On 09/02/12 22:26, Tudor Hughes wrote:
On Feb 9, 9:28 pm, Liam wrote:
I know this forum is populated by various people with meteorology
backgrounds, so thought I'd ask here first...

For my PhD I am assimilating data into a general circulation model. Now
what I want to do is compare the assimilation result (a simple lat-lon
plot of whatever) with a lat-lon plot of the observations.

I can simply subtract one from the other to get a difference plot, but I
was wondering if there was a statistical way to compare the 'fit'
between two 2D plots? The model calculates the average rms error over
all lat-lon points, but this isn't very helpful, as one particularly bad
gridpoint can skew the whole result.

If you can give any helpful advice I shall endeavor to mention you in
the acknowledgements of my thesis. )

Many thanks,

--
Liam (Milton Keynes)http://physics.open.ac.uk/~lsteele/

You mention a lat-long plot of the observations as the base to
compare the assimilated figures with but surely observations are not
made on a lat-long grid but at irregularly spaced points. If one of
these observations is grossly in error it can be weeded out before
attempting the fit but to do so would require the meteorological
expertise acquired through experience.
Alternatively, all the observations could be used, whether
suspect or not and if the error at one point exceeds the rms error of
the entire grid by some specified amount it could be removed at that
stage though that would make the comparison invalid though it may well
be more use operationally.
I can't see any way round using the rms error even if a rogue
point or two spoils it. December 2010 "spoilt" the (fairly) smooth
winter warming of the last few decades but it happened and can't be
ignored.

Tudor Hughes, Warlingham, Surrey

Hi Tudor.

Yes, I thought about restricting the domain that is used for quantifying
the analysis, but that seems a bit sneaky, as though I don't trust the
model results! What seems to be happening is that the model predicts
broadly the right features (e.g. clouds) but these might be shifted
slightly so they are in an adjacent grid box to the observations. Thus,
when calculating the rms, the model is penalised twice: once in the grid
box with too much cloud and again in the adjacent grid box with too
little cloud.

What I would ideally like is something that can compare magnitudes as
well as the spatial offset of model predictions. A look on the link Andy
posted suggests I need something like the Displacement and Amplitude
Score. How I would implement something like this is not clear though. I
imagine it's not simple!

--
Liam (Milton Keynes)
http://physics.open.ac.uk/~lsteele/