Volume 24, Issue 2 p. 191-205
RESEARCH ARTICLE
Free Access

High resolution air temperature climatology for Greece for the period 1971–2000

A. Mamara

A. Mamara

Hellenic National Meteorological Service, Hellenikon, Greece

Laboratory of Atmospheric Physics, Department of Physics, University of Patras, Patras, Greece

Search for more papers by this author
M. Anadranistakis

M. Anadranistakis

Hellenic National Meteorological Service, Hellenikon, Greece

Search for more papers by this author
A. A. Argiriou

Corresponding Author

A. A. Argiriou

Laboratory of Atmospheric Physics, Department of Physics, University of Patras, Patras, Greece

Correspondence: A. A. Argiriou, Laboratory of Atmospheric Physics, Department of Physics, University of Patras Campus, GR-265 00 Patras, Greece. E-mail: [email protected]Search for more papers by this author
T. Szentimrey

T. Szentimrey

OMSZ – Hungarian Meteorological Service, Budapest, Hungary

Search for more papers by this author
T. Kovacs

T. Kovacs

OMSZ – Hungarian Meteorological Service, Budapest, Hungary

Search for more papers by this author
A. Bezes

A. Bezes

National Observatory of Athens − Institute for Environmental Research and Sustainable Development (IERSD), Athens, Greece

Search for more papers by this author
Z. Bihari

Z. Bihari

OMSZ – Hungarian Meteorological Service, Budapest, Hungary

Search for more papers by this author
First published: 11 April 2017
Citations: 8

ABSTRACT

Climate atlases provide an excellent overview of a region's climate in the form of maps and disclose information about climate change. They constitute a valuable tool for easy access and management of climate information for a wide range of users including scientists, policymakers, resource managers and urban planners.

The main aim of this work was the production of high resolution monthly mean air temperature climatology for Greece, determined from a high resolution homogeneous mean temperature dataset. Temperature data were obtained from 52 meteorological stations of the Hellenic National Meteorological Service. High resolution temperature maps were obtained by interpolating the homogenized data. MISH (Meteorological Interpolation based on Surface Homogenized Data), an interpolation method developed for meteorological purposes, was applied twice, once using as topographic model variables the elevation derived from a 90 m digital elevation model (DEM) and the first 15 AURELHY (Analyse Utilisant le Relief pour les Bésoins de l'Hydrométéorologie) principal components, and then using some additional model variables besides the elevation and the 15 AURELHY principal components: latitude, incoming solar irradiance, Euclidean distance from the coastline and land to sea percentage. The spatial interpolation of monthly mean temperature was performed with a 0.5′ resolution (∼730 m at 38 ° N). The results showed that modelling with only the AURELHY variables is not sufficient and the use of additional model variables is necessary.

1 Introduction

Nowadays, the demand for high resolution climatic atlases has increased since most environmental disciplines are related to climate. Climatic atlases contribute to the general global climate monitoring puzzle by integrating meteorological data into high resolution maps that clearly delineate the climatic conditions of an area. They also contribute to the economy and everyday life by providing important input to economic and political decisions related to renewable energy production, the building industry, transport logistics and management. Their preparation is complex because it involves many disciplines such as meteorology, climatology, statistical analysis and geographical information systems (GIS).

Many European meteorological services and institutes, following the recommendation of the World Meteorological Organization (WMO) for generating grids that show the spatial distribution of observations (WMO, 2014), have published national climatic atlases (e.g. Cegnar, 1995; Sachweh and Enders, 1996; Kirchhofer, 2000; Müller-Westermeier et al., 2003; Tolasz et al., 2007; Zaninović et al., 2008; AEMET and Instituto de Meteorologia in Portugal, 2011). Some countries produced digital climatic atlases (e.g. DWD, 1999; Meteo-France, 1999, 2004; Auer et al., 2001; LUBW, 2006); others published web-based digital climatic maps (e.g. the German Climate Atlas http://www.dwd.de/bvbw/appmanager/bvbw/dwdwwwDesktop?_nfpb=true&_pageLabel=dwdwww_result_page&gsbSearchDocId=1070304; the Digital Climatological Atlas of Spain http://opengis.uab.es/wms/iberia/mms/, the Finnish wind Atlas http://www.tuuliatlas.fi/creators/index.html). Some high resolution climatology studies for high elevation regions such as the greater Alpine region (Auer et al., 2008; Hiebl et al., 2009) and the Carpathian Mountains (Antolovic et al., 2013; Lakatos et al., 2013; Szalai et al., 2013) have been done. For Greece there are no climatic atlases except that of 1935 by Mariolopoulos and Livathinos (1935) and the project Geoclima (http://www.geoclima.eu) that attempted to map climate data.

The atlas presented here is based on a set of homogenized monthly temperature time series from 52 weather stations covering a 30 year (standard) period from 1971 to 2000. Several studies related to the development of spatial analysis algorithms, seeking the best approach to generate climate surfaces from point data, have been published. Various statistical methods exist for the spatial interpolation of climate data. These are classified in three main categories (Sluiet, 2009): deterministic, probabilistic and other. Deterministic methods, such as the nearest neighbourhood (Thiessen polygons and triangulation), inverse distance weighting, splines and linear regression, use the geometric characteristics of point observations and create a continuous surface. Probabilistic methods such as kriging (e.g. ordinary kriging, simple kriging, co-kriging, universal kriging, residual kriging) are based on probabilistic theory; the interpolated field is one of many realizations. The above interpolation methods are described by Tveito et al. (2005) and Dobesch et al. (2007). Other methods include those developed only for the interpolation of metrological data and use a combination of deterministic and probabilistic methods. These are AURELHY (Analyse Utilisant le Relief pour les Bésoins de l'Hydrométéorologie) (Bénichou and Le Breton, 1987), PRISM (Parameter-Elevation Relationships on Independent Slopes Model) (Daly et al., 1994) and MISH (Meteorological Interpolation based on Surface Homogenized Data) (Szentimrey and Bihari, 2007). The MISH method was used in this work. The temperature maps cover the period 1971–2000, deemed by the WMO to constitute a standard climatological period. In order to maximize accuracy during modelling, the period was extended to 45 years (1960–2004).

The objective of this work was to produce modern high resolution temperature climatology for Greece. This is a first step towards the preparation of a comprehensive climatic atlas that will include additional meteorological parameters in the future. The improvements of the proposed climatology over that provided by Geoclima are that the interpolated data were homogenized, improved geographical and topographical data were used, a different interpolation method suitable for meteorological parameters was applied and the data produced are at a higher spatial resolution.

2 Data and methodology

2.1 Examined area

Due to its unique location in the Mediterranean and its fairly diverse topography, Greece is characterized by several climatic zones. Greece (total area 131 957 km2, land borders 1180.71 km long (HSA, 2011)) belongs to the temperate continental climate zone of the northern hemisphere, at the southernmost end of the Balkan Peninsula, between latitudes 34–42 ° N and longitudes 19–30 ° E, with coastlines (coastline length of 15 021 km) on the Aegean Sea (east), the Ionian Sea (west) and the Libyan Sea (south). The main geographical areas are the mainland, the islands and the Aegean basin. The mainland covers about 80% of the total area; the remaining 20% is shared among about 6000 islands and islets. The landscape is mainly mountainous or hilly. The land is dry and rocky; only 20.45% of it is arable. The elevation ranges up to 2904 m above sea level. The Greek climate is typical Mediterranean: mild and rainy winters, relatively warm and dry summers and extended periods of sunshine throughout most of the year (http://www.hnms.gr/hnms/english/climatology/climatology_html).

2.2 Data and their homogeneity

The data come from 52 meteorological stations (Figure 1; details in Table 1) of the National Meteorological Service of Greece. The stations are part of the WMO Global Observing System. Weather observations are frequently exposed to artificial influences caused by station relocations, changes in the instrumentation introducing inhomogeneities, so the WMO recommends homogenization (WMO, 2014). The 52 mean temperature series used were homogenized (Mamara et al., 2014) using the HOMER software (Venema et al., 2012). Eighty-five per cent of the temperature time series suffered from inhomogeneities and were corrected accordingly. This is one of the basic differences between this work and the Geoclima project.

Details are in the caption following the image
Locations of the meteorological stations.
Table 1. List of stations, station IDs, geographical co-ordinates (longitude, latitude) and landscape factors (elevation and distance from coast).
Station ID Longitude (°) Latitude (°) Elevation (m) Distance from coast (km)
16606 23.529 41.077 32 39.47
16607 24.150 41.150 104 31.59
16608 24.408 40.938 7 0.64
16609 24.887 41.137 84 18.62
16611 26.300 41.200 15 46.87
16613 21.400 40.780 695 106.54
16622 22.972 40.527 2 0.86
16627 25.947 40.857 4 1.69
16632 21.842 40.290 621 62.03
16641 19.914 39.608 1 0.42
16642 20.810 39.700 484 51.64
16643 20.769 38.922 2 0.39
16645 21.760 39.510 110 75.81
16648 22.460 39.646 73 34.00
16650 25.233 39.923 2 1.55
16662 23.710 39.110 11 2.27
16665 22.793 39.224 13 2.27
16667 26.604 39.054 4 0.52
16672 21.350 38.600 25 11.47
16674 23.100 38.380 110 19.56
16675 22.430 38.860 14 8.24
16682 21.287 37.923 10 5.78
16684 24.491 38.963 22 0.56
16685 20.505 38.120 21 0.65
16687 21.422 38.149 11 1.42
16689 21.730 38.080 1 10.82
16693 22.530 38.410 590 3.91
16699 23.563 38.335 138 7.37
16701 23.730 38.050 138 12.15
16706 26.142 38.345 5 0.10
16707 21.440 37.670 36 3.69
16710 22.397 37.525 651 28.28
16715 23.780 38.107 237 17.31
16716 23.742 37.890 43 1.63
16717 23.630 37.950 5 0.36
16718 23.522 38.068 27 2.46
16719 20.888 37.751 3 1.71
16723 26.916 37.691 6 0.67
16726 22.023 37.069 6 4.77
16732 25.373 37.101 9 0.06
16734 21.700 36.810 52 0.52
16738 24.430 36.730 165 0.87
16742 27.088 36.802 129 4.27
16743 22.980 36.150 167 0.37
16746 24.145 35.529 148 2.89
16749 28.088 36.402 7 1.00
16754 25.182 35.335 39 0.74
16756 25.730 35.010 10 0.10
16757 26.103 35.216 114 0.65
16758 24.500 35.360 5 1.08
16759 24.762 35.067 6 0.88
16765 27.147 35.428 11 0.12

2.3 Methodology

The goal of interpolation techniques is to provide country-wide information given a selection of points of the studied area. The relationships used correlate the meteorological or climatological variables to geographical or topographical parameters. These relationships reflect the impact of the topography on the climate of the area. One of the most important tasks in this procedure is to identify the geophysical and topographical parameters affecting climate. The geophysical characteristics of a region do not affect all climatological parameters and, of course, they do not affect every point on the planet in the same manner.

In this study, elevation data were used as a temperature predictor, thus accounting for the atmospheric lapse rate. Elevation data come from the digital elevation model (DEM) (http://srtm.csi.cgiar.org) originating from the NASA Shuttle Radar Topography Mission (SRTM). The SRTM provided DEM data for all land between 56 ° S and 60 ° N covering almost 80% of the Earth's total land mass. The horizontal resolution is 3 arcsec, about 90 m along the Equator (1 arcsec approximately 30 m) data are available for use only in the USA. The vertical error of the DEM data is >16 m. Elevations (in metres) are referenced to the WGS84/EGM96 geoid and horizontal data are geo-referenced to the WGS84 ellipsoid using a geographical projection. The SRTM 90 m DEM data are provided in a mosaic of 5° × 5° tiles. The following tiles were used: srtm_40_04, srtm_40_05, srtm_41_04, srtm_41_05, srtm_41_06, srtm_42_04, srtm_42_05, srtm_42_06.

Besides the elevation, there is no universal rule to dictate the choice of other geographical and topographical variables; variables commonly used in such interpolations are slope angle, aspect, distance to the coastline, longitude and latitude (Agnew and Palutikof, 2000; Ninyerola et al., 2000; Vicente-Serrano et al., 2003; Claps et al., 2008; Feidas et al., 2013). Since many meteorological parameters are strongly influenced by topography, the AURELHY method which models the spatial interpolation of climatic variables using topography was developed (Bénichou and Le Breton, 1987). AURELHY proposes an objective method based on principal component analysis (PCA) for the determination of topographic variables. The first step in AURELHY is terrain analysis; each grid point is described by its altitude and by the elevation differences between the central point and the neighbouring grid points in a squared matrix (e.g. 11 × 11 DEM grid points). Consequently each grid point is assigned to 121 data values. PCA is then applied to a large number of elevation vectors representing small areas (e.g. 11 × 11 DEM grid points) in order to condense the large amount of total information. The principal component (PC) loading of the first few PCs provides a good description of the most important patterns of orographic variability on a scale similar to the size of 11 × 11 (Gyalistras, 2003). Thus, the grid points are represented by their elevation and the values of the first few PCs.

In this work, the so-called AURELHY PCs were used as variables to model monthly mean temperature. However, instead of the typical AURELHY version, a modified version was used; each grid point is characterized by its altitude and elevation differences between the central point and the 1550 neighbouring grid points, i.e. 2 × 16 neighbours in the latitudinal direction and 2 × 23 neighbours in the longitudinal direction (1550 = (2 × 16 + 1) × (2 × 23 + 1) − 1). This modification is necessary in order to obtain approximately equal distances in both latitudinal and longitudinal directions in the mid-latitudes. This large amount of information was condensed via PCA. The PCA showed that the grid points can be represented by their elevation; the values of the first 15 PCs explain more than 90% of the total variance due to orography. The first five PCs can easily be interpreted geometrically as shown in Figure 2. PC-1indicates peaks (positive black coloured values) and valleys (negative white coloured values), PC-2 east−west slopes, PC-3 north−south slopes, PC-4 north−south saddles and PC-5 northeast−southwest saddles. Additional PCs account for subtle structures.

Details are in the caption following the image
The first five principal components (PCs): (a) PC-1 indicates peaks and valleys, (b) PC-2 east–west slopes, (c) PC-3 north–south slopes, (d) PC-4 north–south saddles and (e) PC-5 northeast–southwest saddles.

Besides the first 15 PCs some additional variables were used as temperature predictors, according to the study by Feidas et al. (2013); this study demonstrated that the distance to the coastline, land to sea percentage and latitude are among the factors affecting temperature. The sea has a great influence over weather and climate in maritime regions; it takes much longer to heat up but retains that heat far better than the land. Consequently inland areas are often much warmer during the day than coastal areas. This situation is reversed at night and continental regions are subjected to large temperature fluctuations. The influence of the sea diminishes when moving further inland. Here the Euclidean distance to the nearest sea coast (in kilometres) was used as a predictor in order to investigate the temperature dependence on the distance to the coastline.

The large number of islands in Greece increases the climate variability. The maritime influence on climate is very strong. In order to capture maritime influences, a field of land to sea percentage of area coverage within a 10 km radius was calculated, with the land taking the value 1 (100%) and the sea the value 0 (0%). Also, the latitude ϕ (decimal degrees), which is related to the general climate and temperature patterns, was used to describe the spatial distribution of mean temperature.

Solar irradiance was also used as a temperature predictor since it is the primary driver of atmospheric processes. Greece enjoys about 250 days of sunshine annually. Variation in elevation, orientation (slope and aspect) and shadows cast by topographic features affect solar irradiance impinging at different locations and consequently affect temperature. Here the monthly solar irradiance on the Earth's surface was calculated via libRadtran, a detailed radiative model for solar and terrestrial radiation in the atmosphere (Mayer and Kylling, 2005). The linear regression between the mean temperature in September of the 52 stations and the corresponding solar irradiance illustrates this dependence (Figure 3); an air temperature increase of about 0.78 °C per 10 solar irradiance units (W m−2) and the R statistic indicate that the model as fitted explains 41.7% of the temperature variance. The correlation co-efficient equals 0.65 (p value < 0.05) indicating a statistically significant (95.9% confidence level) moderate relationship.

Details are in the caption following the image
Linear regression model between mean temperature in September for the period 1971–2001 (all stations included) and solar irradiance. The external boundaries show the 95.0% prediction intervals for new observations and the inner boundaries the 95.0% confidence intervals for the mean of the 52 observations.

Twenty geographical and topographical variables were used in total, all of them (except the land to sea percentage coverage) calculated within an 800 m radius at a spatial resolution of 0.5 min (0.0083333333°), this resolution corresponding to a range from 689 m (at 42 ° N) to 769 m (at 4 ° N). Apart from the homogenized temperature series used in this work, another difference with Geoclima (Feidas et al., 2013) is the use of an objective method for parameterization of the topography (AURELHY PCs) and the use of solar irradiance. The higher resolution of the data is an additional improvement. In Geoclima a 5000 m radius was selected for calculation of the mean elevation and an even greater (10 and 20 km) radius for other topographical variables, while the spatial resolution of the Geoclima maps was lower than the maps produced here.

2.4 The interpolation method MISH

As already discussed, there are many methods for spreading discrete measurements over a continuous surface. In meteorology statistical and mainly geostatistical methods, built in GIS software, are applied in order to predict the spatial distribution of climate variables. The main disadvantage of geostatistical methods is the use of a single realization in time for modelling and that they do not take long term time series of climate variables into account (Szentimrey and Bihari, 2014). In order to overcome this disadvantage, the MISH method (Meteorological Interpolation based on Surface Homogenized Data) for producing mean temperature gridded data on a monthly basis was selected. In addition, spatial interpolation with MISH requires homogenized data series; in the work of Feidas et al. (2013) multiple regression was applied in order to estimate climate parameters by using only the topographical and geographical variables as predictors.

MISH developed at the Hungarian Meteorological Service by Szentimrey and Bihari (2007, 2014):
  1. gets information from long term data series by modelling the climate statistical parameters;
  2. calculates the modelled optimum interpolation parameters which are certain known functions of the modelled climate statistical parameters, and
  3. substitutes the modelled optimum interpolation parameters and the predictor values into the interpolation formula.

MISH consists of the modelling and the interpolation part. In the modelling part the meteorological parameter (e.g. temperature in the case of the additive model or precipitation in the case of the multiplicative model) is described as a function of the topographic variables via multiple linear regression equations. The dependent variable is the spatial trend (expected values in space); the long term data series (station mean values) constitute a sample in time and space. The independent variables are the elevation, the first 15 AURELHY PCs, solar irradiance, distance to the coastline, land to sea percentage and latitude.

The main steps of the modelling part are:
  1. modelling of the climate statistical parameters for the 0.5′ grid based on long homogenized monthly series and on deterministic model variables (the deterministic model variables are the topographical and geographical variables; in this work, the data series required were homogenized using HOMER (Mamara et al., 2014)), and
  2. performance statistics of the proposed modelling.
The interpolation part applies on the results of the modelling part; its main features are:
  1. the additive (e.g. temperature) or multiplicative (e.g. precipitation) model and interpolation formula can be used depending on the climate parameters;
  2. daily or monthly values and many years' means can be interpolated;
  3. few predictors are also sufficient for the interpolation;
  4. the expected interpolation error is modelled, and
  5. additional background information such as satellite, radar or forecast data can be used. The capability for application of the background information is very useful for precipitation, for example, because for gridding precipitation a dense station network is necessary; however, no background information was used in this work.

In contrast to the geostatistical methods, the values of variograms must be modelled for each interpolating process. One of the most important advantages of MISH is that the modelling part can be executed only once before the gridding of data on different time scales such as daily, monthly or seasonal (Lakatos et al., 2011). Also, different periods can be used in the modelling and in the gridding part. The modelling part of MISH was performed using the homogenized time series for the period 1960–2004. The implementation of the interpolation part was performed on the homogenized values of mean temperature in the period 1971–2000.

2.4.1 Mathematical background of MISH

Let Z(s1, t). Z(sn, t) be the known predictors (topographical and geographical parameters); the location vectors sare the elements of the given space domain D, t is the time and Z(s0, t) is the unknown predictand (in this case the mean temperature). The unknown predictand value can be estimated from the known predictor values using the appropriate interpolation formula, the type of which depends on the probability distribution of the meteorological parameter. A normal distribution has been assumed for temperature, for which the additive interpolation formula is appropriate (in the case of precipitation a multiplicative model can be applied for a quasi log normal distribution):
urn:x-wiley:13504827:media:met1617:met1617-math-0001(1)
with λ0, λi the interpolation weights. The root mean square interpolation error is given by:
urn:x-wiley:13504827:media:met1617:met1617-math-0002
and the representativity examination of the arbitrary station network is given by
urn:x-wiley:13504827:media:met1617:met1617-math-0003
where E is the expectation and D(s0) the standard deviation of the predictand.

Generally the optimum interpolation parameters minimize the expected interpolation error and can be written as functions of basic statistical parameters. The basic statistical parameters used in the interpolation formulae are deterministic or local parameters (expected values, standard deviations) and stochastic (correlations and covariances or variograms; variograms are preferred in geostatistics and covariances in meteorology). The optimum λ0 depends on E(s0) − E(si) for i = 1, …, n, and the optimumλi and REP(s0) depend on the ratios D(s0)/D(si) and the correlations r(si, sj) for i, j = 1, …, n.

In essence, these statistical parameters are climate parameters meaning that optimal interpolation could be achieved with an a priori knowledge of the climate. The homogeneous long term time series provide this knowledge and consequently a lot of information for modelling. Climate parameters are essentially known by long data series and MISH uses this information. The final interpolation formula is:
urn:x-wiley:13504827:media:met1617:met1617-math-0004(2)
with E(si), i = 0, …, n, the spatial trend values in the linear meteorological model of expected values, E{Z(si, t)} = μ(t) + E(si) for i = 0, …, n and μ(t) the common temporal trend. The form of the weighting factors is:
urn:x-wiley:13504827:media:met1617:met1617-math-0005(3)
with C and c the predictor–predictor covariance matrix and predictand–predictor covariance vector respectively; vector 1 is identically one.

A detailed description of MISH can be found in Szentimrey et al. (2011) and also in the MISH manual (Szentimrey and Bihari, 2014).

3 Results and discussion

3.1 Modelling spatial trends

Initially the modelling part was tested using only the elevation and the 15 AURELHY PCs as predictors of mean temperature. According to the results obtained the correlation co-efficient (r) ranged from 0.52 to 0.75 indicating a moderate correlation between mean air temperature and the variables of the model. Therefore, the AURELHY variables alone are not sufficient for modelling the spatial trend; some additional geographical variables are needed. It is worth noticing that in the project CARPATCLIM (Climate of the Carpathian Region) (Bihari et al., 2014) the temperature series from stations in the Carpathians were modelled with the same 15 AURELHY variables; the correlation was better and provided very good modelling results. It is therefore concluded that in the Mediterranean area, where the maritime influence and solar irradiance are important, it is not possible to model the spatial trends using the AURELHY variables only. This finding is in line with Thomas and Herzfeld (2004) who stated that when AURELHY is applied to a large region, e.g. East Asia, continent-scale effects such as the latitudinal decrease of temperature or the influence of different monsoon air masses (Domros and Peng, 1986) have to be taken into account.

The modelling part was tested using all 20 topographical and geographical variables described in Section 4, by applying the cross-validation test for the interpolation errors. This is a validation technique for assessing the performance of a predictive model, i.e. how well the model predicts the values at unknown locations. In this work all time series were interpolated between each other. Two procedures were implemented: interpolation with optimum parameters and interpolation with modelled parameters. The root mean square interpolation errors RMSE(si) or the representativity REP(si) values obtained from the interpolation with optimum parameters were then compared to the interpolation results obtained with modelled parameters in order to control the modelling. The mean representativity values (based on 52 stations) calculated for the monthly mean temperatures are shown in Figure 4. The representativity values with optimum parameters vary between 0.7 and 0.8 while the corresponding values with modelled parameters are much lower, ranging between 0.5 and 0.6 during winter and autumn and below 0.3 during summer. Much better representativity values were obtained with the optimum parameters meaning that on the basis of the station series modelling with the optimum interpolation parameters provides good estimations.

Details are in the caption following the image
Mean monthly representativity values (representativity = 1 − (root mean squared error)/(standard deviation)) for monthly mean temperature (for 52 stations). The black line denotes the interpolation with optimum parameters and the grey line the interpolation with modelled parameters.

Figure 5 shows the number of topographical and geographical variables per month used as air temperature predictors; Table 2 shows the variables used in the linear regression model per month. PC-1 to PC-15 correspond to the first 15 AURELHY PCs. At first glance, the linear regression models for all summer and winter months use five or six geographical variables while for April, May and September they use only three. Apart from the elevation used as an independent variable in all 12 linear regressions, solar irradiance and land to sea percentage have been taken into account for almost all months. Also latitude ϕ is used from May to August, November and December. The second principal component (PC-2) corresponding to east−west slopes seems to affect air temperature as well since it is used from January to March and from August to November, while the third PC (PC-3) indicating north−south slopes is used only in July, August and December. PC-4, corresponding to the north−south saddles, is used only during the winter months while PC-1, corresponding to peaks and valleys, is used only during the autumn months.

Details are in the caption following the image
Number of topographical and geographical variables per month used as air temperature predictors.
Table 2. Model variables per month applied in the linear regression model.
Month Elevation (m) ϕ (°) Land to sea percentage (%) Solar irradiance (W m−2) Distance from coast (km) PC-1 PC-2 PC-3 PC-4 PC-5 PC-6 PC-7 PC-8 PC-9 PC-10 PC-11 PC-12 PC-13 PC-14 PC-15
January × × × × × ×
February × × × × × ×
March × × × × × ×
April × × ×
May × × ×
June × × × × ×
July × × × × ×
August × × × × × ×
September × × ×
October × × × × ×
November × × × × × ×
December × × × × × ×

The correlation co-efficients (r) and the co-efficients of determination (r2) are shown in Figure 6; correlation is decreasing in late spring and summer while it increases again in autumn and winter. In fact the correlation co-efficients are very high (equal to or greater than 0.9) from October to March, are high (about 0.8) in April and September and good (about 0.7) from May to August. In accordance with the correlation co-efficients, r2 is higher in winter (∼0.9) and lower in summer months (∼0.5). The lower r values in summer should be attributed to the lower spatial variability of the air temperature, while during winter there is a distinct mean temperature difference between stations located in northern Greece and stations located in southern areas, with a temperature increase on average when moving from north to south. This difference is not observed in summer since high temperatures prevail in almost all regions and the mean temperature range is small. This is depicted in Figure 7 showing the frequency histogram of all data from the 52 stations for January (a) and August (b). It is clear that while the mean temperature in January ranges from about −4 to 16 °C, in August it varies only between 19 and 30 °C.

Details are in the caption following the image
Monthly correlation co-efficients (r) and the co-efficients of determination (r2).
Details are in the caption following the image
Frequency histogram of all data values from the 52 stations for (a) January and (b) August.

In order to examine the summer modelling results in more detail, the modelling procedure was repeated for August with the same topographical and geographical variables (elevations, 15 AURELHY PCs and the four additional variables) but for a different coverage: only 29 stations located mostly in the mainland (25 stations located in the mainland, three in the Ionian Sea and one in the Aegean Sea) were used. The area covered (Figure 11) extends between approximately 37.00 ° N–41.48 ° N and 19.90 ° E–23.90 ° E. The crosses in Figure 8 show the locations of the 29 stations. The modelling revealed some interesting results. The variables used in the linear regression were the same as before (Table 2) except for the land to sea percentage, but the performance was better; r and r2 were about 0.8 and the 0.7 respectively. This means that the spatial trend modelling in August is quite good for the major part of Greece including the main continental part and a smaller area of sea and the islands. The lower correlation values during summer may be due to the large part of the sea area and the numerous small islands of the complete coverage.

Details are in the caption following the image
Main continental area of Greece plus a limited sea area.

3.2 Assessing the monthly mean temperature climatology

The following analysis aims to assess whether the long term monthly mean temperatures have been modelled and predicted sufficiently well. The assessment is based on monthly scatter plots (Figure 9) comparing the observed versus predicted temperatures at the station locations. The analysis was carried out on the monthly dataset of the 52 stations. The visual inspection of the graphs reveals a strong relationship between observed and predicted values. The slope for all months indicates that there is neither substantial underestimation nor overestimation of the mean temperatures; the correlation co-efficient (r) is very high, exceeding ∼0.85. The co-efficient of determination (r2) is also high, ranging from 0.76 to 0.97, indicating that the regression line fits the data almost perfectly. It is worth noting that Feidas et al. (2013) also compared the predicted values against the observed and obtained high r values (∼0.75–0.88) for the annual, winter, spring and autumn temperatures but a low one (∼0.5) for summer and stated that all tested interpolation methods showed a consistent overestimation of temperatures for all seasonal and annual means.

Details are in the caption following the image
Observed temperatures versus predicted temperatures at the station locations from (a)–(l) January–December.

The residual values were also computed for all stations; the mean error (bias), the mean absolute error, the mean squared error and the RMSE were calculated (Table 3).The very small bias, ranging from −0.02 to 0.03 °C, indicates that the interpolation method is not seriously affected by systematic errors. Also, the low mean absolute error and RMSE values indicate a very good fit between model and observations.

Table 3. Statistics of the interpolation model.
Month Bias (°C) MAE (°C) MSE (°C) RMSE (°C)
January 0.03 0.39 0.25 0.50
February 0.02 0.38 0.27 0.52
March 0.00 0.35 0.23 0.48
April −0.03 0.35 0.25 0.50
May −0.02 0.41 0.28 0.53
June 0.00 0.41 0.25 0.50
July 0.01 0.45 0.32 0.57
August 0.00 0.42 0.27 0.52
September 0.00 0.35 0.19 0.43
October 0.02 0.36 0.23 0.48
November 0.03 0.36 0.23 0.48
December 0.02 0.44 0.30 0.54
  • MAE, mean absolute error; MSE, mean squared error; RMSE, root mean square error.

High monthly temperature climatology for the 30 year normal period 1971–2000 is illustrated in Figure 10 (December, January, February), Figure 11 (March, April, May), Figure 12 (June, July, August) and Figure 13 (September, October, November). The spatial distribution of temperature reveals various climatological features: (1) all through the year there is a strong dependence of the mean temperature on elevation; temperature decreases, more or less, with altitude and minimum values always occur at the highest mountains; (2) the temperature dependence on latitude becomes noticeable since temperature increases from north to south; (3) during summer the highest temperatures occur in the mainland. In the study of Feidas et al. (2013) a distinct dependence of air temperature on elevation, especially during summer, as well as a dependence of winter and annual temperatures on latitude was also detected. In July, the hottest month in Greece, the highest temperatures are observed in the plains of Thessaloniki, Thessaly, Kopais, Argolis and Aitoliki. The urban heat island in the Attica region is quite clear. The different temperature conditions prevailing in Athens and in the surrounding mountains (Hymettus (∼1026 m) located east of Athens, Pentelicus (∼1109 m) located northeast of Athens and Parnitha (∼1413 m) located northwest of Athens) are illustrated in detail.

Details are in the caption following the image
Mean temperature maps for the period 1971–2000 for the winter months: (a) December; (b) January; (c) February.
Details are in the caption following the image
Mean temperature maps for the period 1971–2000 for the spring months: (a) March; (b) April; (c) May.
Details are in the caption following the image
Mean temperature maps for the period 1971–2000 for the summer months: (a) June; (b) July; (c) August.
Details are in the caption following the image
Mean temperature maps for the period 1971–2000 for the autumn months: (a) September; (b) October; (c) November.

4 Summary and conclusions

In this study the mean temperature surfaces for the Greek area were interpolated at a spatial resolution of 0.5′ (730 m approximately at 38 ° N). Data from 1960 to 2004 come from 52 meteorological stations of the National Meteorological Service of Greece. Temperature data series were homogenized before interpolation.

Mean temperatures were interpolated using the MISH (Meteorological Interpolation based on Surface Homogenized Data) method. MISH was selected because it is based on a purely meteorological procedure and requires all meteorological and climatological information to be combined with model information. The advantage of MISH against other geostatistical methods (e.g. kriging) is the amount of information required for modelling the statistical parameters. MISH uses additional information from long term data series while the other methods use only one realization in time for modelling the covariance (or the variogram) structure.

The potential to use the elevation and the first 15 AURELHY (Analyse Utilisant le Relief pour les Bésoins de l'Hydrométéorologie) principal components as temperature predictors was investigated. It was found that, in the Mediterranean with an important coastline, the use of only elevation and the AURELHY variables cannot describe temperature. Additional topographical and geographical variables, namely the land to sea percentage and the expected solar irradiance, are required. The results revealed that elevation, land to sea percentage and solar irradiance should be used as independent model variables.

Apart from these variables, the east−west slopes seem to be associated with mean temperature during all months except between April and July. Also, latitude seems to affect mean temperature as well. Finally, north−south saddles are related to mean temperature only during the winter months.

The model performs very well from October to March, quite well for April and September and relatively well from May to August. The lower correlation results mainly in summer should be attributed to the lower spatial variability of the mean temperature. However, the modelling results of the spatial trend in August in the main continental part plus a limited part of the coastal area were much better, showing that the dependence of mean temperature on the selected topographical and geographical variables is sufficient. The statistical indicators of the predicted monthly temperatures for the 52 stations against the observed values were very good, providing high correlations and small errors.

Finally, the mapping of the monthly mean temperature normals (1971–2000) is very close to reality. Lower temperatures, mainly in winter, at Mounts Olympus and Voras (north-central Greece), on the Pindos mountain chain (western Greece) and on the Rhodopes mountain range (northeastern Greece) can be clearly distinguished; also, the low temperatures on other mountains such as the White Mountains and Mount Psiloritis in Crete, Mount Parnassus in central Greece and Mount Taygetos in the Peloponnese in southern Greece are clearly shown. The highest temperatures in summer can be observed on the plains of Thessaloniki, of Thessaly, of Kopais, of Argolis and of Aitoliki, as actually occurs.

Compared to a previous climatology, the proposed database has the following improvements: data are provided at a higher spatial resolution; temperature data were homogenized; improved geographical and topographical data were used; an interpolation method appropriate for meteorological parameters was applied; and the statistical results of the observed versus predicted values were better. In future work, additional climatic data will be homogenized and interpolated.