10/27/2014. Content. What is all about Homogenization? What do we homogenize? Gradual inhomogeneities

Transcription

1 Current Homogenization Approaches Applied in Climate Science Strengths and Weaknesses Content The problems of homogenization (and solution paths) Direct homogenization Relative homogenization Absolute homogenization Victor Venema, University of Bonn, Germany. Enric Aguilar, Center for Climate Change, University R.V. of Tarragona, Spain. Some examples Improving through benchmarking: the COST-HOME experience Summary and problems to solve What is all about Homogenization? Homogeneous: means of the same nature (comes f rom the Greek!!) DARE,QC A homogeneous climate time series is defined as one where v ariations are caused only by variations in climate HOMOGENIZATION Conv ersely, an inhomogeneous climate time series is one which contains v ariations (biases) caused by f actors other than climate What do we homogenize? Gradual inhomogeneities Jumps and gradual inhomogeneities: special technique for gradual? Outliers?? Bad Quality Control can ruin homogenization. Typically homogenized with multiple breaks On HOME benchmark with PRODIGE no difference between gradual and breaks But not many gradual inhomogeneities Expected to be a problem for daily data

2 Getting some insights: metadata Multiple approaches are and have been used in homogenization - Homogenization strongly relies on statistical techniques to detect breakpoints - A good stations history helps to identify/confirm their occurrence - If metadata exists, it should be used to compute correction or precise date COST-HOME WG SURVEY (27). INPUT FROM 4 EXPERTS FROM 24 COUNTRIES (MOST OF THEM IN EUROPE) Source: B. Dubuisson (Météo France) The three main approaches The traffic light of homgenization methods Direct: we know were the inhomogeneity (~ breakpoint) happened and we have parallel or experimental measurements to derive a correction or we can obtain it from theoretical models. Relative: we apply detection/adjustment statistical tools over a climate time series using other highly correlated time series to try to avoid the confusion between variability and changes in the climate signal and biases produced by inhomogeneities. Needs a dense network with well correlated neighbors. Absolute: we apply detection/adjustment statistical techniques over a climate time series Direct: we know were the inhomogeneity (~ breakpoint) happened and we have parallel or experimental measurements to derive a correction or we can obtain it from theoretical models. Relative: always needed in addition to detect unknown breaks. We apply detection/adjustment statistical tools over a climate time series using other highly correlated time series to try to avoid the confusion between variability and changes in the climate signal and biases produced by inhomogeneities. Needs a dense network with well correlated neighbors. Absolute: we apply detection/adjustment statistical techniques over a climate time series Absolute homogenization Using absolute methods? (I) Absolute: we apply detection/adjustment statistical techniques ov er a climate time series Dangerous: both detection and adjustment is compromised by the climate signal. It can make the data worse (i.e. less homogeneous) than it was Method of last resort: applied with care and expertise can help to statistically identify large breaks in areas of v ery low density of stations (i.e. with lack of well correlated neighbors) Can be used to aid search f or metadata Ouesso, Rep. of. Congo: in absence of close neighbors, an absolute method was used to conf irm the obv ious inhomogeneity. No attempt to correct the data was made. Aguilar et al., 29, JGR 2

3 Using absolute methods? (II) Direct homogenization JJA DTR for Madrid, Spain (black line) and SNH T tes t values (blue line). The break, encountered at 893, closely matches the information available from metadata (-894: is metadata always true?????) Brunet et al., 26 In the homogenization of the Spanish network SDATS, the optimum detection of the introduction of the Stevenson Screen with relative methods was difficult due to the presence of the effect in most of the network Before metadata was made available, the dates were approached by applying SNHT to the DTR series Metadata confirmed that the approach was most of the time valid. Is metadata always true??? The initial adjustment factors were computed as mean differences of the 5 years before and after the break. We know were the inhomogeneity (~ breakpoint) happened and we have parallel or experimental measurements to derive a correction or we can obtain it from theoretical models. The IF approach: the best IF we now from the metadata when the breaks happened & IF we have enough information to derive a plausible correction model. They depend on the existence of parallel/experimental data and it is arguable if the results can be exported to similar settings. Screen correction from parallel measurements Weather based corrections Brunet et al., 2 NORDHOM WORKSHOP NORRKÖPING, 9-2//23 Auchmann and Brönnimann (22) Relative homogenization Relative homogenization Should always be applied Metadata cannot be assumed to be complete Comparison with neighbouring stations Removes the regional climate change signal Reduces weather noise Needs a dense network with well correlated neighbours The most applied approach: but raises lots of questions and involves solving lots of problems: Network definition Reference series / pairwise / ANOVA correlation required, number of neighbors, etc. Missing values and outliers Detection algorithm: one break at a time + hierarchical splitting or multiple breaks detection? Data resolution for detection: annual; annual + seasonal; monthly; daily Approach for the computation of adjustment factors Network-wide inhomogeneities, etc. 3

4 Relative homogenization: selecting neighbors Using pairw ise comparisons Relative homogenization relies on comparisons between highly correlated series, belonging to the same climatic area. It is necessary to define overlapping areas - either with statistical tools and/or using the climatologist expertise - as the edge ar eas might be problematic. The example shows the different regions determined for the homogenization of temperatures in France. Source: B. Dubuisson. Areas defined to homogenize monthly maximum temperature series Each station is compared with different neighbors using their common period The differences/ratios remove the climate signal Discontinuities detected by the test are due to a potential inhomogeneity either in the candidate or in the reference The break needs to be attributed by comparison and the use of subjective means (Caussinus - Mestre, HOMER-PW) or objective algorithms (PHA) Example: Otavalo, Ecuador: clear break in 2 Using composite references Inhomogeneous reference problem: joint detection Breaks? Ref series: PCA over monthly anomalies; detection: SNH T 2-4 Different approaches: weighted average, regression, PCA Dense network with good correlation is needed Example: Otavalo, Ecuador: break in 2 Computing composite reference series relays on the supposition that potential inhomogeneities may cancel out each other and the resulting series will be a homogeneous reference with no climate signal Potential problems: different length of the reference series; missing values, inhomogeneous reference series. The expected homogeneity of reference potential useful information o Two nearby stations likely have little remaining inhomogeneitie s Composite reference o Compute a weighted average of neighbouring stations o Reduces the influence of inhomogeneitie s in single stations o Careful selection of stations needed (especially for correction) o Difficult case: technological changes in many/all stations o Gaps and different lengths can lead to jumps in composite Pairwise o More noise o Can handle references with different length elegantly o Need to attribute the breaks found in the pairs to a station o Solution to this problem is still ad-hoc or manual, but mathematically tractable Multiple break detection Multiple breakpoint problem: Internal and External Variance HOMER uses a two factors ANOVA model and cghseg routine to simultaneously detect breakpoints and compute adjustment factors. The black triangles are results of individual candidatereference pairwise comparisons. Consider the differences of one station compared to a neighbor reference. Internal variance: within the subperiods External variance: between the means of different subperiods Break position criterion: Maximum external variance Farana Station, Guinea Conakry The green crosses are the results of the joint-detection process. 4

5 Shorter length, less certainty Influence of break variance n = years Ex c eeding probability /28 /64 /32 /6 /8 /4 n = 2 years RMS skill for: Random segmentation + Standard search For SNR = ½: Random segmentation and standard search have comparable skills. Random Lindau, R., V.K.C. Venema. On the multiple breakpoint problem and the number of significant breaks in homogenisation of climate records. Idojaras, Quart. journal H ungarian Meteorol. Service, 7, no., pp. -34, 23. Only for SNR >, the standard search is significantly better. Standard Multiple breakpoint problem More is different Maximisation of break variance plus a penalty term for no. breaks o o Article: Random segmentation includes contribution of breaks: ~half of break variance Manuscript: Future, just started: joint detection of all series simultaneously some examples Integrated approach (IMS, Israel) SDATS (22 long term series in Spain) Yzhak Yusefi, personal communication to the 8th Seminar on Homogenization and Interpolation, Budapest, 24. SNHT + interpolation of monthly factors: cannot well correlated neighbors with non-overlapping breaks for the earliest period (94 and before) 5

6 Homogenization of distribution Austria: , 57 Tmin & 54 Tmax stations Detection: PRODIGE, metadata Annual, winter and summer means Correction: SPLIDHOM (trust the skewness) Significance testing by bootstrapping A blind test of monthly homogenisation algorithms Victor Venema, O. Mestre, E. Aguilar, I. Auer, J. A. Guijarro, P. Domonkos, G. Vertacnik, T. Szentimrey, P. Stepanek, P. Zahradnicek, J. Viarre, G. Müller-Westermeier, M. Lakatos, C. N. Williams, M. Menne, R. Lindau, D. Rasol, E. Rustemeier, K. Kolokythas, T. Marinova, L. Andresen, F. Acquaotta, S. Fratianni, S. Cheval, M. Klancar, M Brunetti, C. Gruber, M. Prohom Duran, T. Likso, P. Esteban, T. Brandsma Nemec et al., 22. Intercomparison study Surrogate temperature section Compare full homogenisation algorithms Detection, correction If applicable: reference, iterations, remove outliers, etc. Benchmark dataset Monthly temperature and precipitation networks Homogenized blind Random small inhomogeneities o Gaussian distribution Sections o Real (inhomogeneous) climate records o Synthetic data o Surrogate data Generated homogeneous surrogate temperature networks Based on statistical properties of homogenized data 5 networks Length: years Number of stations: 5, 9, 5 Missing data Missing data Beginning WWII 6

7 Outliers Breaks Breaks Local trends Simulataneous Participants returned the data Scatterplots monthly CRMSE 25 blind contributions Some algorithms multiple contributions o Test versions o Test influence operator (manual methods) Algorithms/software o USHCN o RhTestV2 o PRODIGE o SNHT o MASH o Climatol o Craddock o ACMANT o AnClim.5 ACMANT.5.5 MASH main.5.5 PRODIGE monthly.5.5 C3SNHT.5.5 USHCN main PMFred abs

8 Monthly CRMSE complete contributions Temperature Inhom. data MASH main PRODIGE main PRODIGE monthly PRODIGE trendy USHCN main USHCN 52x USHCN cx8 AnClim main PMTred rel PMFred abs C3SNHT SNHT DWD ACMANT CRMSE [ C] Precipitation Decadal CRMSE complete contributions Temperature Inhom. data MASH main PRODIGE main PRODIGE monthly PRODIGE trendy USHCN main USHCN 52x USHCN cx8 AnClim main PMTred rel PMFred abs C3SNHT SNHT DWD ACMANT CRMSE [ C] Precipitation Inhom. data Errors in trends Inhom. data MASH main PRODIGE main PRODIGE monthly PRODIGE trendy USHCN main USHCN 52x USHCN cx8 AnClim main PMTred rel PMFred abs C3SNHT SNHT DWD ACMANT Trend difference [ C/a] Contingency scores Pierce Heidke Skill Heidke Special Contribution No stations POD POFD Skill Score Score PRODIGE main PRODIGE monthly PRODIGE trendy USHCN main USHCN 52x USHCN cx AnClim main icraddock Vertacnik PMTred rel PMFred abs C3SNHT SNHT DWD Climatol ACMANT Lessons learnt from COST-HOME experiment (I) Lessons learnt from COST-HOME experiment (II) Moderate correlation between error metrics Contingency scores not good predictor of skill Artefact of validation: good algorithms detect more small breaks, but with inherent uncertainty in position Indicate that solving combinatorial problem for large breaks, more important as finding many breaks Relative homogenisation improves temperature records Absolute homogenisation can make data more inhomogeneous Regional climate variability may not have been realistic Best algorithms (ACMANT), Craddock, MASH, PRODIGE, USHCN Automatic algorithms among the best Function with an inhomogeneous reference Multiple breakpoint methods 8

9 Improvements to benchmarking over HOME Uncertainties Size of the breaks in HOME too large Network density like Europe o Much of the world SNR will be less Length of the series not varied Benchmark more comparable to real data o ISTI benchmarking cycle These points will be fixed in the benchmarking of the International Surface Temperature Initiative o Benchmarking gives qualitative idea of uncertainties Not much work on computing uncertainties from remain in homogeneities Multiple sources of uncertainty: Not all breaks in the candidate station can be detected Uncertainty in the estimation of correction parameters due to insufficient data Uncertainties in the corrections due to remaining inhomogeneities in the references The date of the break may be imprecise (see Lindau & Venema, 23b) Absolute methods would have larger uncertainties Summary: Issues in homogenization Fundamental problems to solve by the homogenization community (by us!) Direct methods should be applied if possible Relative methods preferred to absolute methods. Depending on your data and network, a different solution must be applied Direct methods have strong data demands and absolute methods are a last resort. Most of the time, we will be using relative homogenization Homogenization is hampered by missing data and different lengths of record Unidentified quality control problems can compromise the results of homogenization In sparse networks and in the old parts of the records it is difficult to find adequate neighbors to apply relative methods. This problem is worst in daily data, where most sophisticated techniques require very high correlations Small inhomogeneities (compared to the large variance in the (difference) series) are difficult to detect Gradual inhomogeneities are difficult to detect and sometimes their size is small compared to other breaks It is necessary to improve validation and uncertainty estimates Need uncertainty estimates Trend small compared to breaks Missing data (middle) and incomplete data (especially for composite) Density of networks (SNR) Multiple breakpoints Inhomogeneous references (joint homogenization) 9