Update of Infrared Atmospheric Sounding Interferometer (IASI) channel selection with correlated observation errors for numerical weather prediction (NWP)

Coopmann, Olivier; Guidard, Vincent; Fourrié, Nadia; Josse, Béatrice; Marécal, Virginie

doi:https://doi.org/10.5194/amt-13-2659-2020

Articles | Volume 13, issue 5

https://doi.org/10.5194/amt-13-2659-2020

© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/amt-13-2659-2020

© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 13, issue 5

Research article

|

26 May 2020

Research article |

| 26 May 2020

Update of Infrared Atmospheric Sounding Interferometer (IASI) channel selection with correlated observation errors for numerical weather prediction (NWP)

Olivier Coopmann, Vincent Guidard, Nadia Fourrié, Béatrice Josse, and Virginie Marécal

Download

Final revised paper (published on 26 May 2020)
Preprint (discussion started on 23 Jul 2019)

Interactive discussion

Status: closed

AC: Author comment | RC: Referee comment | SC: Short comment | EC: Editor comment

- Printer-friendly version

- Supplement

RC1: 'Referee Comments', Anonymous Referee #1, 27 Aug 2019
- AC1: 'Response to Referee #1', Olivier Coopmann, 02 Mar 2020
RC2: 'Insufficient explanations of why this strategy was chosen. A bit verbose. But clear a lot of work had been done.', Anonymous Referee #2, 12 Nov 2019
- AC2: 'Response to Referee #2', Olivier Coopmann, 02 Mar 2020
RC3: 'Review of Coopman et al.', Anonymous Referee #3, 22 Nov 2019
- AC3: 'Response to Referee #3', Olivier Coopmann, 02 Mar 2020

Peer-review completion

AR: Author's response | RR: Referee report | ED: Editor decision

AR by Olivier Coopmann on behalf of the Authors (02 Mar 2020) Author's response Manuscript

ED: Referee Nomination & Report Request started (13 Mar 2020) by Thomas von Clarmann

RR by Anonymous Referee #2 (01 Apr 2020)

RR by Anonymous Referee #1 (06 Apr 2020)

Suggestions for revision or reasons for rejection

Update of IASI channel selection with correlated observeration errors for NWP
Coopman et al

Summary

The paper describes a new method for IASI channel selection which is based on a
statistically derived observation error covariance matrix rather than the simple
diagonal models assumed in the past. This representation is intended to allow
for spectral correlations introduced by the forward model errors as well as
instrument noise.

The determination of this OE covariance follows the method suggested by
Desroziers which is essentially to evaluate the NWP model background covariance,
the IASI-background covariance and attribute the difference to observation
error. However, evaluating an appropriate background covariance it itself a
difficult task.

It is noted that OE covariances are now regularly evaluated and used in NWP
assimilation schemes, but these are all based on predefined channel
selections. So the main purpose of this work is to try to incorporate the OE
covariance into the channel selection itself.

The authors demonstrate that their new selection leads to improved performance
in a data assimilation context compared to the 'standard' channel selection of
Collard.

Main Comments
(for these, I do not expect the authors to revise their experiment, but I would like to see either some explicit answer or at least acknowledgement of the questions in their revised paper).

1) The Collard selection had a number of restrictions, particularly
a) exclusion of spectrally adjacent channels
b) selection of CO2-only, then allowing channels with H2O and O3 sensitivity
Both these restrictions are relaxed in this case so it is not clear how much
improvement is simply due to this relaxation and how much due to the more
complex representation of the OE covariance. In particular I would expect (a) to
limit the ability to sound at higher altitudes since most of the high-altitude
CO2 channels are concentrated around the Q-branch at 667cm-1, while (b) would
limit the channel selection to the edge of the main CO2 band rather than
extending across the window region where water vapour starts to contribute.
There should at least be some explicit acknowledgement that removing these
restrictions would, by themselves, be expected to lead to some improvement.

2) Local Jacobian matrices (H) were computed for each of the 60 profiles, but
(it seems) the same global R and B matrices were used. Ideally, local R and B
matrices would also be used unless there is good reason to assume that these are
constant (which seems unlikely). Was this considered?

3) It appears that 60 separate channel selections were derived (Fig 9) but it
was unclear how these were combined into a single selection. It could be that
the selection across all 60 profiles was performed simultaneously using an
aggregated or averaged DFS but this isn't stated.

4) The assimilation formalism (eg Eq 6) assumes that the background and
observation error covariances are completely uncorrelated with each other. That
might be true until the first IASI measurements are assimilated but thereafter
the background covariance updated for the next assimilation step will have some
component from the OE error so no longer strictly independent. Where a large
amount of information for the NWP is coming from other sources this may be
negligible, but it should be stated somewhere that this has been assumed
(although the authors also note P2 L20) that 60% of all observations assimilated
in ARPEGE are from IASI.

5) The authors show the main features of their derived OE covariance (Fig 5) but
there is very little discussion of whether they believe this is truly
representative or how much of the magnitude may be ascribed to uncertainties in
the evaluation of the background covariance. Why, for example, would the
observation error be larger in the ozone band than in the H2O bands? On the
other hand it seems quite plausible that the ozone contribution to the
background error covariance has not been properly characterised. If the
background and observation errors are indeed uncorrelated, is it just
coincidence that where the observation error is large the background error is
also large?

Minor/Typographical Comments

Abstract
P1 L2: IASI = Infrared Atmospheric Sounding Interferometer

P1 L24: I suggest temperature error more conventionally expressed in K rather
than % (also elsewhere)

P1 L40: 'The IASI spectrum ...'

P2 L17: I think "This 'analysis' state is thus ..." would read better.

P2 L25: What does uncorrelated 'vertically' mean with regard to IASI? There is
no vertical coordinate in the measurement space.

P2 L41: For the opening sentence of this paragraph it is not clear whether the
authors are referring to previous work or what they will be presenting
in this paper.

There should be some clarification here: if the only 'observation error'
considered is the instrument noise, then it is quite valid to consider only the
diagonal elements of the covariance matrix, with the additional precaution of
discarding adjacent channels (a side-effect of the Gaussian apodisation being to
introduce correlations in the noise between adjacent spectral points).

P2 L64: Jacobian (with capital J), and numerous other instances.

P2 L88: For completeness, which IASI instrument? And were the profiles
restricted to near-nadir views or did you sample the full across-track
swath? Did you use different FOV elements from the set of four?

P2 L98: At nighttime is the AVHRR cloud flag reliable? How would it distinguish
between clear surface and a stratiform cloud top? Usually there is an
additional test based on a comparison of the retrieved skin temperature
compared with the model forecast.

P3 L33: 'the the'. Also 'Degrees of Freedom' (plural).

P3 L33: Strictly speaking DFS is just one of a variety of different criteria
that could be used and, since it ignores off-diagonal information, isn't
actually the one that provides the 'largest information content'. It is,
however, the one that is conventionally used for channel selection, so I
have no argument with the choice of DFS here as well.

P3 L49: 'represents the Jacobian matrix ...'

P3 L45: Eq (2) should be 'I + ' ....

P3 L61: If you're using 645-2000 that should be 5421 channels, or 5420 without
channel 1194. I assume 645-2019.75cm-1 is intended here.
also P15 L40

P3 Table 1: For completeness give ranges of band 1 and band 2 separately in the
table caption

P4 Fig 2 caption: 'subset of the 60 atmospheric profile database'.

P4 L1: I expect that once m gets beyond a fews 10s this becomes computationally
expensive. Can you give some indication of the time required?

P4 L8: Why is this section titled 'Preliminary' work? That suggests further work
will be presented.

P4 L23: Use consistent font (ie math font) for BT and X in the text and in
Eq(4)

P4 L42: 'also to water vapour' - this, and indeed the whole of panel b, are
inconsistent with Table 1 which suggests H2O is only retrieved from band
2.

P4 L51: Temperature sensitivity is, of course, a necessary accompaniment to
sensitivity to *any* absorber, so every part of the spectrum will be
sensitive either to surface or to atmospheric temperature.

P5 Fig 3: caption missing explanation of panel (d). 'Jacobians' should have
capital J. The pressure axis on panel (d) is inappropriate and would be
better converted to show the size of the Jacobian.

P6 L13: The usefulness of a 'poly-sensitive' channel depends mainly on the
assimilation scheme modelling all the relevant parameters. This would be
true even for a single channel, so inter-channel correlations are no more
(or less) important for 'poly-sensitive' channels than for single
parameter channels.

P6 L35: Apart from the large departure of 3K which I guess is at the centre of
the CO2 Q-branch at 667cm-1

P6 L80: 'B and H are ...'

P7 Fig 5: It would be better if Fig(a) were flipped so IASI channel number was
on the x-axis

P7 L44: It is not clear what use is made of the diagonal R matrix. I assume it
forms part of the pseudo-inversion procedure for the full R?

P10 Fig 8: These would be better with logarithmic x-axes.

P11 Fig 9: y-axis title 'Number of channels' slightly clipped in my PDF file,
although that may just be a local issue.

P11 L8: Since DFS has a physical interpretation as 'number of independent pieces
of information that can be retrieved' (ie effectively independent
profile levels) it makes more sense to present these results as actual
DFS values rather than percentage of the total. For example I would
expect skin temperature to be very close to 1, but '6.5% of total'
means nothing to me.

P11 L12: I am unclear at this point as to whether you now have 60 separate
channel selections, one based on each profile, or just a single
selection based on solving for an aggregated/averaged DFS over 60 profil
es

P11 L30: I believe the Collard selection considered only a single atmospheric
profile (?). Are you using a selection based on the same single profile
here?

P11 L44: It does not surprise me that there is not much overlap with the Collard
channels. With such high redundancy there are usually a large number of
channels available at each step in the iteration, each differing only ver
y
slightly in information content. Of course Collard also excludes adjacent
chanels to those already selected. But does Collard use the same set of
profile levels, which would also modify a selection based on DFS?

P12 L21: 'Jacobians' (capital J).

P13 Fig 11: 'Jacobian' (capital J).

P13 L13: In terms of coverage represented purely by the Jacobians I see no
reason why inter-channel errors should encourage a more homogeneous
coverage. Could this be related to density of profile levels in different
regions of the atmosphere? Or ability to select adjacent channels in the
new algorithm (I imagine this would be a particular limitation for the
Collard algorithm since high altitude CO2 channels are limited to the
667cm-1 region). Also the Collard selection starts with a limitation to
CO2-only channels so would exclude channels which could be used jointly.

P13 L19/20: '... channels in the first ...'

P14 Fig 12: I assume the profile of DFS values are evaluated from
(1 - (sigma^a)^2/(sigma_b)^2 ) for the different selections? So, closely
related to the (sigma^a-sigma^b) plots. It would be helpful to make this
clear.

-------

Hide

ED: Publish subject to minor revisions (review by editor) (09 Apr 2020) by Thomas von Clarmann

AR by Olivier Coopmann on behalf of the Authors (18 Apr 2020) Author's response Manuscript

ED: Publish subject to technical corrections (24 Apr 2020) by Thomas von Clarmann

AR by Olivier Coopmann on behalf of the Authors (25 Apr 2020) Author's response Manuscript

Short summary

The objective of this paper is to make a new selection of IASI channels by taking into account inter-channel observation-error correlations. Our selection further reduces the analysis error by 3 % in temperature, 1.8 % in humidity and 0.9 % in ozone compared to Collard’s selection, when using the same number of channels. A selection of 400 IASI channels is proposed at the end of the paper which is able to further reduce analysis errors.