![Hier klicken, um den Treffer aus der Auswahl zu entfernen](images/unchecked.gif) |
Titel |
Homogenization of climatic series via pairwise comparisons: An automated Bayesian algorithm |
VerfasserIn |
A. Hannart, O. Mestre, P. Naveau |
Konferenz |
EGU General Assembly 2009
|
Medientyp |
Artikel
|
Sprache |
Englisch
|
Digitales Dokument |
PDF |
Erschienen |
In: GRA - Volume 11 (2009) |
Datensatznummer |
250028887
|
|
|
|
Zusammenfassung |
We describe an automated homogenization algorithm based on pairwise comparison of
neighbouring station series. In homogenization, such pairwise comparison techniques avoid
falling back on regional reference series, which potential non homogeneity is often
considered a drawback. While they efficiently mitigates this issue, pairwise comparison
techniques implementation cost is nevertheless higher for several reasons: first, the set of
series to be scanned for shifts is larger since each candidate series may have several
neighbours; second, each shift detected on a paired difference series may be caused by any
two series and must hence be attributed to the culprit series; third, multiple shift locations an
amplitudes estimated on several paired difference series must be reconciled into a unique
position and amplitude to be used for adjusting the candidate series. High resulting
implementation cost of pairwise comparison may thus prohibit its use when stations are
many. To cope with this issue, the algorithm described here deals with each of these three
steps in a completely automatized way; in particular it does not involve visual inspection
in the first step, nor does it defer the second and third steps to a time consuming
manual review by an analyst as most previous methods do (Caussinus and Mestre,
2004). The algorithm rely on a Bayesian multiple change-point detection technique
(Hannart and Naveau, 2008) to cope with the first step, a method which benefits
from expert knowledge on jumps amplitude and frequency introduced through prior
distributions. Then, the algorithm takes advantage of posterior distributions obtained from
this Bayesian method to quantify the uncertainty on the position and amplitude of
shifts. Simple quantitative criterion are derived from these posterior distributions
and probabilities to attribute all detected shifts to a culprit series and reconcile
disparities in estimates of their position and amplitude, both of which are automatically
performed through explicit and simple calculations. The algorithm is tested on real and
simulated series leading to considerably improved processing time yet with a similar
performance level, compared to manually reviewed pairwise comparison techniques. |
|
|
|
|
|