[Esip-infoquality] A different perspective on quality from GODAE

Cornford, Dan D.Cornford at aston.ac.uk
Tue Jun 7 18:15:29 EDT 2011


Greg, others,

  it can be true that adapt assimilation is sensitive to outliers when a Gaussian error distribution is assumed, as is the case for most currently operational systems, but that need not be so. In theory if a more robust characterisation of the errors on the observations were available (for example a Huber norm, or a mixture distribution) it would be possible to assimilate outlier observations, 'optimally' so long as the correct distribution of errors were known.

I also think that in data assimilation (and more generally in observation quality) one needs to very carefully define what is meant by an outlier. Is this the result of a real observation which due to the spatial and temporal sampling properties of the instrument with respect to the real natural variability of the observed field means that a very localised but good observation is made (representativity error) or has there been an instrument failure (probably want to identify these and remove). There are surely other causes of outliers too. 

I sort of agree that the human eye can often 'easily' filter out bad observations with experience (prior knowledge) and when these outliers are somehow clear (e.g. radar clutter or other phenomena which are sort of obvious most of the time), but I think there are also examples where a more systematic approach based on a model of the observation process that includes the processes generating outliers can also be very effective (particularly where the prior knowledge can be expressed as a mathematical model, e.g. a numerical weather prediction model).

I do not agree that it makes sense to equate quality with level ... might be sort of true due to the filtering of errors, but in doing the filtering you are probably also loosing information (unless you have a perfect filter ... but who has that!) and thus in some sense reducing the quality of the data set. 

cheers

Dan

-------------------------------------------
Dr Dan Cornford
Reader, Computer Science and NCRG
Aston University, Birmingham B4 7ET
 
www: http://wiki.aston.ac.uk/DanCornford/
 
tel: +44 (0)121 204 3451
mob: 07766344953
------------------------------------------- 

> -----Original Message-----
> From: esip-infoquality-bounces at lists.esipfed.org [mailto:esip-infoquality-
> bounces at lists.esipfed.org] On Behalf Of Leptoukh, Gregory G. (GSFC-6102)
> Sent: 07 June 2011 13:35
> To: Lynnes, Christopher S. (GSFC-6102); esip-infoquality at lists.esipfed.org
> Subject: Re: [Esip-infoquality] A different perspective on quality from
> GODAE
> 
> I would argue that this is an "assimilation-biased" perspective on data
> quality. Assimilation systems are known to be very sensitive to outliers,
> so for assimilation system is better to have just few data points but be
> sure that they are not outliers. On the contrary, if you monitor ash
> transport, you want to have an image with the best coverage, where a human
> eye can 'filter" outliers easily.
> 
> -----Original Message-----
> From: esip-infoquality-bounces at lists.esipfed.org [mailto:esip-infoquality-
> bounces at lists.esipfed.org] On Behalf Of Lynnes, Christopher S. (GSFC-6102)
> Sent: Tuesday, June 07, 2011 8:19 AM
> To: esip-infoquality at lists.esipfed.org
> Subject: [Esip-infoquality] A different perspective on quality from GODAE
> 
> Interesting...GODAE kind of equates quality level with processing level:
> http://www.godae.org/Data-definition.html.
> 
> I guess there may be some truth to that, in that you discard more and more
> suspect values as you go up the processing level chain:
> L0 = raw data, keep everything
> L1B = probably keep everything, but flag suspect quality based on
> calibration
> L2 = discard individual pixels where retrievals are impossible or
> (sometimes) where they are highly suspect
> L3 = (usually) do not include suspect pixels in grid averages
> L4 = use only the best retrievals for assimilation
> 
> Of course, most levels carry exceptions:
> L0 = delete packets with incorrect timestamps or apparent bit-level
> corruption
> L2 = keep only values that pass strict quality control (e.g., AIRS CO2)
> L3 = keep all retrievals (MODIS Image_Optical_Depth)
> --
> Christopher Lynnes
> Goddard Earth Sciences Data and Information Center, NASA/GSFC
> 301-614-5185
> 
> _______________________________________________
> Esip-infoquality mailing list
> Esip-infoquality at lists.esipfed.org
> http://www.lists.esipfed.org/mailman/listinfo/esip-infoquality
> _______________________________________________
> Esip-infoquality mailing list
> Esip-infoquality at lists.esipfed.org
> http://www.lists.esipfed.org/mailman/listinfo/esip-infoquality
> 
> -----
> No virus found in this message.
> Checked by AVG - www.avg.com
> Version: 10.0.1382 / Virus Database: 1511/3686 - Release Date: 06/07/11


More information about the Esip-infoquality mailing list