1 Workshop Climagri Agricoltura e cambiamenti climatici Analisi, incertezze, controversie, interdipendenze Ancona, 27 e 28 giugno 2005 Facoltà di Agraria Università Politecnica delle Marche Aula A. Bartola SERIE STORICHE DI DATI OSSERVATIVI E DA SCENARIO: INCERTEZZE E CONTROVERSIE NEL CONTESTO DELLA RICERCA CLIMATOLOGICA Maurizio Maugeri Università di Milano
2 Workshop CLIMAGRI Agricoltura e cambiamenti climatici - analisi, incertezze, interdipendenze e controversie Serie storiche di dati osservativi e da scenario: incertezze e controversie nel contesto della ricerca climatologica Ancona, 27 giugno 2005 Maurizio Maugeri Istituto di Fisica Generale Applicata Università degli Studi di Milano
3 Introduction: data quality and homogeneity Data quality and homogeneity are very capital factors in any climate-change-related issue. They are at the moment among the most widely debated and controversial topics in the field of climatology.
4 Introduction: data quality and homogeneity A good example of the problems related to observational data is presented by Redder et al., Radiosonde data, which are generally assumed to be unbiased, are often used to determine and remove biases in satellite observations. Radiosonde temperature data over the United States are shown to have significant and unexplained inhomogeneities in the mid-troposphere. Redder et al., 2004: Unexplained discontinuity in the U.S. radiosonde temperature data, J. Atm. Oc. Tech., vol. 21,
5 Fu et al., 2004: Contribution of stratospheric cooling to satellite-inferred tropospheric temperature trends, Nature, vol. 429, Introduction: data quality and homogeneity Another example highlighting how, in climate analyses, data quality must be continuously brought into question, in order to draw conclusions as more solidly as possible is given by Fu et al., They address the inconsistencies between temperatures observed globally by the mid-tropospheric channel of satellites (channel two) Microwave Sounding Unit (MSU) showing a warming trend of less than 0.1 K per decade and surface temperatures based on the in situ observations (showing a larger trend, of more than 0.17 K per decade). This trend difference - they conclude - is mainly due to differences in data adjustments related to instrument calibration and diurnal drift correction. The problem probably depends on the fact that although the original purpose of MSU measurements was to improve weather forecasts the data are often used to satisfy climate research requirements. So, even if a continuing data analysis effort has been made to assure homogeneity, there are probably a lot of open questions concerning this topic.
6 Introduction: data quality and homogeneity The scientific community is well aware that any reconstruction of past climate variability and change must rely on the most solid, reliable and homogeneous data-sets
7 Soon et al., 2004: Estimation and representation of long-term (>40 year) trends of Nothern- Hemisphere-gridded surface temperature: A note of caution, Geophys. Res. Lett.., vol. 31, L03209, doi:10.129/2003gl Introduction: the problems beyond the data Problems are not always connected with the data themselves, but they can also be related to the meaning we attribute to the data: problems of misinterpretation. Accurate communication of methods and avoidance of data-padding procedures for smoothing and/or filtering of climatic time series should be incorporated in reporting data trends (Soon et al., 2004).
8 McIntyre, and McKitrick, 2003: Corrections to the Mann et al 1998 proxy data base and Northern hemispheric average temperature series Energy & Environment 146, Introduction: the problems beyond the data In some cases the problems depend both on the data and on the analytical techniques. A good example is related to the work performed by Mann et al. (2003 and previous papers) on a large amount of proxy records. Stephen McIntyre and Ross McKitrick claimed various errors in Mann's research, but McIntyre and McKitrick offered no explanation as to why their analysis also differs from other reconstructions. In turn, Mann (supported by Tim Osborn, Keith Briffa and Phil Jones of the Climatic Research Unit) has disputed the claims made by McIntyre and McKitrick, saying they have made critical errors in their analysis that have the effect of grossly distorting the reconstruction of Mann et al., Mann et al., 2003: Global surface temperatures over the past two millennia, Geophys. Res. Lett., vol. 30, 1820, doi:10.129/2003gl
9 McIntyre and McKitrick: Hockey sticks, principal components, and spurious significance. Geophys. Res. Lett., vol. 32, L03710, doi: /2004gl021750, Introduction: the problems beyond the data In 2004 Mann, Bradley, and Hughes published a corrigendum to their article, correcting a number of mistakes in the online supplementary information that accompanied their article but leaving the actual results unchanged. M&M have published another Geophysical Research Letters article on February 12th, 2005, claiming that the "Hockey Stick" shape was a result of a programming error, and that using the same steps like Mann et al., they were able to obtain the Hockey Stick graph in 99 percent of cases even if red noise was used as input. Mann and his collaborators have responded to articles by Stephen McIntyre and Ross McKitrick via various means, including the blog Real Climate reconstructions with significantly more variability than the reconstructions of It is also to be noticed that Moberg et al. have recently generated Mann et al., Mann et al.: Corrigendum: Global-scale temperature patterns and climate forcing over the past six centuries, Nature, 430, 105, 2004.
10 Introduction: what can we conclude? Let s focus on the data First Postulate of Any Researcher involved in studying past climate variability and change: It is not possible to consider any record of climatic data as entirely reliable and homogeneous It is quite likely that the situation will stay as such for some time in the future.
11 NCEP: US National Center for Environmental Predictions Inhomogeneities in Re-Analyses It is to be noticed that also the most widely used globala data-sets (e.g. NCEP reanalyses) are affected by significant biases
12 NCAR/NCEP RE-ANALYSIS Sturaro G., 2003: A closer look at the climatological discontinuities present in the NCEP/NCAR reanalysis temperature due to the introduction of satellite data, Climate Dynamics, 21,
13 NCAR/NCEP RE-ANALYSIS Sturaro G., A closer look at the climatological discontinuities present in the NCEP/NCAR reanalysis temperature due to the introduction of satellite data, Climate Dynamics, 21,
14 M. Maugeri, M. Brunetti, F. Monti, T. Nanni The Italian air force sea level pressure data set ( ). Il Nuovo Cimento, 26C, NCAR/NCEP RE-ANALYSIS: ITALIAN SLP Italian SLP NCAR/NCEP Re-analysis UK Met Office
15 NCEP-ECMWF-MSU Stendel et al. 2000, Assessing level of uncertainty in recent temperature time series, Climate Dynamics, 16, We do not know which of the data set (if any) is correct. There is, however, evidence that inhomogeneities can be identified in both reanalysis, but this does not imply that there are no problems with the MSU
16 Stendel et al. 2000, Assessing level of uncertainty in recent temperature time series, Climate Dynamics, 16,
17 Inhomogeneities in Surface data The awareness of the importance of data quality and homogeneity issues for a correct detection of climate change has increased rapidly in the last few years. Most of the contributions concern upper air data (see e.g. Lanzante et al., 2003; Fu et al., 2004), however errors and inhomogeneities also concern surface ones. At surface level it is often assumed that such inhomogeneities have random distribution and that, considering a sufficiently large number of series, average records with negligible biases can be obtained. This assumption is likely to be correct if global or hemispheric averages are considered (is it really so?), but it may not be correct at a regional scale. Lanzante et al., 2003: Temporal Homogenisation of Monthly Radiosonde Temperature Data. Part I: Methodology, J. Climate, vol. 16, Fu et al., 2004: Contribution of stratospheric cooling to satellite-inferred tropospheric temperature trends, Nature, vol. 429,
18 An interesting example of this problem is given by Böhm et al. (2001) in a paper investigating temperature variability in the Alps and their surroundings. They subjected about 100 secular temperature records of this area to a detailed quality control and homogenisation procedure and performed a systematic comparison between the original and the corrected records. The results clearly display that the original series are biased by non-climatic noise and, even if the average over all the series is considered, the original data display an error of the long-term amplitude of the temperature evolution in the region of about 0.5 K.
19 Estimated bias (K) of the average over all Italian records Adapted from: Böhm, R., Auer, I., Brunetti, M., Maugeri, M., Nanni, T., Schöner W., 2001: Regional Temperature Variability in the European Alps from homogenised instrumental time series. Int. J. Climatol., 21,
20 Estimated biases (K) of the averages over all Austrian (right graph) and Swiss (lower graph) records
21 Homogeneity Homogenisation of regional data-sets: introduction Climate variations Measuring problems Signals in the records of meteorological data Measuring problems Relocations Instrumental errors (changes of the instruments and/or recalibrations) Observation methods Screenings Changes in the environment around the station
22 So, at regional level, systematic biases in the original records can hide a significant part of the long-term temperature trend Second study: homogeneity testing is performed in regional subgroups of 10 series using a procedure that rejects the a priori existence of homogeneous reference series and tests each series against each other. In this case important adjustments are made. What are the consequences? Let s show another example: Two different North Italy regional average records are given by Brunetti et al. (2000) and by Böhm et al. (2001). First study: the homogeneity of the station records is tested comparing each record with a weighted average over all available records (Alexandersson, 1986) and no corrections are made.
23 North Italy long-term temperature evolution (filtered curves) in the period according to Brunetti et al. (2000) and Böhm et al. (2001) 1.0 Po Valley (Boehm et al., 2001) NITA (Brunetti et al., 1999) Adapted from: Brunetti, M., Buffoni, L., Maugeri, M., Nanni, T., 2000: Trends of minimum and maximum daily temperatures in Italy from 1865 to Theor. Appl. Climatol., 66, and Böhm, R., Auer, I., Brunetti, M., Maugeri, M., Nanni, T., Schöner W., 2001: Regional Temperature Variability in the European Alps from homogenised instrumental time series. Int. J. Climatol., 21,
24 What are the linear trends? Po Valley (Boehm et al., 2001) North Italy (Brunetti et al., 1999) Trends etimated by linear regression 1.09 K / 100 years 0.59 K / 100 years Conclusion: At least at regional level, the correct detection of trends in long term series of meteorological variables requires a very critical analysis of the homogeneity of the observations
25 Homogeneity testing and homogenisation The problem is not easy to manage Meteorological series can be tested for homogeneity and homogenised both by direct and indirect methodologies. The direct approach is based on objective information that can be extracted from the station history or from some other sources, the indirect one uses statistical methods, generally based on comparison with other series. Both direct and indirect methodologies have remarkable limits
26 Homogeneity testing and homogenisation Direct methodologies are not easy to use as: 1) it is generally very difficult to recover complete information on the history of the observations (metadata); 2) even when available, metadata hardly give quantitative estimates of the inhomogeneities in the measures. Also indirect methodologies have important deficits: 1) they require some hypotheses about the data (e.g. homogeneous signals over the same region); 2) inhomogeneities and errors are present in all meteorological series, so it is often difficult to decide where to apply corrections and, when the results are not clear, the risk of applying subjective corrections is very high.
27 Homogeneity testing and homogenisation How to overcome the intrinsic limit of indirect homogenisation methods is, at present, still an open question. The possibilities range from homogenising all suspect periods, to correcting the series only if the results of the statistical methods are very clear and also supported by metadata. So, at present, a universal approach to the problem is still missing
28 The approach of the Milan University ISAC CNR Research group Around the mid 1990s, a group of Italian researchers set up a wide research program with the aim of gaining a better understanding of the evolution of the Italian climate in the last 100/150 years. Within this program we have developed a methodology to manage the quality and homogeneity problem
29 Getting started with homogenisation: the preliminary quality checks Example: daily temperatures 1. Removal of absurd daily values, e.g. t_max < -15 C or t_max > 49 C; t_min < -25 C or t_min > 45 C 2. Flipping of the values whenever t_max < t_min (check with neighbouring stations) 3. Check with neighbouring stations whenever daily temperature range DTR > 25 C 4. Removal of daily temperatures if their absolute value differs for more than 10 C from the ones of neighbouring stations 5. Check with neighbouring stations if t_max or t_min stays the same for 5 or more consecutive days 6. Check with neighbouring stations if t_max and t_min stays the same for 3 or more consecutive days 7. Check with neighbouring stations if t_max = t_min
30 ...but the best is to check each individual datum An example from HISTALP Date stat suspect value ori (mm) suspect value hom (mm) suspect value (% of ) suggested by map (mm) sugested by data provider (mm) definitively corrected value (mm) data provider ORANGE MF NANCY MF ARLES MF DIJON MF VIX > MF CLUNY to MF NANCY MF BESANCON MF Checking outliers by representing spatial anomalies patterns
31 Checking outliers by representing temporal evolution gen 1884 gen 1889 gen 1894 gen 1899 gen 1904 gen 1909 gen 1914 gen 1919 gen 1924 gen 1929 gen 1934 gen 1939 gen 1944 gen 1949 gen 1954 gen 1959 gen 1964 gen 1969 gen 1974 gen 1979 gen 1984 gen 1989 gen 1994 gen 1999 gen DIFF-pad DIFF-bel DIFF-tri DIFF-rem DIFF-par DIFF-ven DIFF-fer DIFF-bol DIFF-man Rovigo T_max
32 Our methodological approach to homogenisation 1) Collecting as many data and information (metadata) as possible: History of the National network and of the involved institutions; History of instruments and observation methods; History of every meteorological station 2) Performing a first homogenisation by means of direct methodologies; 3) Performing final homogenisation by means of indirect methodologies
33 Homogenisation: the metadata Get a clear picture of: history of the observation (i.e. to understand the evolution of the meteorological network and to reconstruct the history of each available station in the data-set); Data sources and actual data availability. The reconstruction of the network evolution is not only important to get a picture of the data quality but is also useful to check data homogeneity, as homogenisation is generally based on statistical methods. This methods often fail in identifying breaks that affect a high fraction of stations within a short period. This happens when breaks are due to changes in instruments and methods caused by new standards imposed by the network management, for example, as a consequence of new national or international standards
34 Corrections by means of metadata: an example hpa /1 1/2 1/3 1/4 1/5 1/6 1/7 Date 1/8 1/9 1/10 1/11 1/ hpa Maugeri, M., Buffoni, L., Delmonte, B., Fassina, A., 2002: Daily Milan temperature and pressure series ( ): completing and homogenising the data, Climatic Change, 53, /1 1/2 1/3 1/4 1/5 1/6 1/7 Date 1/8 1/9 1/10 1/11 1/12 Corrections applied to Milan daily air pressure data to eliminate the bias introduced by calculating daily means using observations taken at A: 8 a.m., 2 p.m. and 7. p.m. and B: sunrise and mid-afternoon. The corrections A apply to the period December 1 st, December 31 st, 1987, corrections B to
35 Homogenisation by means of indirect methods The indirect methods make use of meteorological data from neighbouring stations. Formally, data of a given series can be represented as a sum of more terms. Be X(t) the meteorological variable s value X at the time t. Therefore it can be written: X(t) = N + A(t) + IH(t) (t = 1, 2,..., n) (1) where N is X s normal value (it is defined by considering the mean value over a suitable time interval like, for example, the period ), A(t) is the anomaly related to the instant t (it defines the departure of the variable X from its normal value) and IH(t) is the possible inhomogeneity lying in the measured value X(t) (in the simplest case, IH(t) is a step function that equals to 0 until the inhomogeneity-inducing event takes place, and then that equals to a constant value which represents the effect of the inhomogeneity in fact). By using an analogous notation, a reference series which is constituted, for example, by the data of a neighbouring station can be written as follows:
36 Homogenisation by means of indirect methods R(t) = N + A (t) + IH (t) (t = 1, 2,..., n) (2) If the two series belong to the same climatic area, it can be assumed that A(t) = A (t) for each value of t. Moreover, if you postulate the reference series as homogeneous, it will be always true that IH (t) = 0. Therefore, the series of the differences will be: Z(t) = X(t) - R(t) = (N - N ) + IH(t) (t = 1, 2,..., n) (3) In other terms it can be assumed that, unless there are possible inhomogeneities, the series of the differences must result as constant. The same approach is followed for the series of the ratios. The latter approach is particularly used for precipitation series. Possible deviations from Z(t) constant path are therefore assumed as being due to inhomogeneities.
37 Homogenisation by means of indirect methods The application of indirect methodologies is actually much more complicated than what the previous relations seem to suggest. In fact, whenever in a relation like the (3) there is a signal which is characterised by one or more steps, it is usually very hard to understand whether it is due to the station under exam or to the station used as a reference. Then, if you consider not too short periods, it can also happen that both the stations present some significant inhomogeneities and that there are several step-shaped signals. So, the question of the identification of a reference series is actually very problematic
38 How do we select the reference series? A procedure that rejects the a priori existence homogeneous reference series is used. Each series is tested against each other series in subgroups of 10 series. Subsequently, the break signals of one series against all others are collected in a decision matrix and the breaks are assigned to the single series according to metadata and/or to probability. of
39 How do we compare the test and the reference series? The comparison between a test series and a reference series can be performed by a number of different mathematical techniques. We use of them: the Craddock homogeneity test
40 Homogenisation: the Craddock statistical test One among the most commonly used statistical tests is the Craddock test. At first it was developed for analysing the precipitation series and subsequently it has been widely updated, improved and extended to thermometric records. It accumulates the normalized differences between two series (a and b) according to one of the following formulas: s n = s + a + ( b a ) b b = s 1 + a b (4B) m n 1 n m m n (4A) n n n n am s where the mean values of the series are calculated over the entire period in which the comparison is performed and where the choice of the proper formula depends on the underlying hypothesis, such as on considering as a constant the difference either the ratio between stations of the same area. See also the example presented on the craddock.xls Excel File.
41 All results are displayed in the Excel files Craddock_TMED_1+2 and Craddock_PREC_1+2. In order to display the ability of the Craddock homogeneity test to identify some typical inhomogeneities, we have made use of records generated by means of random numbers. In particular, we have generated some records with the features of Milan yearly mean temperature and yearly total precipitation. TEMPERATURE Series length: 240 data Average: 13.3 C St. Dev.: 0.9 C PRECIPITATION Series length: 240 data Average: 1015 mm St. Dev.: 202 mm Then we have applied the Craddock test to A) some pairs of completely random temperature/precipitation records and B) some pairs of records obtained partially from random series and in part from the series to test itself (i.e. we introduce a 0.7 correlation between the pair of series to subject to the Craddock homogeneity test). Then we have added to the series some typical errors as step functions, trends,
42 Homogenisation: statistical test and metadata Craddock test - Bologna precipitation record All inizio del 1857 a questo pluviometro, 6000 ridotto in cattivo stato pel lungo uso, ne venne sostituito un altro di migliore costruzione, e lavorato con molta precisione News about a damage to the pluviometer. In corrispondence with repairing the damage, the cause of the underestimation of precipitation has been removed for the period Change in data origin: from Osservatorio Astronomico to Istituto Idrografico Introduction of a new pluviometer (Fuess recorder):... fu collocato a cura del prof Bernardo Dessau nel periodo CRADD-FER CRADD-VAL CRADD-FIR CRADD-ARE CRADD-PIA CRADD-PAD CRADD-PAR CRADD-REM CRADD-MAN CRADD-ROV
43 Basic problems: what has to be corrected? How to correct? a) All the periods found by statistical methods b) Only the periods for which there is evidence in metadata The problem is, in part, still open
44 Basic problem: what has to be corrected? Our methodology: When the signal is not so clear our philosophy is to homogenize the data only in the following cases: 1) when there is some information in the metadata 2) when more reference series give coherent adjustment estimates and their scattering around the mean value is lower than the break amount In our opinion, only in these cases the corrections really improve the data quality, whereas in other cases there is a high risk of introducing corrections whose associated errors are higher than the corrections. The point concerning the fundamental question of when actually perform correction is an important open issue of any research concerning the reconstruction of the past climate.
45 Basic problem: how to correct? Once we decide to correct one break, the series used to estimate the adjustments are chosen among the reference series that result homogeneous in a sufficiently long sub-period centred on the break year, and that well correlate with the candidate one. We chose to use several series to estimate the adjustments to be sure about their stability and to avoid unidentified outliers in the reference series from producing bad corrections. Moreover, it often happens that homogeneous sub-intervals between two detected breaks are so short that the signal-to-noise ratios of the adjustments obtained with only one reference series are very low. So, the use of more series allows us to correct a great number of short sub-periods that would have to be left unchanged otherwise. The adjustments from each reference series are calculated on a monthly basis, and then they are fitted with a trigonometric function in order to smooth the noise and to extract only the physical signal (the adjustments often follow a yearly cycle). The benefits of using smoothed adjustments instead of the rough ones are well described in Auer et al. (2005). The final set of monthly adjustments is then calculated by averaging all the yearly cycles, excluding from the computation those stations whose set of adjustments shows an incoherent behaviour compared with the others. When a clear yearly cycle is not evident, the adjustments used to correct the monthly data are chosen as constant through the year and are calculated as the average among the monthly values for temperature, and as the weighted average for precipitation, where the weights are the ratios between monthly mean precipitation and total annual precipitation.
46 Why is homogenisation so difficult? Graphs correlation-distance 160 mean decorrelation of precip-series of different time-resolution to a common variance of 50% (after Scheifinger, 2003) km daily monthly seasonal annual Brunetti, M., Maugeri, M., Monti, F., Nanni, T.: Temperature and precipitation variability in Italy in the last two centuries from homogenized instrumental time series, Int. J. Climatol., in press.
47 Why is homogenisation so difficult? Some examples of the results of systematic homogenisation activity: H I S T A L P - T E M P E R A T U R E D A T A S E T : B R E A K A N A L Y S I S series area km 2 area relative to E urope 6.9 % available data (incl. closed gaps) years available data (incl. closed gaps) m onths m ean length of series years detected breaks (total) 711 m ean hom ogeneous subperiod 23.1 years m ean num ber of breaks per station RM S -break K m axim um positive break 4.6 K m axim um negative break -5.1 K detected real outliers 4175 m ean outlier rate 1.8 % corrected overshooting adjustm ents 1663 m ean overshooting rate 0.7 % closed gaps m ean gap rate (% ) 5.8 % CHA NG E S V E RS US 2001-RE LE A S E series avialable years m ean length release release without length changes with additional early parts new series 6085 additional years Updated from: Böhm, R., I.Auer, M.Brunetti, M.Maugeri, T.Nanni and W.Schöner 2001: Regional temperature variability in the European Alps from homogenized instrumental time series. Int. J. Climatol. 21:
48 Homogenisation (Italy): statistics of the breaks Mean T Maximum T Minimum T Precipitation N. of years (excluding filled gaps) N. of breaks N. of break per series N. of break per year per series Mean homogeneous sub-period (years) Brunetti, M., Maugeri, M., Monti, F., Nanni, T.: Temperature and precipitation variability in Italy in the last two centuries from homogenized instrumental time series, Int. J. Climatol., in press.
49 Homogenisation: number of detected breaks (ITALY) Number of detected breaks per year. From a) to d) absolute values (for mean temperature, maximum temperature, minimum temperature and precipitation respectively); form e) to h) in relation to the available series (for mean temperature, maximum temperature, minimum temperature and precipitation respectively) Brunetti, M., Maugeri, M., Monti, F., Nanni, T.: Temperature and precipitation variability in Italy in the last two centuries from homogenized instrumental time series, Int. J. Climatol., in press.
50 TEMPERATURE HISTALP_TEMPERATURE 350 HISTALP-T01: Frequency distribution of detected and removed breaks permille minimum: -5.1 deg maximum: +4.6 deg deg C (class width 0.5 deg) histalp T01: Detected and removed breaks per year breaks per year nr. of stations Updated from: Böhm, R., I.Auer, M.Brunetti, M.Maugeri, T.Nanni and W.Schöner 2001: Regional temperature variability in the European Alps from homogenized instrumental time series. Int. J. Climatol. 21:
51 TEMPERATURE HISTALP_TEMPERATURE Böhm, R., I.Auer, M.Brunetti, M.Maugeri, T.Nanni and W.Schöner 2001: Regional temperature variability in the European Alps from homogenized instrumental time series. Int. J. Climatol. 21:
52 Homogenisation: mean annual adjusting series (ITALY) Mean annual adjusting series obtained by calculating the yearly average differences (ratios) between the homogenised and the original temperature (precipitation) series. a) Mean temperature, b) Maximum temperature, c) Minimum temperature, and d) Precipitation. Standard deviations (thin lines) and total correction ranges (dotted lines) are indicated too. The standard deviations were not calculated before 1803 for mean temperature and before 1814 for minimum and maximum temperature, due to a sample size less than 5 stations Brunetti, M., Maugeri, M., Monti, F., Nanni, T.: Temperature and precipitation variability in Italy in the last two centuries from homogenized instrumental time series, Int. J. Climatol., in press.