant code for msl12 is here:
(operationally, only the FRESNSEN, FRESNSOL, FOQMOREL options are active).
A. Morel, D. Antoine, and B. Gentilli, "Bidirectional reflectance
of oceanic waters: accounting for Ramen emission and varying
particle scattering phase function," Appl. Opt. 41, 6289-6306
H. R. Gordon, "Normalized water-leaving radiance: revisiting
the influence of surface roughness," Appl. Opt. 44, 241-248
M. Wang, "Effects of ocean surface reflectance variation with
solar elevation on normalized water-leaving radiance," Appl.
Opt. 45, 4122-4128 (2006).
>>The correction is actually derived through iteration, as we need nLw to compute chlorophyll (i.e., nLw -> chl -> nLw_ex -> chl' -> nLw_ex' -> chl'' -> nLw_ex'').
I thought one would need that iteration, if Chl is needed for f/Q correction. But you have already done it! That's simply excellent! (To be frank, I am amazed)
>>Alternatively, the PML code could use Rrs_ex and solz=0, senz=0, relaz=0 as the path geometry, since that is the appropriate geometry (theoretically) for Rrs_ex.
I have thought of that option, but I was not sure if it solves the issue about "inconsistency" between our and your f/Q, because our f/Q(solz=0,senz=0) could still be different
from your f/Q(solz=0,senz=0). If the difference could be viewed as a result of our approximation to your f/Q (rather than as an "inconsistency"), then it wouldn't be a problem anymore.
I was not quite sure about interpretation of the difference...
>> I would stress, though, that this workshop is not intended to be an algorithm shoot-out.
Sure. I just thought that this issue of f/Q applies to any (present or future) IOP model which considers the angular geometry in itself, and that this kind of discussion would be useful.
I think the discussion of f/Q approach versus Gordon approach is extremely useful. My understanding is that the PML model is very similar in approach to the QAA model, but PML uses f/Q and "current" QAA uses a re-tuned formulation of Gordon. Correct? It sounds like f/Q is the direction Ping is planning to take "future" QAA. Correct?
I hope that one or both of you (or Tim S.) can lead a discussion on this point at the workshop (and activities prior).
Yes, I will talk about our NASA project about this BRDF thing. As everyone knows, it really depends on the shape of VSF, in addition to absorption and scattering coefficients. I will show how we deal this this based on our extensive measurements of VSF.
I am looking forward to seeing your new results.
Meanwhile, we (I and Tim) also examine a sensitivty of PML model to VSF.
Let's discuss more later, when results become available.
Have a good Monday start...
Does the PML code expect input Rrs to be corrected for gothic-R (Fresnel reflection/refraction) dependence? I want to run a global test. I can give it Rrs without f/Q, with or without gothic-R.
Powered by mwForum 2.29.7 © 1999-2015 Markus Wichitill