From rls@email.unc.edu Sun Feb 23 20:48:26 2003
Date: Sun, 23 Feb 2003 20:48:06 -0500 (EST)
From: Richard Smith <rls@email.unc.edu>
To: env-email@samsi.info
Subject: Atmospheric science subgroup meeting and topics



To remind everyone, we have a meeting of the "atmospheric sciences"
subgroup at SAMSI at 11am on Wednesday (2/26), and there is parallel
meeting of the "porous media" subgroup at 1pm.

For the atmospheric sciences subgroup, what I hope we can focus on is the
identification of some "test problems" that we can make the theme of
smaller subgroups that will meet at regular intervals for the remainder of
the program. The two main criteria for a test problem are (1) it should be
consistent with the interdisciplinary focus of the program, (2) there
should be a realistic prospect of measurable progress, such as the
submission of one or more research papers, during the time frame of the
program. The latter criterion should serve as a constraint on proposing
problems that are too broad.

To get the ball rolling, I would like to propose a number of possible
candidates for test problems. Some of these are my own idea, but most have
arisen from discussions with one or more colleagues over the past six
weeks. There is a bias towards problems with a strong statistical
component, but I hope the applied mathematicians and atmospheric
scientists will counter that by proposing their own problems. Roughly what
I would like to do is (a) between now and Wednesday, invite others to
propose their own problems or possible modifications of these, (b) during
Wednesday's meeting, agree which problems we are actually going to work on
as well as defining working groups and a leader for each group. If there
are people receiving these emails who would like to participate but cannot
make the 11am Wednesday meeting time, please let us have your suggestions
and preferences as well. Once the groups are defined, it will be the
responsibility of each group leader to work out a schedule that allows the
participation of all those who want to participate.

Richard


Suggested problems:

1. Large-scale ozone models (proposed by Prasad Kasibhatla (Duke School of
the Environment) and Jim Zidek).

Prasad has hourly data collected on 36x36 km grids cells for a six-month
period in 1995. The data are generated by an atmospheric chemistry model
that includes a meteorological model (MM5) and information about sources.
A major uncertainty is the actual emissions from the various sources. The
model data has been compared with real data from monitors, to assess the
performance of the models and to improve information about the sources.
Other objectives include the optimal combination of model and monitor data
for ozone forecasting. The problem would seem an excellent representative
of a whole class of problems in this field of research (e.g. those
generated by the EPA's Models-3 group) and could therefore serve as a
test-bed to develop new methods for a broad range of issues.

2. Source apportionment (also proposed by Prasad and Jim Zidek).

The problem here is to estimate emissions at a known set of sources, given
measurements at a set of monitors, and a (physical) transport model that
relates the monitor data to the emission. The effective problem is of the
form y = Ax + e where y is a set of monitor readings, x is a set of
unknown emissions, A is a transport matrix (calculated from the transport
model and assumed known) and e is error. Bayesian solutions of this
problem depend critically on the form of spatial prior assumed for x. A
very similar problem was discussed by Tapio Schneider in his workshop
presentation whose approach was more in the spirit of applied mathematics
(regularization of ill-posed problems) but the essential issues are the
same. It would seem to be a good test-bed for the comparison of
statisticians' and applied mathematicians' techniques.

Note: I have placed on the web page a separate note (written by Jim with
input from Prasad) describing these problems in more detail. See
http://www.stat.unc.edu/faculty/rs/envmod/NotesJan03.txt

3. Incorporation of dynamics into spatial-temporal statistical models
(proposed by Sandra McBride and various other people)

In recent years, statisticians have developed a number of new approaches
to spatial-temporal data but relatively few of them have made a direct
attempt to incorporate physics. Among those that have are the papers of
Chris Wikle (Sandra mentioned Chris in an earlier email but go to Chris's
web page, reached via http://www.stat.missouri.edu/, for the 2001 JASA
paper and numerous others) and the papers and book of George Christakos
and Marc Serre of UNC's ESE department. This would seem another excellent
area to combine the expertise of statisticians, applied mathematicians and
atmospheric scientists. We would need to identify some test data sets and
it may be that Prasad's ozone data set (see problem 1 above) could be a
candidate for this.

4. Design of monitoring networks (Jim Zidek, Montse Fuentes, Zhengyuan
Zhu, Dave Holland, ...)

This refers to the very broad problem of specifying the locations at which
an environmental variable is to be monitored; possible objectives are very
varied but include prediction at locations away from the monitors
(assessed by some criterion such as integrated mean squared error),
inference about unknown parameters of the model, or possibly some more
specialized objective such as maximizing the probability of detecting
threshold crossings in the context of enforcing a standard. There are
possible links with the optimization activity being led by Tim Kelley in
the porous media group. (To make progress we would have to start by
defining more specific objectives than this, but I think if we had a
meeting of all interested parties, we would have a good chance of doing
that.)

5. Trends in means and extremes in climate models (Amy Grady, Gabi Hegerl)

Much recent climatological research has moved beyond modeling trends in
mean temperatures towards alternative "indices of climate change", many of
which involve possibly changing probabilities of extreme events (e.g. are
hurricanes becoming more frequent as a result of anthropogenic climate
change?). Data are available from both climate models and observational
networks; one interpretation of the big-picture problem is to combine the
two courses of information to obtain realistic projections of extreme
events 50 or 100 years ahead, with associate uncertainty estimates.

6. Health effects of air pollution (Robert Wolpert, Sandra McBride, Jim
Zidek...)

I won't try to define a specific problem here since there seem to be a
wide range of views about this topic, but a number of our group have
worked in this field and it should be possible to define something
specific for us to look at during this program

7. Mathematical and statistical methods for data assimilation

This came up a couple of times during the workshop. Data assimilation
broadly refers to problems of integrating data from a real observational
network into a physical model. Most of the work in this field has been
done by applied mathematicians but some of the methods are closely
associated with various statistical themes such as non-linear extensions
of the Kalman filter and the use of Monte Carlo methods such as particle
filter. There may also be some merit in discussing associated network
design problems, as in topic 4 above (there is an interesting and not very
well known paper on this subject coming out of the NCAR group:  Berliner,
L.M., Lu, Z.-Q. and Snyder, C. (1999), Statistical design for adaptive
weather observations. {\it J. Atmos. Sci.} {\bf 56}, 2536--2552).