[Go-essp-tech] Global attributes and DRS extensions for downscaled datasets

Wed Mar 27 03:23:02 MDT 2013

Hello Karl,

Thanks for clear response – I probably should have been able to work that out if I had followed the email thread carefully.

There is a separate document for CORDEX which describes the intended mapping of attributes onto facets in the ESGF User Interface – I’ll try to send that to you later.

I understand your reservations about some aspects of the CORDEX data requirements, but the aim was to use terms which are in usage in the community (particularly for the “region” attribute which actually combines region and resolution) and thus, hopefully, improve compliance and acceptance of the standard.

There is an aspiration to include statistically downscaled data in the CORDEX archive (and it is another shortcoming of the CORDEX document you refer to, I think, that it only deals with dynamically downscaled data and does not leave a hook to allow extension to statistically downscaled). The system you’ve described could presumably be used for statistically downscaled data in CORDEX. We have a new European Union project starting next week which funds Bruce Hewitson’s group in Cape Town to do some coordination and networking on data standards for CORDEX downscaling, so I’ve copied in Chris Jack who is, I think, leading their effort.

Regards,
Martin
From: Karl Taylor [mailto:taylor13 at llnl.gov]
Sent: 27 March 2013 00:51
To: Juckes, Martin (STFC,RAL,RALSP)
Cc: galina at ucar.edu; obc at dmi.dk; colin.jones at smhi.se; ncpp_core at list.woc.noaa.gov; ncpp_tech at list.woc.noaa.gov; go-essp-tech at ucar.edu; laura.e.carriere at nasa.gov; gerald.potter at nasa.gov; williams13 at llnl.gov; denis.nadeau at nasa.gov; Pascoe, Stephen (STFC,RAL,RALSP)
Subject: Re: Global attributes and DRS extensions for downscaled datasets

Dear Martin,

I'm not advocating changing the CORDEX requirements; it's probably much too late for that.  There are are limitations to the generality of the CORDEX specifications, which means they might not be applicable to downscaling efforts outside of CORDEX.  The document I prepared was to try to address the more general issue of what descriptors are needed for downscaled datasets.

I have proposed that a single additional "descriptor" be added to the already defined components of the DRS:

Source of predictor data ⇒ driving_model_id - driving_model_rip (e.g. “GFDL-CM3-r1i1p1”)   In some cases the driving_model_rip might be omitted (e.g., when using reanalysis output to drive the downscaling).

In CORDEX this descriptor could be formed by joining with a hyphen your GCMModelName and CMIP5EnsembleMember.

I have also proposed expanding the "ensemble member" descriptor to include an indication of the "nominal resolution".  The idea here is that output might need to be regridded or be made available at various resolutions, so we would like to be able to distinguish among these closely related datasets.  Here is the description of the 'riph' designator:

 Ensemble member⇒  ‘riph’ designator, where the “rip” form is defined as in CMIP5 (which for downscaled data would usually be “r1i1p1”), and the “h” is followed by nominal resolution expressed in kilometers.   (For backward compatibility the DRS would consider the “h” segment as optional, but it is required for downscaled datasets.)  The last part of the 'riph' designator is of the form “hnXXXX” or “hiXXXX” where XXXX is the nominal horizontal resolution of the downscaled data, expressed in kilometers (rounded to the nearest km with leading zeros dropped).  “hn” indicates that the data is stored on the model’s “native” grid, while “hi” indicates that the data has been interpolated from a model’s native grid to a different grid.  (Statistically downscaled data would normally be recorded on a so-called “native” grid.)  Data on a native grid at a nominal resolution of 5 km, for example, would be identified as “hn5”, while regridded data at 11 km resolution would be identified as “hi11”. The XXXX should be calculated as follows:  XXXX = sqrt(domain area / (number of grid cells)), expressed in km/grid cell and rounded off to the nearest km.

CORDEX has chosen to include resolution information as part of a domain name (e.g., CAM-44 or SAM-44i), but the resolution doesn't seem to me to belong as part of the region identification.

I should note also that CORDEX specifies a directory structure and/or filenames where in the CORDEX document some of the DRS categories are renamed.  I've attached a table that shows the DRS elements and corresponding CORDEX identifiers, along with global attributes.  (I'm going to try to get NASA / NOAA to be consistent with the DRS.)  I also provide a table of additional global attributes.  CORDEX is mostly consistent with this table, except for using "CORDEX_domain and omitting driving_model_tracking_ids.

Finally, I note that in the example found in the CORDEX document for global attributes:

1)  experiment_id = "evaluation",  but in the directory structure and filename templates, this is presumably used as "CMIP5ExperimentName",  but of course "evaluation" is not a CMIP5 experiment.  I think a better term for "CMIP5ExperimentName" is simply "experiment", which in the case of CORDEX is usually the same as the CMIP5 experiment_id.

2)  CORDEX requires "contact", but this was left out of the example.

Please let me know what you think.

Best regards,
Karl
On 3/26/13 4:58 AM, martin.juckes at stfc.ac.uk<mailto:martin.juckes at stfc.ac.uk> wrote:
Hello Karl,

I’m puzzled about how this fits in with CORDEX. We went through this discussion some time ago, and agreed on some data requirements in the document you cite below which we believed to be appropriately consistent with the CMIP5 requirements. This document was then discussed at a WCRP meeting and has been circulated as the requirements for groups submitting CORDEX data to ESGF. Since then, modelling groups have been preparing data and we are expecting to start publication soon.  Do you think there are problems with uniformity in the way the CORDEX requirements are specified?

Regards,
Martin

From: Karl Taylor [mailto:taylor13 at llnl.gov]
Sent: 25 March 2013 21:50
To: Galia Guentchev
Cc: ncpp_core at list.woc.noaa.gov<mailto:ncpp_core at list.woc.noaa.gov>; NCPP TECHNICAL TEAM; go-essp-tech at ucar.edu<mailto:go-essp-tech at ucar.edu>; laura.e.carriere at nasa.gov<mailto:laura.e.carriere at nasa.gov>; Potter, Gerald Lee. (GSFC-606.2)[UNIVERSITY OF MARYLAND]; Dean Williams; Nadeau, Denis (GSFC-610.1)[R S INFORMATION SYSTEMS, ]; Juckes, Martin (STFC,RAL,RALSP); Pascoe, Stephen (STFC,RAL,RALSP)
Subject: Global attributes and DRS extensions for downscaled datasets

Dear all,
I have spent considerable time reviewing the following four documents:
A. The email (copied below) sent by Galia and Aparna, which proposed attributes, filenames, and directory structures for downscaled data.
B.  http://cmip-pcmdi.llnl.gov/cmip5/docs/cmip5_data_reference_syntax.pdf which describes the corresponding CMIP5 metadata.
C. http://cordex.dmi.dk/joomla/images/CORDEX/cordex_archive_specifications.pdf<http://cordex.dmi.dk/joomla/images/CORDEX/cordex_archive_specifications_121022.pdf> which describes the corresponding CORDEX metadata.
D.  http://cmip-pcmdi.llnl.gov/cmip5/docs/CMIP5_output_metadata_requirements.pdf which specifies all the CMIP5 metadata requirements.
I hope that document A above could be made compatible with the others and in general could provide a sound basis for establishing more uniformity moving forward.  Toward that end, I have prepared the attached document describing for downscaled data a minimal set of  global attributes needed to augment those used in CMIP5 and also the extensions needed to the DRS document to accommodate downscaled data.
I hope at least a few of you will take the time to study this document and provide feedback.
Best regards,
Karl

Mail sent by Galia Guentchev 3/12/13

**********************************************************************
Details of each element of the proposed directory structure:

Proposed elements -
/projectID/sub-project/product/institution/predictorModel/experimentID/frequency/realm/MIPtable/Pred
ictor_experiment_rip/predictorversion/downscalingMethod/predictand (variableName)/region/DownscaledDataversion/file_name.nc

Example:

/ncpp2013/perfectModel/downscaled/NOAA-GFDL/GFDL-HIRAM-C360-coarsened/amip/day/atmos/day/r1i1p1/v20121024/GFDL-ARRMv1/tasmax/US48/v20120227/tasmax_day_amip_r1i1p1_downscaled_US48_GFDLARRMv1_19790101-19831231.nc

The new element sub-project (in blue above) gives the opportunity to indicate to users that in the one case the method was trained on observations (standard setting), and in the other on model that was considered to be the truth (perfect model setting);
The options there could be: PerfectModel or Standard - where possibly there could be a different name instead of 'standard' for the standard downscaling setting.

For NASA datasets some of the directories could be:
project = NEX
product = downscaled
institution = NASA-Ames
predictorModel - original model value
experimentID = historical
frequency = mon
realm = atmos
Predictor_experiment_rip - original model value
variable = precipitation or temperature
region = CONUS

DownscalingMethod will also be included as a directory to allow for search on method.

**********************
There are a set of sub-directories that refer to the PredictorModel - presented in bold - /predictorModel/experimentID/frequency/realm/MIPtable/Pred
ictor_experiment_rip/predictorversion

Where:
·         predictor model - is the specific GCM which is the source of the predictor data set - GFDL-HIRAM-C360-coarsened - in the above example
·         experimentID - the specific experiment - amip in this case
·         frequency - refers to the temporal scale of the predictor fields - daily
·         realm - the realm of the predictors - in this case atmos(phere)
·         MIPtable - name of the model intercomparison table - daily in this example, could be amon - for atm monthly data;
·         Predictor-Experiment-rip - follows the standard notation from CMIP5
·         version - the version date of the global model that provided the predictor dataset

The elements above follow quite closely the structure for CMIP5 model output directory elements.
There is a set of sub-directories that refer to the Downscaling method - presented in italics -
downscalingMethod/predictand (variableName)/region/DownscaledDataversion

Where:
·         downscalingMethod - is the downscaling method abbreviation - in this case GFDL-ARRMv1 - the GFDL in the name indicates that this is a setting applied by GFDL where there were two sets of predictors, based on the ARRM method of K.Hayhoe; also v.1 indicates which version of the ARRM method was used (the original version) - more details about the method are given in the global attributes of the file;
·         Predictand (variableName) - the specific predictand variable that was downscaled; tasmax in this case;
·         region - indicates that the method was applied to the US48
·         DownscaledDataversion - the version of the downscaled dataset

For the purposes of standardization there are two directions to consider:

1) One is to have one standard directory structure that will be used by all - for example, following the example of GFDL to have the details of the predictor model first and then the downscaling method details:
·         ProjectID - sub-project - product - Institution - Predictor dataset details - Downscaling method details - Filename

Having a standardized approach would help any automated service/web service to detect the directory path for a particular dataset.

2) During our last teleconference there was a proposal to follow the downscaling practice and describe the downscaling method first and then the predictor model. This leads to two paths:

        • ProjectID - Standard or Perfect Model sub-project facet - product - Institution -  then see below:
               -  (if Perfect model setting) Predictor dataset details - Downscaling method details,
               -  (if Standard setting) - Downscaling method details - Predictor dataset details

The NCPP Core team accepts that it may be reasonable to have a directory structure - where the method description is first; and another directory structure - where the predictor description is first and then the methods that are applied are described; NCPP will support either approach (one overall directory structure, or two separate pathways) and if the second approach is chosen (with two different sub-directory sequences) - we would like to promote and to support the standardization of these different directory pathways - meaning - we will support two standardized directory structures to accommodate two common practices.

******************
Additional details:

Variable level attributes-
The published dataset should also conform to CF-standards.
eg-

                tasmax:long_name = "Downscaled Daily Maximum Near-Surface Air Temperature" ;
                tasmax:units = "K" ;
                tasmax:missing_value = 1.e+20f ;
                tasmax:_FillValue = 1.e+20f ;
                tasmax:standard_name = "air_temperature" ;
                tasmax:original_units = "K" ;
                tasmax:downscaling_method: GFDL-ARRMv1

Global attributes- listing a few here, several CMIP-style attributes will be inherited.

"predictorModel" will replace "model_id"
  For the 'downscaling model', as agreed with Luca on the call it would be 'downscalingMethod'

                :Conventions = "CF-1.4" ;
                :references = "info about model, training datasets etc will be provided here"
                :info = "additional info about the downscaling method"
                :creation_date = "2011-08-19T21:57:06Z" ;
                :institution = "NOAA GFDL(201 Forrestal Rd, Princeton, NJ, 08540)" ;
                :history = "info on file processing. Eg" processed by toolX." ;
                :projectID = ncpp2013
                :subprojectID = perfectModel
                :product = downscaled
                :institution = NOAA-GFDL
                :predictorModel = GFDL-HIRAM-C360-coarsened
                :experimentID = amip
                :frequency = day
                :modeling_realm = atmos
                :Predictor_experiment_rip = r1i1p1
                :region = US48
                :table_id = day
                :version = v20120227
                :downscalingMethod = GFDL-ARRMv1
**************************************************

Best regards,
Galia and Aparna

--

Galia Guentchev, PhD

Project Scientist

National CLimate

Predictions and

Projections

Platform (NCPP)

NCAR RAL CSAP

FL2 3103

3450 Mitchell Lane

Boulder, CO, 80301

phone: 303 497 2743

--
Scanned by iCritical.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.ucar.edu/pipermail/go-essp-tech/attachments/20130327/dd5837ae/attachment-0001.html 