Content. Many solutions (for monthly data)! Benchmark dataset ADVANCES IN HOMOGENISATION METHODS OF CLIMATE SERIES: AN INTEGRATED APPROACH

Transcription

1 The COST-HOME monthly benchmark dataset with temperature and precipitation data for testing homogenisation algorithms Victor Venema, Enric Aguilar, José A. Guijarro and Olivier Mestre COST Action Content Benchmark Why Generation of homogeneous data Insertion of inhomogeneities Preliminary results Meteorological Institute Bonn ADVANCES IN HOMOGENISATION METHODS OF CLIMATE SERIES: AN INTEGRATED APPROACH Olivier Mestre Météo-France France Enric Aguilar U. Rovira i Virgili Spain Ingeborg Auer ZAMG Austria Anders Grimvall Linköping University Sweden Petr Stepanek CHMI Czech Republic Tama's Szentimrey OMSZ Hungary Victor Venema University Bonn Germany Benchmark dataset Survey: surveys returned Monthly temperature and precipitation Additive vs. multiplicative process Test complete homogenisation algorithms Many solutions (for monthly data)! DETECTION Visual, Craddock test, Student t-test, Likelihood ratio test (SNHT), Potter test, Bayesian procedures, Local contrast test, Pettitt test, penalized likelihood, CORRECTION Composite reference series, interpolated reference series, multiple non-homogeneous series (), ANOVA (Mestre)

2 Benchmark dataset Survey: surveys returned Monthly temperature and precipitation Additive vs. multiplicative process Test complete homogenisation algorithms Statistical detection Correction methods Computation reference (if any) Handle outliers and missing data Complete system Benchmark dataset ) Real (inhomogeneous) climate records Most realistic case Investigate if various HA find the same breaks ) data Empirical distribution and correlations Insert know inhomogeneities Test performance ) Synthetic data For example, Gaussian white noise Insert know inhomogeneities Compare to synthetic data: test of assumptions Real data section France (Bourgogne): 9 rain rate France (Brittany): Tmin, Tmax The Netherlands: rain rate*, 9 Tmean Norway: 9 rain rate*, Tmean* Two precipitation networks with 9 and stations Two temperature networks with stations Catalonian region: rain rate, Tmin, Tmax * Some meta data available Outline creation benchmark ) Start with homogenised data ) Multiple surrogate and synthetic realisations ) Mask surrogate records ) Add global trend ) Insert inhomogeneities in station time series ) Published on the web ) Homogenize by COST participants and third parties ) Analyse the results and publish ) Start with homogenised data Austria: rain rate, Tmean France (Bourgogne): 9 Rain rate France (Brittany): Tmin, Tmax Catalonian region: Tmin, Tmax ) Multiple surrogate realisations Seven networks: need about networks Solution: data Multiple surrogate realisations Temporal correlations Station cross-correlations Empirical distribution function Advantage compared to autoregressive modelling Non-Gaussian Correlations on long time scales Rust, Mestre & Venema. Fewer Jump, less memory... JGR, Oct..

3 The iterative IAAFT algorithm Schreiber and Schmitz Flow diagram Time series Distribution IAAFT algorithm smoothes jumps Bounded Cascade time series of Bounded Cascade Time series LWP or LWC 99 LWP or LWC Time or space 9 Time or space One station with annual cycle One station anomalies Multiple stations year zoom Multiple stations year zoom - low cross correlation high cross correlation

4 Station Scatterplot stations monthly rain anomalies Station Station anomalies Station Synthetic data From the surrogate networks: synthetic networks are computed No temporal correlations Difference or ratio time series are Gaussian Do have almost the same spatial correlations Good comparison with surrogate data Test only influence of structure Same settings are used to generate inhomogeneities ) Mask surrogate records Beginning of records jagged (rough) Linear increase in number of stations Three stations in 9 Last station after % of full time End of record all stations are measuring Influence of jagged edge on detection and correction ) Mask surrogate records. ) Add global trend ) Insert inhomogeneities in stations Random breaks Frequency of breaks to / a Temperature Standard deviation breaks:. C Standard deviation seasonal cycle:. C Precipitation Standard deviation breaks: % Standard deviation seasonal cycle:.%.9..

5 ) Insert inhomogeneities in stations Simultaneous breaks Difference and ratio IH precipitation - 9 In % of network Frequency of breaks: in % of stations Examples inhomogeneities temperature.. ) Insert inhomogeneities in stations Outliers Frequency: per station, i.e. per a Size: 99 and 99.9 percentiles Outliers temperature and precipitation - 9 ) Insert inhomogeneities in stations Local trends (only temperature) Linear increase or decrease in one station Duration:, a Maximum size: same as breaks Frequency: once in % of the stations

6 Examples of local trends ) Published on the web Inhomogeneous data will be published on the COST-HOME homepage Final version: published in May ) Homogenize by participants Everyone is welcome to download and homogenize the data Homogenised by the end of the year Returned homogenised data should be in COST- HOME file format Also homogenising one or a few networks is fine Start with surrogate data Only fully automatic algorithm need to do synthetic Multiple homogenisation algorithms welcome Influence implementation and operator ) Analyse the results Detailed analysis will be performed in the working groups Detection Correction Preliminary analysis of previous version of the benchmark Contributions No. homogenised networks - algorithm Participant Algorithm Remarks Table. Number of homogenised networks per algorithm. José Guijarro. Péter Domonkos. Michele Climatol CM-D, -D, NSHT-D Versions with different settings different detection algorithms Detection Craddock based Homogenisation alg. All networks Real netw. netw. Synthetic netw.. Dubravka Rasol & Olivier Mestre. Matthew Menne & Claude Williams. Christine Gruber & Ingeborg Auer. Gregor Vertacnik. Petr Stepanek Automated pairwise hom. Versions Craddock Climatol A Climatol C Climatol D Climatol E Climatol F ClimatolG APHa APHa Lucie. Enric Aguilar NSHT CM-D -D SNHT-D

7 No. homogenised networks input data Mean no. outliers per station Table. Mean number of outliers per station for every algorithm Table. Summary data: Number of homogenised networks per network Network No. networks Temp. netw. Precip. netw. Homogenisation alg. All networks.. Real netw. netw... Synthetic netw... All.... Real Climatol A Climatol C Climatol D Climatol E Synthetic Climatol F ClimatolG APHa.9... APHa.9... CM-D.. -D.. SNHT-D.. Mean no. outliers per station Table. Mean number of outliers per station for every algorithm Homogenisation alg. All networks Real netw. netw. Synthetic netw Climatol A.... Climatol C Climatol D.... Climatol E.... Climatol F....9 ClimatolG.. APHa.9... APHa.9... CM-D.. -D.. SNHT-D.. Histogram no. outliers per station Temperature No outliers per station Temperature No outliers per station Mean no. breaks per station Table. Mean number of breaks per station for every algorithm Homogenisation alg. All networks Real netw. netw. Synthetic netw Climatol A.... Climatol C.... Climatol D.... Climatol E..9.. Climatol F.... ClimatolG.. APHa.... APHa...9. CM-D.. -D.9.9 SNHT-D.. Histogram no. breaks per station Temperature No breaks per station.... Precipitation No breaks per station

8 Root mean square error surrogate... RMSE Temperature surrogate.... RMSE Conclusions Interesting benchmark dataset Realistic structures Realistic inhomogeneities Preliminary results show large spread Automatic algorithms are still less accurate Different implementations of algorithm can produce very different results Everyone is invited to join the effort More information COST Benchmark venema/themes/homogenisation data venema/themes/surrogates Question slides Everyone is invited to join the effort ) Pre-processing of homogenised data Annual cycle removed before, added at the end Number of stations between and Cross correlation varies as much as possible Data is few decades long good statistics Generated networks are a long Detrended the input stations Mirrored them to more a Cropped to a Larger scale correlations are small Detected break distribution, tn, tx