|
Titel |
Improved source term estimation using blind outlier detection |
VerfasserIn |
Marta Martinez-Camara, Benjamin Bejar Haro, Martin Vetterli, Andreas Stohl |
Konferenz |
EGU General Assembly 2014
|
Medientyp |
Artikel
|
Sprache |
Englisch
|
Digitales Dokument |
PDF |
Erschienen |
In: GRA - Volume 16 (2014) |
Datensatznummer |
250089375
|
Publikation (Nr.) |
EGU/EGU2014-3574.pdf |
|
|
|
Zusammenfassung |
Emissions of substances into the atmosphere are produced in situations such as volcano
eruptions, nuclear accidents or pollutant releases. It is necessary to know the source term –
how the magnitude of these emissions changes with time – in order to predict the
consequences of the emissions, such as high radioactivity levels in a populated area
or high concentration of volcanic ash in an aircraft flight corridor. However, in
general, we know neither how much material was released in total, nor the relative
variation of emission strength with time. Hence, estimating the source term is a crucial
task.
Estimating the source term generally involves solving an ill-posed linear inverse problem
using datasets of sensor measurements. Several so-called inversion methods have been
developed for this task. Unfortunately, objective quantitative evaluation of the performance of
inversion methods is difficult due to the fact that the ground truth is unknown for practically
all the available measurement datasets.
In this work we use the European Tracer Experiment (ETEX) – a rare example of an
experiment where the ground truth is available – to develop and to test new source estimation
algorithms. Knowledge of the ground truth grants us access to the additive error term. We
show that the distribution of this error is heavy-tailed, which means that some measurements
are outliers. We also show that precisely these outliers severely degrade the performance of
traditional inversion methods.
Therefore, we develop blind outlier detection algorithms specifically suited to the source
estimation problem. Then, we propose new inversion methods that combine traditional
regularization techniques with blind outlier detection. Such hybrid methods reduce the error
of reconstruction of the source term up to 45% with respect to previously proposed methods. |
|
|
|
|
|