Physically based data assimilation and fuzzy verification
Gad Levy
NorthWest Research Associates
1:30 pm March 17 in Room 2551
Abstract: Ideally, a verification and validation scheme
should be able to evaluate and incorporate lower dimensional
features (e.g., discontinuities) contained within a bulk simulation
even when not directly observed or represented by model
variables. Nonetheless, lower dimensional features are often
ignored. Conversely, models that resolve such features and the
associated physics well, yet imprecisely are penalized by
traditional validation schemes. This can lead to (perceived or real)
poor model performance and predictability and can become
deleterious in model improvements when observations are sparse,
fuzzy, or irregular. I present novel algorithms and a general framework
for using information from available satellite data through fuzzy
verification that efficiently and effectively remedy the known
problems mentioned above. As a proof of concept, we use a sea-ice model
with remotely sensed observations of leads in a one-step initialization
cycle. Using the new scheme in a sixteen day simulation experiment
introduces model skill (against persistence) several days earlier than
in the control run, improves the overall model skill and delays
its drop off at later stages of the simulation. Although sea-ice models
are currently a weak link in climate models, the appropriate choice of
data to use, and the fuzzy verification and evaluation of a
system’s skill in reproducing lower dimensional features are
important beyond the initial application to sea ice. Our strategy
and framework for fuzzy verification, selective use of information, and
feature extraction could be extended globally and to other disciplines.
As time permits, examples from tropical climate classification,
oil spill detection and modeling will also be presented.