During the last decades due to the increased interest about climate change, many studies have been conducted trying to detect shifts in climatic series. The necessity of the homogenization of meteorological observations becomes obvious to all these studies. In practice, inhomogeneities are hardly ever avoided, because the meteorological station networks are constantly changing. A myriad of methods for detecting and adjusting inhomogeneities in climate series have been developed. In this study two homogeneity methodologies, namely MASH and Climatol, were applied to 49 monthly temperature series of synoptic stations, covering almost all climatic zones of Greece, belonging to the operational weather network of the Hellenic National Meteorological Service (HNMS). Time series cover periods ranging from 35 to 45 years. Only 8.2% of the stations passed both tests successfully, another 10.2% passed the MASH homogeneity test successfully without any breakpoint, but not the Climatol test. On the other hand, 14.3% of the stations passed the Climatol test successfully but not the MASH test. The remaining stations presented one or more breaks or outliers in both homogeneity methods. Due to lack of metadata only 15% of the breaks could be explained by the stations' history. Station relocations as well as changes in observation practices caused most of the temperature anomalies. The adjustments in seasonal series were in the range from −2.54 to 1.39 °C for MASH and from −2.30 to 1.50 °C for Climatol. Seasonal and annual mean temperature trends were analyzed and their statistical significance calculated. The most pronounced seasonal trends were recorded in summer. Also, the differences of climatological normals for the period 1961–1990 between raw and homogenized annual series were computed. The absolute values of differences ranged from 0.0 to 0.8 °C for MASH and from 0.0 to 1.0 °C for Climatol.
Climate variability and change are in the epicenter of global interest following assessments that most of the temperature change observed over the last 50 years can be attributed to anthropogenic impacts (IPCC, 2007). Calculations for the detection of climate change require reliable and good quality long-term time series. However, the reliability of these datasets strongly depends on their homogeneity. A climate time series is defined as homogeneous when its variations are caused only by variations in weather and climate (Peterson et al., 1998). Unfortunately, the majority of climate records have been affected by a number of nonclimatic factors such as station relocations, changes in the instrumentation and recalibrations, new formulae used to calculate mean temperature, changes in land use, changes in observation practices, etc. (Peterson et al., 1998), making these datasets unrepresentative of the actual climate variation. Quite often the magnitude of these biases is as large as the climate variation signals that we try to detect (Della-Marta et al., 2004).
Numerous homogenization techniques have been developed so far in order to detect inhomogeneities on monthly time series (Easterling and Peterson, 1995; Alexandersson and Moberg, 1997; Peterson et al., 1998; Vincent, 1998; Szentimrey, 1999; Tuomenvirta, 2001; Lund and Reeves, 2002; Ducre-Robitaille et al., 2003; Wang, 2003; Caussinus and Mestre, 2004; Menne and Williams, 2009; Guijarro, 2011). Also several studies concerning the homogenization of long-term time series were performed all over the world (Böhm, 1998; Böhm et al., 2001; Vincent et al., 2002; Alexandrov et al., 2004; Della-Marta et al., 2004; Aguilar et al., 2005; Auer et al., 2005; Begert et al., 2005; Staudt et al., 2007; Kuglitsch et al., 2009; Zhen and Zhongwei, 2009, Stastna, 2010). Nevertheless until this date the relative studies over the Greek area are very few. Greece is a small country but very interesting from the point of view of climate conditions because of the variety of ground morphology and because it is located in the Mediterranean, an area where it is anticipated that the impact of climate change will be important (IPCC, 2007).
The aim of this work is to detect abrupt changes in mean monthly temperature series from almost the whole Greek meteorological stations network, using two different methods, to homogenize these time series and to identify roughly differences (e.g. in annual series, in seasonal trends, in climatological normals, in the climatic type). Data are described in Section 'Geographical information and data'. Section 'Methodology' presents the methodology. The results are discussed in Section 'Results' and finally, the conclusions are summarized (Section 'Conclusions').
2. Geographical information and data
2.1. Geographic location and geomorphology
Greece occupies the southernmost end of the Balkan Peninsula and lies approximately between latitudes 34° and 42°N and longitudes 19° and 30°E. The country is surrounded on the east by the Aegean Sea, on the west by the Ionian Sea and on the south by the Libyan Sea. The country can be divided in three main geographical areas, the mainland, the islands and the Aegean basin and has a total area of 131 957 km2 (National Statistical Service of Greece (NSSG), 2008). The mainland covers about 80% of the total area while the remaining 20% is divided among nearly 6000 islands and islets. Due to its highly indented coastline and a vast number of islands, Greece has the 12th longest coastline in the world with 15 021 km in length, while the length of its land boundary is 1180.71 km.
Despite the numerous islands, two-thirds of Greece are largely covered by mountains of medium height (highest mount Olympos 2917 m) making the country one of the most mountainous in Europe. As a result, the country presents a considerable climatic variability.
The climate of Greece is predominantly Mediterranean, with mild, wet winters and warm, dry summers. However, due to the country's unique geography, Greece has a remarkable range of microclimates and local variations (Aiginitis, 1907; Mariolopoulos, 1938; Carapiperis, 1963; Zambakas, 1981).
The cold and rainy period lasts from the mid of October until the end of March, and the warm and dry season from April until September.
The average mean winter temperature varies from 6.0 to 11.0 °C in the mainland, from 9.0 to 13.0 °C in coastal areas and from 2.0 to 6.0 °C in northern Greece (for the needs of this work, averages for the period 1960–2004 used as reference were calculated). January and February are generally the coldest months. However, in the end of January and during the first fortnight of February, a period of sunshine with mild weather and calm winds often prevails, known as ‘halcyon days’.
During the warm and dry period, the sky is clear and the sun is bright. However, there are scarce intervals with showers or thunderstorms of small duration mainly in mainland areas. The average mean summer temperature ranges from 26.0 to 28.0 °C in the mainland, whereas it is approximately 1.0 °C lower in islands and 2.0–3.0 °C lower in the north (for the needs of this work, averages for the period 1960–2004 used as reference were calculated). Highest temperatures are observed during the last 10-day period of July and the first 10-days of August. However, the high temperatures are dampened from the fresh sea breezes in the coastal areas and from the north winds blowing mainly in the Aegean – a combined result of a depression over Asia Minor and an anticyclone over the Balkans – also known as ‘Etesians’.
A set of monthly temperature series from 49 Greek stations for the period 1960–2004 has been used. The main geomorphologic characteristics of Greece and weather types affecting the country are such that, the selected stations are representative and cover almost all climatic zones except the mountainous areas with polar climate. Time series cover periods ranging from 35 to 45 years. The first two figures are dedicated to the description of raw data series. Figure 1 shows the overall number of available time series and the histogram in Figure 2 shows the frequency of mean temperature values. Table 1 shows the names of the stations, their location and altitude. All weather data used here were provided by the operational weather network of the Hellenic National Meteorological Service (HNMS) and only for the purposes of this study. These data are available after request; fees may apply.
Table 1. Geographical coordinates and altitude of the 49 stations
Latitude (decimal degrees)
Longitude (decimal degrees)
For the needs of this work, the Greek area is divided into climatically homogeneous subregions. Due to the variable topography of Greece the climatic classification is not easy. Taking into account the Köppen climate classification (Köppen, 1918) as well as the correlation between stations, seven regions with similar climate characteristics resulted. All stations in a region are highly correlated with correlation coefficients of daily data higher than 0.9. Also, to avoid the impact of inhomogeneities, correlation coefficients of monthly data from the first differences of the series were computed and all available pairs of observations have been used. These coefficients were not lower than 0.8 even for distances between stations of the order of 450 km.
Homogenization methods were applied separately on the datasets of each of the above climatic regions. Climatic regions are presented in Figure 3: the first region (A) comprises four stations located in Central and Eastern Macedonia (Alexandroupoli, Kavala, Serres, Macedonia airport), the second region (B) includes the mountainous stations (Florina, Kozani, Ioannina, Desfina, Tripoli), the third region (C) includes stations from the western areas (Corfu, Aktio, Agrinio Zakynthos, Argostoli, Patra, Pyrgos, Araxos, Andravida, Kalamata, Methoni), region (D) includes stations located in Central Greece and Thessaly (Larisa, Trikala, Aghialos, Lamia, Aliartos, Tanagra, Tatoi, N. Filadelfeia, Eleysina, Piraeus, Helliniko), region (E) includes stations located in the North and Central Aegean (Limnos, Skyros, Skopelos, Naxos, Milos, Kythira), region (F) includes stations located in the Eastern Aegean (Mytilini, Chios, Samos, Kos, Rhodos) and the last region (G) includes stations located in Crete and South Aegean (Souda, Rethymno, Heraklio, Ierapetra, Tympaki, Siteia, Karpathos).
Different statistical tests can be used for the detection of artificial changes or inhomogeneities in weather time series. Some old methods relied on tests checking the nonstationarity of a single climatological series. These absolute methods must be avoided, because they are based on the unrealistic assumption that climate is stable (Guijarro, 2011). Relative homogenization methods can be used instead in which stationarity tests are applied not on individual series, but on series of ratios or differences between the station under study and one or more reference stations.
The relative homogeneity methods used in this study are two: (1) multiple analysis of series for homogenization (MASH) and (2) Climatol.
3.1. MASH method
The MASH method is developed in the Hungarian Meteorological Service (Szentimrey, 1996; Szentimrey, 1999; Szentimrey, 2000; Szentimrey, 2008). This is a relative homogeneity test procedure based on multiple comparisons between the climatically similar series and does not assume a homogenized reference series. The time step of comparisons may be annual, seasonal or monthly. Also, the software is developed for both automatic or manual processing. During the automatic procedure of MASH the user can select only the month or season to be homogenized and the number of iteration steps. In the manual mode the user decides the candidate station for homogenization, excludes a reference station (e.g. bad test statistics of reference station) if necessary, checks the results through graphical output, examines test statistics before and after homogenization and accepts or not a given breakpoint. The process can be repeated if necessary (e.g. remaining bad test statistics after homogenization) and a breakpoint can be corrected manually if necessary. In this study the version MASH 3.02 and the manual processing were used and not all breaks and outliers MASH detected have been accepted. A monthly time step was chosen, while seasonal and annual time series were analyzed in parallel.
The series candidate for homogenization is selected among the available time series while the remaining series are considered as reference series. The role of each series changes at every step of the procedure.
Assuming that temperature follows the normal distribution, the additive model is applied
where X is the examined series, μ is the unknown climate change signal, E is the spatial expected value, IH is the inhomogeneity signal with T breakpoints and IH(T – 1) – IH(T) shifts and ϵ is the normal noise.
To filter out the unknown climate signal μ(t) and to increase the signal-to-noise ratio (power), several difference series are constructed from the candidate and weighted reference series.
for (j = 1, 2, …, N) where Zj(t) difference series, Zj(t) candidate series and and reference series constructed for the Xj(t) candidate series.
The optimal weighting is determined by minimizing the variance of the difference series to increase the efficiency of the statistical tests. This means that the power can be increased by decreasing the variance of the noise term. Provided that the candidate series is the only common series of all the difference series, breakpoints detected everywhere in the difference series can be attributed to the candidate series. The optimal weighting factors λji in vector form are
The optimal difference series are also obtained using the generalized-least-squares estimation for the climate signal μ(t).
The MASH method for multiple breakpoints detection is based on a hypothesis test for a given significance level, here equal to 5%. The difference series
where IHZ(t) inhomogeneity with K breakpoints and T1 < T2 < … < TK, are independent. The estimated breakpoints are
From the hypothesis, testing is defined:
H0: an estimated breakpoint is false breakpoint.
H1: an estimated breakpoint is real breakpoint.
There are two types of errors:
Type one: detection of a false inhomogeneity.
Type two: neglecting a real inhomogeneity.
Then the probability of these errors for the number of breakpoints is assessed. The inhomogeneity test, which can be characterized by the test statistic, is applied for Z(t) above all intervals, TSZ[k,l] ≥ 0, ∀ k, l: 1 ≤ k < l ≤ n.
For the given significance level the test statistic can be compared to the critical value α (by a Monte Carlo method) and in case of homogeneity it should be smaller.
The monthly series correction is based on confidence intervals. Once a first correction has been performed, if inhomogeneities are still detected the corrected series is corrected again. After that, missing data are completed using spatial interpolation techniques. The optimum interpolation parameters minimizing mean standard error are uniquely determined by the differences between the expected values and the covariances.
The MASH method also can use the metadata information automatically and metadata always have priority during the detection procedure. Moreover, the quality of the metadata can be verified by statistical tests. In this study we took benefit of the metadata existing in the archives of the HNMS.
3.2. Climatol method
The Climatol method developed at the Spanish State Meteorological Agency (AEMET) (Guijarro, 2011) is dedicated to the problem of homogenizing monthly climatological series. The methodology has been developed under the R programming language. The version Climatol 2.1 was used for this work.
As for the MASH methodology, Climatol was applied on a difference series between the tested station and a reference series constructed as an (optionally) weighted average of series from nearby stations. The selection of these stations is based not on the proximity criterion only, but also on a correlation criterion, because the anomalies of highly correlated time series are essentially synchronous. This implies, however, that climate varies smoothly throughout the studied region, because the presence of sharp geographical boundaries (e.g. high mountains) can lead to the use of nearby but badly correlated stations to compute the reference series.
In the Climatol package original data are normalized using proportions (ratios) or differences depending on the climatological variable. Proportions to normal climatological values are appropriate for zero-limited meteorological parameters with L-shape probability distributions (e.g. precipitation), while differences to normal are most suited to normally distributed variables (e.g. temperature). From the statistical point of view, this is equivalent to apply a type II linear regression model, instead of the commonly known type I. In ordinary linear regression (type I), the goal is to minimize the deviation between the observations to the regression line vertically. In that case, the underlying assumption is that the independent variable x is either controlled by the investigator or measured with negligible errors. But when adjusting regression lines to pairs of series of a climatological network, where the errors are a priori similar in all stations, the goal is to minimize the deviation perpendicularly from the data points to the fitted line. This case is the orthogonal regression (type II).
Once the original data are normalized, every term of each series is estimated as a weighted average of a prescribed number of the nearest available data. The weights to be applied to the reference data can be all the same (plain average) or be computed as an inverse function of the distance d between the observing sites. The function is formulated as 1/(1 + d2/h2) where h becomes the distance at which the weight is half that of a station placed in the same location of the data being estimated. If time series are not complete then means and standard deviations for the whole study period cannot be computed. In that case these parameters are computed from the available data only, the estimated series (after undoing normalization) are used to fill the missing data, and then the means and standard deviations are recomputed, the data are renormalized and new estimates of the series are obtained. This process is repeated until the maximum change in a mean is less than a chosen amount (0.005 units by default).
After having estimated all the data, for every original series a series of anomalies (differences between the normalized original and estimated data) is computed. For the detection of:
Outliers: The series of anomalies is standardized, and anomalies >5 (by default) standard deviations will result in the deletion of their corresponding original data.
Shifts in the mean: The standard normal homogeneity test (Alexandersson, 1986) is applied to the anomaly series in two stages:
On windows of 120 terms moved forward in steps of 60 terms (by default).
On the whole series.
In this study, the default threshold values were adjusted in each region empirically, because the default threshold values have been derived from synthetic white-noise series and have been set considerably higher than those obtained in Monte Carlo simulations. But in the real world, the anomaly series unavoidably show some degree of auto-correlation and local or general trends, depending on the type of climatic variable, its spatial variability, the density of the weather stations network, and the kind of data (annual, seasonal, monthly, daily etc.).
The maximum SNHT test values and their locations for every series are retained, and the series with the greatest value, if higher than the default threshold, is split at the point where this maximum has been computed. Values after this breakpoint are transferred to a new series (with the same coordinates) and deleted from the original series. Ideally, after the first split of a series, the whole process should be repeated, but this can lead to a very long process when dealing with a big number of stations with many inhomogeneities, and therefore a tolerance factor is provided to allow several splits at a time.
When all inhomogeneities detected over the prescribed threshold in the stepped SNHT test have been removed through the split process, the SNHT is applied again to the whole series, possibly generating more breaks in the series.
The last step of Climatol refers to missing data completion (including the data removed by the outliers and shifts detection stages). This applies to all the series, either original (not split series or first fragments of the split series) or derived (new series created by the split process).
In this study, 49 monthly series of mean temperatures over more than 30 years have been organized into seven regional groups and homogenized. All correlation coefficients in one group were higher than 0.9 for the daily data and not lower than 0.8 for monthly data from the first differences series even for distances between stations of the order of 450 km.
The analysis of data quality resulted in homogeneity problems and thus adjustments were necessary for most of the series. The results shown in Table 2, give the years of shift and the correction terms per season, by using MASH and Climatol. When a breakpoint is detected in year j (j = 1960, …, 2004) then the adjustment value is applied to the original series for the period [1960, j] or for the period [j – k, j], (k = 1, 2, 3,…, 45) where j – k corresponds to year of previous breakpoint. When an outlier is detected in year j (j = 1960, …, 2004) then the adjustment value is applied to the original series only for the year j. Bold marks represent outliers and underlines are verified from metadata.
Table 2. Shifts and adjustment factors (°C) per season. Bold marks represent outliers and underlines are verified from metadata
According to the results given in Table 2, the adjustments for MASH were in the range from −2.54 to 1.39 °C and for Climatol from −2.30 to 1.50 °C. The first year of correction is for the temperature series of Patras for the year 1966 using MASH and for the year 1961 using Climatol, while the last year of correction for both methods is 2002 for the time series of Tripoli. Eight stations for MASH and twelve stations for Climatol were corrected by more than 1.0 °C. Greatest correction factors for both methods are found in the summer and winter time series. The highest average negative adjustments of winter series are found in the time series of Kavala for both methods and the highest average positive adjustments of winter series are found in the time series of Kalamata for both methods. Also, the greatest average negative adjustments of summer series are found in the time series of Samos for both methods and the greatest average positive adjustments of summer series are found in the time series of Karpathos for both methods.
4.2. Analysis and comparative results
Only 8.2% of the stations passed both tests successfully and therefore these time series may be considered as homogenized, and 14.3% of the stations passed the Climatol homogeneity test without any split but the MASH test detected one or more breaks in at least 1 month. Another 10.2% of the stations passed only the MASH homogeneity test successfully but not the Climatol test and the remaining time series showed one or more breaks or outliers in both homogeneity methods. All the above percentages are summarized in Figure 4.
The distribution of the number of breaks (outliers included) per station is given in Figure 5 (all monthly breaks in a certain year are combined to one break per year). Most of the series according to both homogeneity methods had at least one break or outlier.
The distribution of the inhomogeneities (outliers not included) over the period 1960–2004 is shown in Figure 6 (all monthly breaks in a certain year are combined to one break per year). Many breaks are accumulated between 1980 and 1984. This may be explained by some relocation of weather stations from the city centre to nearby airports, (e.g. in Zakynthos and Kos), and also by changes in the observation rules (e.g. Pyrgos).
Comparing the results of MASH and Climatol, it was found that for the 32.7% of the stations (including stations with no detected inhomogeneities) the two algorithms agree to the years of break with only a few months difference, e.g. both methods detected a break in the time series of Chios between 1973 and 1974, MASH at the end of 1973 and Climatol at the beginning of 1974. For the 20.4% of the stations the two methods detected nearby years of break with the difference between the year of break detected MASH and that detected Climatol being less than 3 years, e.g. The time series of Naxos has a breakpoint at the end of 1969, according to MASH and in summer of 1972 according to Climatol. For the 14.3% of the stations at least one break was common, e.g. except the common break at 1986 that both methods detect for the time series of Kavala, MASH reports more breaks. Lastly, for the 6.1% of the stations the two methods detected completely different breakpoints with the difference between years of break being greater than 3 years, e.g. for the time series of Andravida Climatol splitted the series at 1975, while MASH at 1980. For the 26.5% of the stations only one method, either MASH or Climatol, detected breaks. Figure 7 visualizes the above percentages.
On average, only 15% of the breaks could be explained by the stations history because most of the stations have poor metadata; only major station relocations and some changes in the schedule of observations are reported.
At this point, it is worth to say that metadata are very important to assess the homogeneity of a series and check the detected breakpoints. Unfortunately, metadata are usually scarce and sometimes erroneous: both homogenization methods detect a breakpoint in the time series of Trikala somewhere between 1985 and 1986 but according to the metadata the station of Trikala has been relocated in 1973. Also, the station of Kavala except a documented relocation in June 1986, operated according to the metadata, with different observation practices up to 1997 (the exact dates are not available), while MASH detects abrupt changes between 1992–1995 and 1999–2002. This means either the metadata are erroneous or that the homogenization methods provide inaccurate results.
Apart from possible erroneous metadata information, metadata can serve to check if the location of a break is correct, or if it should be eventually shifted a couple of months backwards or forward, e.g. it was known that the station of Argostoli was relocated sometime but the date was unknown. MASH detects a break in 1984 but the confirmation from metadata is still pending. These cases demonstrate the necessity that the various meteorological services gather and digitize detailed metadata. On the other hand, homogenization helped incomplete or unknown metadata to be found. Metadata were not available for the time series of Rhodos and Helliniko, but the homogenization results helped to find that the station of Rhodos was relocated in July 1977 and the station of Helliniko in May 1968.
Besides unknown metadata, not all available metadata have been verified by the homogenization procedure, for example in the Nea Filadelfeia station different observation practices were applied between 1976 and 1979. MASH detects a break in summer 1976 and Climatol in May 1977 but neither detects a break in 1979. Probably both methods identified a small shift in 1979 and considered it as negligible.
As a conclusion, two different algorithms were used to homogenize mean monthly temperature series of 49 stations. The produced homogenized series present some differences. Both algorithms are based on relative homogenization methods, where a candidate series is compared to some estimation of the regional climate. MASH uses for comparison of multiple reference series that are not assumed to be homogeneous, whereas Climatol uses one composite reference series calculated from the data of 10 (default value) neighbouring stations and is assumed to be homogeneous. For the detection of breakpoints MASH uses the statistical criterion of maximum likelihood ratio (MLR) and hypothesis test, whereas Climatol uses the SNHT test in two stages: on stepped overlapping windows first (for multiple break detection) and a final application on the whole series (more powerful).
Both methods can effectively detect major change points in climatic series, e.g. those caused by station relocations, with MASH which takes advantage of metadata being more precise, whereas Climatol that does not use metadata may split the series a couple of months earlier or later, e.g. according to metadata Desfina operated with different observation practices up to12/1991 and Helliniko relocated at 5/1968. MASH correctly adjusts the series of Desfina until the end of 1991, whereas Climatol split the series until the spring of 1990. Similarly to the previous example MASH correctly adjusts the series of Helliniko until the spring of 1968, whereas Climatol until the winter of 1968.
A main difference between the two methods is in the way they analyze the monthly series. The 12 monthly series are analyzed independently in MASH, whereas in Climatol sequentially as one time series. Thus MASH detects more monthly breakpoints in total than Climatol which combines monthly results to one date per break, e.g. MASH detects in the time series of Tripoli five consecutive breaks in total (1998 in autumn and winter, 1999 in summer, 2000 in summer and autumn, 2001 in all seasons except in the summer, 2002 in all seasons), whereas Climatol results in two breaks (1999 and 2002).
Also, the primary operation of the two methods is different. Climatol is an automatic package, where all required parameters are specified when calling the homogenization function whereas MASH is interactive. Thus MASH detected few scattered anomalies among all months in the time series of Araxos, Corfu, Macedonia and Rethymno, but we did not correct them because we considered that these inhomogeneities may be affected result of anomalies from the nearest stations, since important anomalies in a neighbouring station may result to an apparent inhomogeneity in the tested station.
Another essential difference between the two methods is that Climatol is very tolerant to highly missing data (requires a minimum of 5 years of data), and therefore can use the information of the short-time series of a network as well, whereas MASH cannot be used if missing data is higher than 30%. This could explain in some extend the different homogenization results for stations such as that of Skopelos.
Concerning the differences in the homogenization corrections between the two methods, it is important to notice that MASH uses the smallest estimation from multiple comparisons and once a first correction has been performed, additional corrections are applied to the corrected time series until no further break is found, whereas according to Climatol significant breaks split the series, and all missing data are filled at the end of the process.
4.3. Impacts of homogenization on the temperature series
To assess the efficiency of the two homogeneity methods, standard deviations in the original and homogenized time series were calculated. Standard deviations of mean annual temperature series for the whole network and for the complete period 1960–2004, before and after homogenization, are given in Figure 8. Both homogenization methods lead to lower standard deviation values. Values of standard deviations in nonhomogenized series varied from 0.4 to 0.9 °C, whereas the range of standard deviations in homogenized series varied from 0.4 to 0.6 °C.
Trying to evaluate the impact of homogenization on annual time series some individual examples are given here. Figure 9 illustrates the annual temperature differences between the station of Tripoli and all reference series used for its homogenization, before and after homogenization. No metadata where available for Tripoli, but an important temperature decrease between 1998 and 1999 and an equally important rise in 2001 are obvious. Trying to understand better the impact of the homogenization on the final output, homogenized mean annual temperature series of Tripoli were transformed into standardized anomaly series using the period 1961–1990 as reference. The same calculations were carried out also for the original series. Results are aggregated in Figure 10. It is clear that Tripoli's annual temperature series has a better temporal behaviour and coherence after the homogenization.
The impact of homogenization on mean annual temperature series of five stations located in the eastern Aegean region is given in Figure 11, which displays the cumulative sum of anomalies before and after homogenization. The difference between each mean annual temperature record and the average value of the whole period 1960–2004 is calculated, and this is cumulatively summed up. Because the average is subtracted from each value, the cumulative sum always ends at zero. The range of cumulative sum of anomalies of nonhomogenized series is clearly wider compared with the homogenized ones. All stations represent a quite similar behaviour after homogenization, a decreasing trend from 1971 up to 1992, approximately, showing that mean annual temperatures were below its average value, a period from 1993 until 1997 approximately, where values are equally distributed around the average and an increasing trend after 1998, approximately, indicating that mean annual temperatures tend to be above average.
4.3.1. Trend analysis
To estimate the impact of homogenization in linear trends, seasonal and annual trends for each station have been computed for raw, MASH and Climatol data series. Additionally, seasonal and annual regional trends for the period 1960–2004 have been computed before and after homogenization. The trends of the mean air temperature before and after homogenization were evaluated using the two sided Kendall test (Kendall, 1976) and examined in terms of sign (positive or negative), magnitude and significance level (95%). Seasonal and annual mean surface temperature trends for each station of raw and homogenized series are shown in Figures 12 and 13.
Trends in winter mean air temperature during the period 1960–2004 present a negative slope in central and southern Greece and a positive slope in the northern part of the country. After homogenization trends for most of the stations were found not to be statistically significant and some stations present major changes compared to the raw time series. The greatest variation in trend size and significance before and after homogenization was observed for the station of Kavala located in northern Greece (climatic region A), where winter mean temperature trend changed from 0.67 to 0.13 °C/decade after MASH and 0.10 °C/decade after Climatol but not statistically significant after homogenization with both methods. Moreover, mean winter temperature trend for the station of Rhodos (climatic region F) changed both in sign, magnitude and significance, significant trend before changed from 0.34 to −0.01 °C/decade after MASH and −0.08 °C/decade after Climatol with no statistical significance with both homogenization methods. Also winter temperature trends for Zakynthos and Kalamata in western Greece (climatic region C) decreased by 0.32 °C/decade after homogenization and Limnos (climatic region E) and Siteia, Ierapetra (climatic region E) decreased around 0.20 °C/decade after Climatol and around 0.15 °C/decade after MASH. Comparing the winter mean air temperature trends after MASH and Climatol, the main difference focuses on three stations located in western Greece, climatic region C (Corfu, Aktio, Andravida) and two stations in northern Greece, climatic region A (Serres and Kozani) which show positive trend after MASH and negative trend after Climatol but these trends are not statistically significant and have weak increase or decrease.
On the contrary, summer temperature trends present a positive slope and are statistically significant for almost all the stations. The main difference between trends before and after homogenization is in magnitude for most of the stations and regarding their sign for the stations of Karpathos, Kalamata and Siteia. The highest increase of mean summer temperature was observed for the station of Samos (climatic region E), where the trend is statistically significant and changed from 1.04 °C/decade before homogenization to 0.36 °C/decade after MASH and 0.43 °C/decade after Climatol. Additionally, a great variation in trend is observed at the station of Karpathos (climatic region G), where the trend changed from −0.49 °C/decade before homogenization to 0.06 °C/decade after MASH and 0.08 °C/decade after Climatol. Also, Argostoli, Pyrgos, Serres and Piraeus present major change in trends with a decrease of around 0.5 °C/decade after homogenization for the first two stations and around 0.4 °C/decade for the last two stations.
Springtime mean temperature trends show some indication of spatially heterogeneity concerning the sign but show weak (not statistically significant) increasing or decreasing trends. Trends range before homogenization from approximately −0.3 °C/decade at Siteia, Kalamata and Kos to approximately 0.3 °C/decade at Serres, Piraeus and Skopelos. After homogenization the mean temperature trends in spring are reduced about 0.1–0.2 °C/decade for most of the stations.
Autumn mean temperature trends show heterogeneity concerning the sign and most of them are not statistically significant especially after homogenization. The only station that has a statistically significant trend with no changes in magnitude and sign before and after homogenization is Ioannina (in climatic region B) with slope −0.20 °C/decade. It is worth to remind that both homogenization methods found the Ioannina series as homogeneous and the only season that Ioannina series has a significant trend is in autumn. Before homogenization the highest positive and statistically significant autumn mean temperature trends were found in Kavala, Serres, Aktio and Rhodos ranging from 0.5 to 0.3 °C/decade, whereas the highest negative and statistically significant autumn mean temperature trends were found in Limnos and Tripoli with −0.35 °C/decade. After MASH, trends for these stations maintained their sign but with much smaller value with no statistical significance. Similarly to MASH, trends for the above stations maintained their sign after Climatol, with the exception of Serres, but with much lower magnitude and no statistical significance.
The visual inspection of annual series shows that after homogenization the regional coherence of the trends is higher compared with the raw series, both in size, sign and significance. No station after homogenization presents a statistically significant trend with the exception of the stations of Macedonia station in northern Greece and Patra in western Greece that show a statistically significant annual temperature trend after MASH.
Seasonal and annual regional mean temperature trends for the period 1960–2004, before and after homogenization, are summarized in Table 3. Bold marks denote the statistically significant trends (c.l. 95%). The analyses confirm a statistically significant warming trend only in summer.
Table 3. Seasonal and annual regional trends (°C/decade) before and after homogenization. Bold marks are statistically significant trends (c.l. 95%)
Summer time regional mean temperature trends of raw series are significant for all climatic regions with the exception of regions C (Western Greece) and G (Crete) and range from 0.07 to 0.35 °C/decade, while trends after MASH are statistically significant and range from 0.22 to 0.32 °C/decade and trends after Climatol are statistically significant with the exception of region C and range from 0.15 to 0.31 °C/decade. Finally, the annual trend for region A (Northern Greece) is statistically significant before homogenization but this is not confirmed after homogenization.
4.3.2. Climatological normals
After the homogenization of the 49 mean monthly temperature series, climatological normals of mean annual temperature series for the period 1961–1990 were computed. Climatological normals are very important for the assessment of climate change. They serve two principal purposes: as a reference against which observations at a particular time are compared and as a prediction (implicit or explicit) of the conditions most likely to be experienced at a given location (Trewin, 2007). World Meteorological Organization recommended a 30-year period to be used as a worldwide standard for the calculation of normals (which at that time meant 1921–1950).
The differences of climatological normals for the period 1961–1990 per station between raw (after missing values completion) and homogenized data are shown in Figure 14. The absolute values of differences ranged between 0.0 and 0.8 °C for MASH and between 0.0 and 1.0 °C for Climatol. These differences of normals strongly show that when nonhomogenized temperature data are used for the calculation of normals, then some or all of the data used for the calculations are not fully representative of the current observations at a given location. This reduces both the predictive ability and the appropriateness of the normals to be used as reference values against which current records can be compared.
4.3.3. Köppen climate classification
The classification of climate originally developed by the German scientist, Wladimir Köppen, is still in widespread use and various updates and new classifications have been reproduced (Kottek et al., 2006; Peel et al., 2007).
The climate classification of each station based on the work of Wladimir Köppen as presented by Kottek (Kottek et al., 2006) has been examined here before and after homogenization. For the needs of this work mean monthly temperature and precipitation time series for the period 1960–2004 have been used. Precipitation series were not checked for inhomogeneities.
According to raw data three stations Serres, Florina and Kozani in northern Greece belong to warm temperate climate, fully humid with hot summer (type Cfa), two stations Macedonia and Larisa located in northern and central Greece respectively, have arid, cold steppe climate (type BSk), Piraeus station located in Attica region has arid, hot steppe climate (type BSh) and the rest of the stations belong to warm temperate climate with dry and hot summer (type CSa). Using homogenized mean temperature series resulted either by MASH or Climatol only one station, that of Kavala located in northern Greece, changes climate type, from warm temperate climate with dry and hot summer (before homogenization) turned into arid, cold steppe climate (after homogenization). Besides Kavala, if the homogenized mean annual temperature for the whole period 1960–2004 would have been higher by 0.05 °C, two more stations, Eleusina and Helliniko located in the Attica region would have changed climate type, from warm temperate climate with dry and hot summer to arid, hot steppe climate.
This paper presents a study of homogenizing monthly temperature series of the Greek weather station network for the period 1960–2004, using two different methods, the MASH and Climatol. The two methods differ in the software type (MASH binary; Climatol sources in R), in the primary working mode (MASH used here is interactive; Climatol is automatic), in the type of reference series (MASH multiple references; Climatol composite reference), in the detection method (MASH uses maximum likelihood ratio and hypothesis test; Climatol SNHT test), in the time series analysis (MASH examines and reports the breaks in the monthly time series separately; Climatol synthesizes monthly breaks to one), in the metadata management (MASH uses metadata; Climatol does not), in the missing data tolerance (MASH around 30%; Climatol very high) and in the correction method (MASH multiple comparisons; Climatol missing data filling). Due to all of these differences the two methods produce different but homogenized series.
Temperature time series of almost 92% of the stations were found to be nonhomogeneous. One consequence of homogenization exercises was the retrieval of missing metadata. Both tests improved the homogeneity of the temperature data, reduced standard deviations and improved the temporal consistency of the time series.
The main homogenization result is that seasonal and annual mean temperature trends present significant changes before and after homogenization. The most noticeable changes are found in summer followed by winter where trends for many stations changed both in sign, magnitude and statistical significance. The northern regions show a small increase of mean temperature in winter while the other regions have decreasing trends. In summer, mean temperature trends after homogenization are increasing for all regions. The most strongly increasing summer temperature trend (0.43 °C/decade after Climatol and 0.34 °C/decade after MASH, corresponding to an increase of 1.29 and 1.02 °C respectively over 30 years) is observed for the station of Samos, whereas the smallest increasing summer temperature trend (0.07 °C/decade after Climatol and 0.04 °C/decade after MASH, corresponding to an increase of 0.21 °C and 0.12 °C respectively over 30 years) is found at the Aktio and the Kalamata stations, respectively. In general, trends after homogenization present better spatial coherence.
Another significant homogenization result is that the climatological normals for the period 1961–1990 were strongly affected in some cases by the inhomogeneities of the raw data series, with the difference between raw and homogenized normals reaching even 1.0 °C. This implies that normals calculated by nonhomogeneous data sets are not suitable as benchmarks, as they can lead to wrong estimations of climate conditions at a given location.
Lastly, all stations were classified before and after homogenization according to the Köppen climate classification, (Kottek et al., 2006). Only one station changed climate type after homogenization, that of Kavala from warm temperate climate with dry and hot summer (before homogenization) turned into arid, cold steppe climate (after homogenization). Nevertheless, further work is still required toward this direction because the precipitation data used were not homogenized. Our future plan is to homogenize the same data series using the new method and software from the research action cost home es0601.
As a general conclusion, the use of homogenization is not to improve absolute values but to produce series with temporal and spatial consistency. No homogenization method is perfect, but both methods improve the original series and allow better analysis of climatic variability (spatial and temporal) than with the original nonhomogeneous series.
The authors would like to thank the HNMS for providing the temperature and precipitation data as well as the J. Guijarro and T. Szentimrey for providing the homogenization algorithms.