date: Wed Apr 9 14:58:15 2003
from: Mike Hulme <m.hulmeatXYZxyz.ac.uk>
subject: Re: CRU interpolated climate
to: Tim Mitchell <t.mitchellatXYZxyz.ac.uk>
Tim - see my comments at the end .......
At 16:59 07/04/2003 +0100, you wrote:
I'll answer as best I can, but please do not quote these answers, as I ought
to collaborate with co-authors before giving any quotable comments.
> I am now thoroughly confused and would be very grateful if you could
> sort me out! I have read your guidelines on the web-site, and need
> help with interpreting the following:
> "These choices mean that while this data-set is suitable for using as
> an input to environmental modelling, it is NOT suitable for use in
> detecting climate change. It is NOT a legitimate use of this data-set
> to attempt to prove or disprove the existence of climate change at an
> individual grid-box."
> My questions are:
> 1. Is the 1960-2000 climate time series really not to be used at all
> to detect climate change, even over aggregated, regional areas?
It depends on the region, period, and climatic variable! For 1961-1990, say,
and for the European mainland, there will probably not be a problem. The
density of stations is sufficient that individual stations coming in and out
are not likely to substantially affect the values over this large area.
However, over central Africa this is probably not true.
Climate change detection is a specialised subject. It demands either
individual station time-series, or carefully assembled (usually
low-resolution) grids. See Q4 of the FAQ.
> can it be used as input for environmental modelling if it is not
> accurate enough to show real phenomona of change?
The high-resolution grids do show "real phenomona of change". However, it is
not the long-term changes for which the grids are optimised; the grids are
optimised for high-resolution 'snapshots', month by month.
Perhaps it would help if I gave an example of how the data-sets can be best
used in data-sparse regions.
1. Constructing a trend at a grid-box is not a good idea, as we explained in
the Nature paper.
2. It would be legitimate to use linear regression to compare (say) April
precip over a few grid-boxes (or perhaps even one grid-box if the data at
that box seems to warrant it) with some comparable areal (not point!!)
environmental index from 1981-2000, to derive an estimate of the
relationship between interannual variability in precip and interannual
variability in the environmental index.
> Where does this
> leave all the previous publications from CRU on regional climate
Largely unchanged, as I see it. These high-resolution grids are our 'best
estimate' of the climate, at a high spatial resolution, in each month in
1901-2000. Perhaps the risks of temporal inhomogeneities at the level of
individual grid-boxes could have been made clearer in the past - that is a
matter of judgement I guess. The coarser-scale grids produced by CRU, such
as Phil Jones' work, are not affected because they use different methods.
> 2. Over what scale do you consider it legitimate to make spatial
> comparisons? Again, some of the publications show, for example, maps
> of Africa with different climate anomalies over about 1000km. With
> greater densities of met stations in Europe, is the spatial
> resolution any better there?
I find it hard to give a definitive answer, because the spatial scale over
which the climate information is temporally homogenous varies with region,
period, and climatic variable. My answer above provides some hints.
> I absolutely appreciate the problem of the changing input from met
> stations through time - we face the same sorts of irregular
> sequential data input from satellite sensors. And I equally
> appreciate that interpolation must blur the differences between
> neighbouring grid-boxes - but over what distance relative to the
> spatial distribution of input met stations?
This depends on the spatial scales over which different variables vary. See
the New et al (2000) paper for the precise values used.
> We are being asked again and again to analyse patterns and causes of
> "emerging" diseases in many parts of the world, and we are really
> concerned to make real sense of the subject, which involves having an
> accurate idea of the degree of climate change within land masses the
> size of Europe. I am myself about to send off a paper for a
> conference proceedings concerned with tick-borne diseases in Europe.
> I have no agenda at all - I am as happy to discover that there has
> been, or has not been, any relevant climate change to account for the
> variety of temporal and spatial patterns of disease change across
> Europe, but I am desperately keen to get it right as a basis for
> further work.
> Looking forward to a fruitful dialogue with you.
PS if any co-authors cc'd want to comment, please feel free!
Tim - worth distinguishing between two types of problems with the New et al. data set:
(a) it is specifically *not* designed for climate change detection/attribution in the
classic IPCC anthropogenic GHG context because for environmental simulation we wish to
capture *all* the changes in regional/local climate whether or not an artefact of urban
development or land use change (this is the exact *opposite* of data sets for GHG detection
since all such datasets should remove such influences - there is a string of papers going
back 10 years or more criticising CRU/Phil's work on these very grounds - urban
heat/desertification influences, etc.).
(b) a largely unrelated weakness in the dataset is the inhomogeneity introduced due to
changing station coverage over time. And here you are right to point out that the
"accuracy" depends on place, season, variable and scale of aggregation. Mark has some
error grids I believe and publishing maps of # stations in interpolation range would help,
but in the end the data set relaxes to 1961-90 in the absence of actual station anomalies.
This is what you mean by space-optimised, but space-optimised inevitably implies it becomes
inhomogenous over time (increasingly so as scales become smaller in data sparse areas).
The other point worth advising people is if they really want to look at very local scale
(certainly sub-grid-scale, but maybe even supra-grid scale in data poor areas) issues -
whether trends or environmental modelling - then they would be best advised to approach
GHCN (or CRU) for access to the underlying station data. Then of course, people need to
pay attention to the credibility and homogeneity of individual station series, in itself
not a trivial task and one that dozens of papers have been written about.
Hope this helps - share these comments with Phil or whoever else is appropriate.
Dr. Tim Mitchell
Tyndall Centre for Climate Change Research
phone: +44 (0)1603 59 3904
fax: +44 (0)1603 59 3901
post: Tyndall, ENV, UEA, Norwich, NR4 7TJ, UK