Hi Stefan,
yes its still the same remaining difference, which I think is acceptable.
Regards
René
On 22.06.2011 09:03, Stefan Buehler wrote:
> Hi,
>
> how large is the remaining difference? Still the 0.1-1.0 K that you reported
> earlier? I agree with Patrick's judgement that it is likely to be an
> interpolation issue. My earlier suggestions were also in that direction. So,
> in case you want to investigate further, use finer integration step and
> pressure grid in both models, then they should converge to the same result.
>
> But up to you to decide if you want to dig further into this, or just accept
> the slight difference.
>
> Glad that things over all seem to work ok. :-)
>
> /Stefan
>
> On Jun 21, 2011, at 19:13 , Rene Bleisch wrote:
>
>
>> Hi Stefan and Patrick,
>>
>> Forward model:
>> - the remaining difference between the forward model is highly correlated
>> with vmr at surface (corr=0.98)
>>
>> Retrieval:
>> - the qpack2 retrieval seems to converge better than qpack1. With Qpack2 511
>> of 511 cases converged, with Qpack1 only 434 of 511. The number of
>> iterations tends to be higher for Qp1.
>> - an interesting fact concerning Qp1: if I use the old iteration_ok
>> condition (iteration=ok if 1 | cost_new< cost_old) instead of the correct
>> one (iteration=ok if cost_new< cost_old), all cases converged.
>>
>>
>> QPack-doc:
>> - in sections 6.1.2 and 7.1, Q.T_ATMDATA should be changed to Q.T.ATMDATA
>>
>> Regards
>> René
>>
>>
>> On 20.06.2011 20:37, Stefan Buehler wrote:
>>
>>> Hi Rene,
>>>
>>> the RT schemes in ARTS 1 and 2 are not identical, so some differences are
>>> expected. 1 K seems a bit on the large side, but .2 K can easily be due to
>>> this.
>>>
>>> Basically, the two programs should converge if ppath_lmax and the
>>> corresponding ARTS1 variable l_step are small enough. (And, especially for
>>> ARTS1, p_grid also has to be fine enough.) Some of the differences have to
>>> do with how the atmosphere and the absorption are assumed to behave between
>>> the grid points, so it is to some extent a matter of definition.
>>>
>>> If you set both p_grid step and the integration step variables to 100 m,
>>> both versions should give very similar results. The new version should
>>> achieve the same accuracy with somewhat coarser p_grid than the old
>>> version, because absorption is interpolated in a smarter way.
>>>
>>> /Stefan
>>>
>>> On Jun 20, 2011, at 18:15 , Rene Bleisch wrote:
>>>
>>>
>>>
>>>> Hi Stefan and others,
>>>> in the meantime I could locate the major source of the difference between
>>>> the forward models:
>>>> I forgot to include the O2, N2 and CO2 absorption part in Qpack2/Arts2.
>>>>
>>>> I then made a series of tests with different setups:
>>>> - backend/sensor part and "other tags" (=O2, N2, CO2) disabled
>>>> - backend/sensor enabled
>>>> - "other tags" enabled
>>>> - all enabled
>>>>
>>>> The used H2O a-priori profiles are calculated from a monthly climatology
>>>> from radiosoundings combined with actual surface data from our meteo
>>>> station, as I use it for my retrievals. The O2, N2 and CO2 a-prioris are
>>>> also climatologies but kept constant. The forward model calculations were
>>>> performed for H2O data from Mai to mid-June in a 2hour-resolution (~550
>>>> different cases).
>>>>
>>>> "dfy.eps" and "rfy.eps" show timeseries of the absolute resp. relative
>>>> differences between QPack1/Arts1 and QPack2/Arts2 Forward Model at line
>>>> center (22.23 GHz). "fy.eps" shows timeseries of the absolute values, the
>>>> color code is the same as for dfy.eps and rfy.eps, circles show
>>>> QPack1/Arts1, lines Qpack2/Arts2.
>>>>
>>>> In summary, the results are quite satisfactory, but still a small
>>>> difference of 0.2 to 1 K (or 1-2%) remains, nearly independent of whether
>>>> the backend or "other tags" are enabled.
>>>>
>>>> I will continue testing the Forward model, and currently I'm performing a
>>>> set of retrievals to see the influence of the changes on the forward model.
>>>>
>>>> Regards
>>>> René
>>>>
>>>>
>>>> PS
>>>> Here the main setup of my forward models/retrievals (if not written
>>>> otherwise, the setup is identical in both versions):
>>>> - frequency range: 22.235 GHz +/- 1GHz
>>>> - Lines: from HITRAN (Arts1: fmin=0, fmax=10^18)
>>>> - Lineshape: Lorenz (quadratic, cutoff=-1)
>>>> - H2O-model: H2O-PWR98
>>>> - Sensor: Altitude 905m, zenith angle: 30°
>>>> in Qpack/Arts1, groundSet{z = 905, e = 0}, z_plat { value = 905};
>>>> za_pencil=30
>>>> in Qpack/Arts2, sensor_pos= 6378905m (r_geoid is added to z_platform
>>>> by default in Qpack2), z_surface=905m; sensor_los=30
>>>>
>>>> - Vertical grid setup:
>>>> - ppath_lmax=500 (corresponding to Q.STEPLENGTH_RTE = 500 in QPack1,
>>>> resp l_step=500 in Arts1)
>>>> - H2O a-priori: snd climatology+surface value from meteostation,
>>>> vertical grid resolution ~1km;
>>>> - pTz: ECMWF+surface value, vertical grid resolution: ECMWF-grid
>>>> (increasing from ~200m at 1km to 500 m at the tropopause); identical in
>>>> Qpack1 and 2
>>>> - p-grid (retrieval grid): vertical resolution ~1km;
>>>>
>>>> Notice: in Qpack1/Arts1, the interpolation of the pTz and vmr grids to the
>>>> final grid seems to be done in Arts1 (all vmr grid-files contain also
>>>> their corresponding p-coordinates), whereas Qpack2 already performs the
>>>> interpolations and the xmls for Arts2 need to contain the data on a unique
>>>> vertical grid. (here all a-priori vmr-profiles are merged to a tensor in
>>>> one xml.file, without the corresponding vertical coordinate).
>>>> Could this difference in interpolation lead to a difference in forward
>>>> model output?
>>>>
>>>> On 20.06.2011 09:26, Stefan Buehler wrote:
>>>>
>>>>
>>>>> Hallo René,
>>>>>
>>>>> Things to look at, perhaps:
>>>>>
>>>>> - Are the absorption settings really the same (same H2O continuum, etc.)?
>>>>> - What is the vertical grid spacing of your atmoshere? If this is
>>>>> relatively coarse, it may be a good idea to set: "NumericSet( ppath_lmax,
>>>>> 250 )" in the controlfile. (The default value for this in general.arts is
>>>>> 1000, which is a bit large.)
>>>>>
>>>>> /Stefan
>>>>>
>>>>> On Jun 16, 2011, at 14:21 , Rene Bleisch wrote:
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> Hi Patrick,
>>>>>> - Sensor/backend: removing the sensor/backend part lead to no
>>>>>> improvements, the differences are still the same
>>>>>> - Forward model: For the same example a-priori vmr, Arts1 delivers a
>>>>>> spectrum with 43.1 K at the center frequency of 22.235 GHz, Arts2 a
>>>>>> spectrum with only 39.7 K (Rayleigh-Jeans BT) resp. 40.2 K (Planck BT)
>>>>>> at 22.235 GHz using the settings from the retrievals, the deviation
>>>>>> slightly increases with frequency. Thus choosing Planck or RJ matters,
>>>>>> but the difference is small (only 0.5 K in the example).
>>>>>>
>>>>>> Hence the problem obviously seems to be in the Forward modelling and not
>>>>>> in the retrieval itself. I will now carefully go through the entire
>>>>>> forward model setup for both versions (which should be the same).
>>>>>>
>>>>>> Regards,
>>>>>> René
>>>>>>
>>>>>>
>>>>>> PS: During the tests with no sensor/backend part, I discovered that
>>>>>> Qpack2 ends with an error, if y is chosen in L2_EXTRA and the
>>>>>> sensor/backend are disabled.
>>>>>> Qpack2 then crashes because qp2_l2 takes L2.f from Q.SENSOR_RESPONSE_F
>>>>>> (line 137). But as the backend is disabled, neither this variable
>>>>>> (containing the name of the xml-file with the sensor-response frequency
>>>>>> grid) nor the xml-file itself are created during the processing.
>>>>>>
>>>>>>
>>>>>>
>>>>>> On 15.06.2011 14:47, Patrick Eriksson wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>>> Hi René,
>>>>>>>
>>>>>>> My answer is in line with the one of Stefan. The first step is to
>>>>>>> check if you get the same spectrum from the two forward models, for
>>>>>>> the same input. That is, no inversions involved. First test without
>>>>>>> sensor. And if OK, include also the sensor.
>>>>>>>
>>>>>>> Looking a bit on the retrieval part. The logrel unit is the most
>>>>>>> tricky one. If the tests above are all OK, please compare weighting
>>>>>>> functions for rel/frac.
>>>>>>>
>>>>>>> Just ask if anything is unclear. I want of course to know if there is
>>>>>>> a bug soemwhere. That can happen even in ARTS/Qpack ;-)
>>>>>>>
>>>>>>> Bye,
>>>>>>>
>>>>>>> Patrick
>>>>>>>
>>>>>>> On 06/15/2011 12:47 PM, Rene Bleisch wrote:
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>> Hi all,
>>>>>>>> I use Qpack to retrieve tropospheric water vapour profiles from spectra
>>>>>>>> of our 22GHz radiometer MIAWARA. The setup is like this:
>>>>>>>> - nonlinear Marquardt-Levenberg
>>>>>>>> - polyfit (1st grade, coefficients are part of the state vector and are
>>>>>>>> retrieved)
>>>>>>>>
>>>>>>>> Till recently, I used the old Qpack1 and it worked well. Some months
>>>>>>>> ago
>>>>>>>> I started trying to setup the same retrieval with QPack2. As it didn'
>>>>>>>> work well and I had a lack of time, I gave it then up. Now I retried
>>>>>>>> it
>>>>>>>> and it works quite well with QPack2 after setting the retrieval unit
>>>>>>>> to
>>>>>>>> logrel (implicitly the Qpack1-retrieval was set to logrel, what I only
>>>>>>>> discovered thanks to Patriks suggestions).
>>>>>>>>
>>>>>>>> Still the results with QPack2 differ from the results with QPack1, as
>>>>>>>> the vmr is generally up to 20% too low in upper troposphere and 10-20%
>>>>>>>> too high in lower troposphere. More detailed analyses revealed that
>>>>>>>> there is a difference between the weighting functions in Qpack1 and
>>>>>>>> Qpack2 (even in the first iteration step), the tropospherical maxima of
>>>>>>>> the weighting functions in Qpack2 are generally up to 10% lower than in
>>>>>>>> Qpack1.
>>>>>>>>
>>>>>>>> Does anyone have an idea where this difference could come from?
>>>>>>>> (spectroscopy and pTz setup are identical)
>>>>>>>>
>>>>>>>> Maybe it has to do something with the sensor/backend-part, which should
>>>>>>>> in principle be the same for both. In Qpack1 the H-matrix (y=H*F(x,b))
>>>>>>>> summarizes the entire sensor/backend stuff. I wanted to compare H with
>>>>>>>> its equivalent in QPack2, but I could not find it. Does there exist a
>>>>>>>> similar H-matrix in Arts2/QPack2?
>>>>>>>>
>>>>>>>> Regards
>>>>>>>> René
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>>
>>>>>> --
>>>>>> René Bleisch
>>>>>> Institute of Applied Physics
>>>>>> University of Bern
>>>>>> Sidlerstr.5
>>>>>> 3012 Bern
>>>>>> Switzerland
>>>>>>
>>>>>> Phone: +41 31 631 89 59
>>>>>> Mail: [email protected]
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> qpack mailing list
>>>>>> [email protected]
>>>>>> https://www.sat.ltu.se/mailman/listinfo/qpack
>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>>
>>>> --
>>>> René Bleisch
>>>> Institute of Applied Physics
>>>> University of Bern
>>>> Sidlerstr.5
>>>> 3012 Bern
>>>> Switzerland
>>>>
>>>> Phone: +41 31 631 89 59
>>>> Mail: [email protected]
>>>>
>>>>
>>>> <dfy.eps><fy.eps><rfy.eps>
>>>>
>>>>
>>>
>>>
>> --
>> René Bleisch
>> Institute of Applied Physics
>> University of Bern
>> Sidlerstr.5
>> 3012 Bern
>> Switzerland
>>
>> Phone: +41 31 631 89 59
>> Mail: [email protected]
>>
>>
>>
>
>
--
René Bleisch
Institute of Applied Physics
University of Bern
Sidlerstr.5
3012 Bern
Switzerland
Phone: +41 31 631 89 59
Mail: [email protected]
_______________________________________________
qpack mailing list
[email protected]
https://www.sat.ltu.se/mailman/listinfo/qpack
