# Graduation Talks

## title

*Adaptive thinning of weather observations*

## speaker

## date & place

Wednesday, 18.07.2007, 15:00 h

Room C252

## abstract

Numerical weather prediction (NWP) is akin to solving an initial value problem: a weather
forecast is computed as a time evolution of an estimated initial weather state. The forecast
error is to a large extent due to the miss-specification of the initial state estimate. To
reduce the estimation error, a precise information about the atmospheric state is required.
The modern meteorological observation systems, in particular satellite instruments, produce
ever increasing amounts of weather measurements with high spatial and temporal density.
Nowadays, the extremely large size of the measurement sets becomes prohibitive for their
complete incorporation into a weather forecast model. Therefore, a preprocessing routine,
called observation thinning, is commonly applied at many of the NWP centers to reduce
the size of the observational data sets prior to the state estimation.
Observation thinning is a procedure that maps a complete observation set onto one of
its subsets. Most of the NWP centers nowadays apply a simple thinning strategy resulting
in uniform spatial distribution of thinned observations regardless of a given meteorological
situation. In contrast to this non-adaptive method, the adaptive observation-thinning
algorithms attempt to identify regions characterized by a sensitive meteorological state
and adapt the spatio-temporal distribution of thinned observations accordingly. Two such
methods are presented here: a cluster-based thinning scheme that is basically a multidimensional
clustering algorithm and an Estimation-Error-Analysis algorithm that makes
use of the leave-one-out cross-validation technique. A test study of these algorithms in
the operational NWP model of the National German Weather Service yielded controversial
results. Therefore a simplified experimental NWP-framework was developed to gain a
deeper insight into the properties of the thinning algorithms with respect to both estimation
and forecast errors.