Saturday, June 9, 2012


cc: "Phil Jones" <>
date: Wed, 26 Mar 2008 12:32:16 +0000
from: Ian Harris <>
subject: Re: further CRUTS2.1 vs 3.0 comparisons

Hi Tim,

No criticism inferred! Sorry if I was snappy - as you can imagine I'm
heartily sick of this dataset! In a professional way of course.

My proposal is as follows:

� Derive TMN/TMX from gridded DTR and TMP anomalies.

Special cases:

� Where TMP is missing, TMN/TMX/DTR marked missing.
� Where DTR is missing, TMN/TMX = TMP, DTR = 0.

Please agree or adjust these (I've probably misunderstood somewhere),
then I'll code up a converter later today.



On 25 Mar 2008, at 9:17, Tim Osborn wrote:

> Hi Harry,
> sorry for any implied criticism of you, Harry -- I appreciate the
> difficulties in working from incomplete documentation (having just
> been
> modifying ClimGen over the last week!). Just trying to get through
> all
> this fast, since some partners have collaborators arriving TODAY
> and need
> the observational data to calibrate their hydrological model with.
> The preference for gridding TMP and DTR and then deriving TMN and TMX
> arises in regions where only TMP exists or early on when only TMP
> exists.
> Suppose one such month has a TMP ANOMALY of -4 C. With no DTR, TMN
> or TMX
> data, gridding each of these would lead to them being relaxed to their
> climatological normals, i.e. anomalies of 0 C. Yet users will ask,
> how
> can both TMN and TMX have zero anomaly at the same time that TMP
> anomaly
> is -4 C?!!! The answer might be "well look at the station coverage
> data
> and you'll see that we don't in fact know what TMN and TMX are, so
> just
> don't use them at all in this instance". But the alternative would
> be to
> say that our best guess in the absence of real information is that
> TMN and
> TMX anomalies are also -4 C, which is what we'd get if we derived them
> from TMP and DTR=zero. Not likely to be completely correct, but
> the guess
> will have some skill over assuming their anomalies are zero.
> I'm not sure what you want to do for CRU TS 3.0, but for ClimGen I
> shall
> go ahead with TMP, DTR as provided, and then calculate TMN and TMX as
> described earlier.
> Cheers
> Tim
> On Sun, March 23, 2008 10:53 pm, Ian Harris wrote:
>> Hi,
>> I did indeed follow a different process from 2.10. This was
>> remiss, but
>> can be explained.
>> Firstly, I don't really think the published papers reflect the
>> actualit� of how the dataset was produced. Neither do the read me
>> files. They contain elements of truth but there are glaring
>> inaccuracies (including the gridding method, possibly the most
>> important single issue). For this reason I've not been as slavish to
>> them as I should have been.
>> Secondly, the tmin and tmax databases are very good. I've removed
>> many
>> duplicates and 'badly augmented' stations and the two databases I
>> used
>> are completely aligned. I think that if stations are measuring a
>> parameter then that's the truth of the matter (conceding the
>> complications arising from measurement times).
>> There are climatologies for DTR, TMN and TMX. This also pushed me
>> towards treating them as primary parameters.
>> I derived a DTR database from the TMN and TMX databases, and gridded
>> all three. That's what you're playing with.
>> If you want TMN and TMX to be derived from TMP and DTR, so be it
>> (though it seems superfluous since it's simply derived).
>> The problem comes when one considers this process - when we are
>> talking
>> about TMN and TMX as theoretical constructs that may not have existed
>> at all.
>> The TMN, TMX (and therefore DTR) databases each have 13654 stations.
>> Even if they require further cleaning, and I'm sure that they do,
>> that's still a lot more than TMP.
>> Incidentally, sorry but I don't really follow this deduction:
>>>> Since TMN and TMX are both correlated with TMP, the v2.1 method is
>>>> clearly the right way to go.
>> When either are correlated with TMP, then the strength of that
>> correlation is of great interest, isn't it? But the approach of TMN =
>> TMP - DTR/2, TMX = TMP + DTR/2 just creates dummies with a
>> correlation
>> of 1.0. And although the correlation between TMP and DTR is very
>> important, DTR is not the only information contained in the TMN
>> and TMX
>> databases.
>> Cheers
>> Harry
>> On 21 Mar 2008, at 18:17, wrote:
>>> Tim,
>>> As far as I'm concerned, the T fields that get gridded
>>> are Tmean and DTR. It might be that the climatology
>>> is for Tmean, Tx and Tn. If this is the case then we
>>> need a DTR climatology.
>>> Then what you say below is how I would calculate as you say.
>>> Hopefully this is what Harry has done.
>>> It might be that Harry has done a Tx and Tn gridding as the
>>> two datasets aren't exactly the same - so you lose a bit of
>>> data getting DTR, but it is very, very small.
>>> Cheers
>>> Phil
>>> Phil -- if Harry says "yes", then will I get exactly the "correct"
>>>> result if I ignore the TMN and TMX files that Harry's made and
>>>> instead make my own from TMP and DTR, using:
>>>> TMN=TMP-0.5DTR and TMX=TMP+0.5DTR?
>>>> I tested this and now get much better correlations with v2.1 for
>>>> TMN,
>>>> and presumably (not checked yet) for TMX. The standard deviations
>>>> are now much more similar too.
>>>> Here's the comparison when I make my own TMN from TMP and DTR for
>>>> v3.0 and compare with TMN from v2.1.
>>>> Tim
>>>> --------------------------
>>>> Phil & Harry,
>>>> Next problem! :
>>>> I've now made comparisons for TMN, TMX and DTR for v3.0 vs. v2.1.
>>>> Attached is the result for TMN. You'll see we now have 4 plots per
>>>> month, for Jan-Apr-Jul-Oct. Plot 1 is the temporal
>>>> correlation. Plot 2 is the SD of v2.1. Plot 3 is the ratio of
>>>> v2.1
>>>> SD to v3.0 SD. Plot 4 is the SD of v3.0.
>>>> Clearly the correlation is rather weak in many areas and the
>>>> ratio of
>>>> standard deviations exceeds 1.5 across most of S America, Africa
>>>> and
>>>> India, plus Greenland and N. Russia.
>>>> I think that this may be due to the way you've made the data. I've
>>>> checked Mitchell and Jones (2003) and it says very clearly that TMN
>>>> (and TMX) are derived variables, taken entirely from the grids
>>>> of TMP
>>>> and DTR (presumably TMN = TMP - 0.5*DTR and TMX = TMP +
>>>> 0.5*DTR???).
>>>> Now if you haven't followed this approach for making v3.0 but have
>>>> instead "independently" made gridded fields of TMP, DTR, TMN and
>>>> TMP,
>>>> relaxing to climatology where there are no nearby observations, so
>>>> that they are all primary variables and none are derived, then
>>>> if you
>>>> have TMP data but not DTR, TMN or TMX values, the latter 3 will be
>>>> relaxed to climatology. But the v2.1 methodology would relaxed DTR
>>>> to climatology, but TMN would be TMPactual - 0.5*DTRclimatology,
>>>> and
>>>> hence would still have variations that paralleled the observed
>>>> variations in TMPactual.
>>>> Since TMN and TMX are both correlated with TMP, the v2.1 method is
>>>> clearly the right way to go. The only time when something
>>>> different
>>>> to both approaches might be useful is if you have lots of
>>>> stations/months with only TMN or TMX but not both. But Harry says
>>>> that you generally have both or none. In which case v2.1 will
>>>> be the
>>>> best we can do.
>>>> Harry -- now that I've confirmed what Mitchell & Jones did to make
>>>> v2.1, can you confirmed that you have made v3.0 in the way I
>>>> described
>>>> above?
>>>> I'll attach the plot with the next
>>>> email (too big for this one!).
>>>> Given the time and effort I've put in to CRU TS 3.0, I shall expect
>>>> to be a co-author when the paper describing CRU TS 3.0 is written!
>>>> Cheers and happy Easter,
>>>> TimDr Timothy J Osborn, Academic Fellow
>>>> Climatic Research Unit
>>>> School of Environmental Sciences
>>>> University of East Anglia
>>>> Norwich NR4 7TJ, UK
>>>> e-mail:
>>>> phone: +44 1603 592089
>>>> fax: +44 1603 507784
>>>> web:
>>>> sunclock:

Ian "Harry" Harris
Climatic Research Unit
School of Environmental Sciences
University of East Anglia
Norwich NR4 7TJ
United Kingdom


No comments:

Post a Comment