Dynamics of precipitation regions of Turkey: A clustering approach by K-means methodology in respect of climate variability

ACRONYM

DEFINITION

AO

Arctic Oscillation

AWOS

Automated weather observing stations

HBSC

Humid Black Sea Coastal

CA

Cluster analysis

COV

Coefficient of variation

EA

Eastern Anatolian

ENSO

El Niño–Southern Oscillation

MK

Mann–Kendall

MO

Mediterranean Oscillation

NAO

North Atlantic Oscillation

NCP

North Sea Caspian Sea Pattern

NWA

Northwestern Anatolia

PCA

Principal component analysis

SC

Spectral clustering

SCCA

Semi-Humid Continental Central Anatolian

SCM

Semi-arid Continental Mediterranean Zone

SNHT

Standard Normal Homogeneity Test

SOI

Southern Oscillation Index

SQMK

Sequential Mann–Kendall

SS/HCM

Subtropical Semi-Humid/Humid Coastal Mediterranean

TSMS

Turkish State Meteorological Service

Understanding the variable character of climate today is essential and imperative for many economic activities and sectors, especially energy, water resources, and agriculture. Ascertaining how precipitation patterns change spatially and temporally can provide crucial information for planning and management activities in various water-dependent sectors. Moreover, an insight into the statistical characteristics of precipitation variability is crucial not only for water-dependent sectors but also for mitigating risks of hydrological extremes such as floods and droughts. Clustering based on the properties of climatological data facilitates in determining the regions reflecting similar climate characteristics and provides a scientific basis for the planning and management of climate data-reliant sectors. Furthermore, clustering large spatial domains into regions of comparable climatic characteristics can provide vital information for climate change and climate variability studies (Machiwal et al. 2019). Nevertheless, delineating homogeneous precipitation regions is a challenging and complicated task owing to the inherent temporal and spatial variability of precipitation. The identification of homogeneous rainfall regions can become even more troublesome with sparse data and data with reduced temporal resolution (Sankranti et al. 2013; Irwin et al. 2017). In a country like Turkey, where physiographic features, such as elevation and orography, affect climate significantly and different weather systems prevail throughout the year, identifying distinct climate regions becomes even more challenging and complex. Many techniques were developed and implemented over the years with the availability of more reliable historical data and improved computational capabilities, enabling the use of different approaches for clustering climate parameters (Ünal et al. 2003; Netzel & Stepinski 2016). Tuel & Martius (2021) argued that using re-analysis dataset is becoming common for cluster analysis (CA) due to shortcomings of observational data. They further claimed that the re-analysis data also allow for a better assessment of climate change in clustering of the climate parameters. Several statistical methods are available for spatio-temporal characterization of precipitation patterns and identifying their groupings, including L-moments, harmonic analysis, multivariate regression, spatial correlation functions, spatial interpolation, empirical orthogonal functions, and regional frequency analysis (Guttman 1993; Ünal et al. 2012; Sankranti et al. 2013; Wazneh et al. 2015; Zaifoglu et al. 2018). Various spatial pattern recognition techniques such as CA and principal component analysis (PCA) are also other favoured approaches applied by researchers to determine homogenous rainfall regions (Darand et al. 2014; Marston & Ellis 2021). Among those, CA is one of the most preferred data analysis techniques to investigate similarities and dissimilarities in data and categorize the data as groups or clusters (Gore 2000; Jajuga et al. 2002). A cluster is a group of data points aggregated together and is established based on the similarities of the data points using statistical criteria, such as average difference or correlation. The purpose of the clustering procedure is to classify or group a specific set of objects into separate individual groups. Similar objects are targeted to be placed in the same cluster while different objects are positioned in the other clusters based on the degree of similarity (or dissimilarity) between the individual objects or patterns (Rokach & Maimon 2005; Berkhin 2006). Hierarchical and non-hierarchical clustering methods are common approaches employed in clustering studies. K-means is one of the fundamental non-hierarchical clustering techniques which acquired widespread use over the years for the classification of climate data as the method is easy to implement, computationally faster than hierarchical clustering, and suitable for large datasets (Berkhin 2002; Wilks 2011). It became one of the popular unsupervised machine learning algorithms favoured by researchers in recent years with the availability of various computing tools (Rodriguez et al. 2019). Further details of the K-means method are discussed in Section 2.2 of this study.

The teleconnection patterns also affect precipitation variability over the Mediterranean region, including Turkey. It is shown that the North Atlantic Oscillation (NAO) and Arctic Oscillation (AO) influence the spatial and temporal variability of precipitation over Turkey, causing wetter and drier periods inter-annually (Türkeş & Erlat 2003, 2006; Sezen & Partal 2019). The relationship between the teleconnection patterns and Turkish precipitation is discussed more in detail in Section 5 of this study. Several studies exist in the literature discussing the clustering of Turkish climate parameters. Six precipitation regime clusters were identified by Sarış et al. (2010) by employing the hierarchical, agglomerative CA using Ward's method based on monthly precipitation totals from 107 stations. In another study by Sönmez & Komuscu (2011), six main precipitation clusters were determined as optimum classification using the K-means methodology, and the resulting clustering was discussed in connection to the NAO pattern. They used monthly precipitation data from 148 stations covering the 1977–2006 period. In another study, Ünal et al. (2003) used hierarchical CA to redefine the climate zones of Turkey. They examined monthly precipitation and temperature data from 113 climate stations for the period of 1951–1998 and found that the Ward method yielded more accurate results as compared to the other methods. In the same study, Ünal et al. (2003) identified six different precipitation regions but found considerable differences in the boundaries as in many other similar clustering studies. Türkeş & Tatlı (2011) applied the spectral clustering (SC) technique involving singular value decomposition (SVD) and the K-means method to precipitation totals of 96 stations across Turkey for the 1929–2007 period. Their research concluded a total of eight clusters of coherent precipitation zones. In another study, Iyigun et al. (2013) used the Ward method hierarchical clustering technique for grouping common features of different climate parameters covering the 1970–2010 period at 244 stations in Turkey and determined 14 clusters of climate zones.

Although Turkish precipitation has been classified using a number of different methods over the last three decades, none of these studies covered the period after 2015 when the impact of climate change and variability gained considerable interest and visibility. In addition, most of those studies lacked using a sufficient number of stations and time period length to represent precipitation variability over the country, and more importantly, they did not investigate the dynamic behaviour of the precipitation clusters in both time and space. Most previous studies have identified different results of climate zones in Turkey using a varying number of stations and time periods spanning several decades. The clustering techniques used in those studies were typically only used within time or space domains. These results and the diverse nature of the precipitation distribution over Turkey necessitate a better understanding of the classification groupings for various societal needs and to further recognize the variability as time progresses. This study, therefore, aims to explore the spatial dynamics of precipitation regions across Turkey by clustering precipitation station data at varying time scales (5- and 10-year time scales) to capture the decadal and inter-decadal variability. It is our view that studies only analysing the long-term precipitation data will not reveal the exact nature of dynamics of the precipitation regions over time. In other words, we seek to simultaneously cluster homogenous precipitation regions of Turkey spatially over time, producing sub-clusters for each period selected at 5- and 10-year time scales along with the time period of 1980–2020. Such findings can provide valuable information to assess the implications of climate variability for the country's water resources, energy, and agricultural sectors considering the variability of Turkish precipitation even at a 5-year time scale. The variable behaviour of the precipitation regions at short temporal scales can cause further challenges, particularly for water-dependent sectors such as agriculture and hydropower generation in addition to the irregularities and seasonality of the monthly precipitation of Turkey. Therefore, this study makes significant progress over previous studies by encompassing a wider temporal and spatial coverage with additional stations, including the recent decade, and avoiding the assumptions made with the typical application of the K-means method. Moreover, connections between monthly precipitation totals and prominent teleconnection patterns, including AO, NAO, and El Niño–Southern Oscillation (ENSO), are investigated to explore whether any significant correlation exists between them. To relate and compare the spatial variability of the clusters with the relative variability of the station-observed precipitation data, the coefficient of variation (COV) analysis is also performed for the whole period as well as 5- and 10-year sub-periods. Here, we aimed to demonstrate how the clustering structures relate to the COV patterns in order to describe their spatio-temporal variability at 5- and 10-year time scales. The study also aimed to underline the importance of avoiding assumptions about choosing a single distance measure for the K-means clustering approach and argued that choosing an appropriate distance measure should be decided based on the silhouette analysis.

The monthly time series of the teleconnection indexes used in the analysis were obtained from the National Oceanic and Atmospheric Administration's (NOAA's) Climate Prediction Center (https://www.cpc.ncep.noaa.gov/data) and the University of East Anglia, Climate Research Unit (https://www.uea.ac.uk/web/groups-and-centres/climatic-research-unit/data).

A common requirement of the clustering technique in meteorological applications is the necessity to group measurements into meaningful categories based on observations. In this study, the K-means algorithm is used to partition the monthly precipitation data representing 234 stations across Turkey into a set of spatial clusters. The aim here is to group the monthly precipitation data into clusters based on similarities between the data points and explore underlying spatial patterns. K-means clustering generates a single level of clusters rather than a multilevel hierarchy of clusters. Therefore, it is considered to be more suitable than hierarchical clustering for large amounts of data (Bock 2007). Despite its easy implementation and suitability for datasets with a large number of variables, the K-means method has its own drawbacks. One of the main drawbacks of the algorithm is the uncertainty with the choice of an initial point, which considerably influences the final results (Nisbet et al. 2009; Rodriguez et al. 2019). Different selections of the initial point directly affect the accuracy of aggregation. Another drawback of the method is its sensitivity to outliers.

With the implementation of the K-means method, observations are treated as objects having locations and distances from each other. The objects are assigned to mutually distinctive clusters, and each cluster is designated based on its centroid or centre point. As the distances used in clustering often do not represent spatial distances, a silhouette analysis is performed to decide the partitioning distance between the clusters. A silhouette plot specifies a measure of how close each object is to its assigned cluster in comparison to other clusters, and in this manner, the plot provides a way to assess the appropriate number of clusters and cluster quality visually (Rousseeuw 1987; Kaufman & Rousseeuw 1990). The thickness of the silhouette plot with similar sizes, the maximum silhouette value, and the lowest negative silhouette number are the main criteria to determine an optimum cluster number.

Based on the proposed methodology above, the precipitation data were processed for the entire 1980–2020 period and 5- and 10-year sub-periods as well to capture the decadal and inter-decadal structure of the precipitation clusters. In the next phase of the study, a correlation analysis was performed in order to establish links between monthly precipitation totals at 234 stations and prominent teleconnection patterns which have a distinctive influence on Turkish precipitation. To quantify the relationship between the teleconnection indices and precipitation totals, we calculated linear correlations between the monthly teleconnection indices for the three teleconnection patterns, namely the AO, NAO, and ENSO, and the precipitation time series at each station. The resulting connections were analysed to gain some insights into the strength of the linear relationship between the station precipitation and the teleconnection indices. The spatial distribution of the stations that correlated significantly with the single and/or multiple teleconnection indices and the percentage of the stations in each region that correlated with each teleconnection pattern is also examined.

Lastly, in order to relate the spatial variations observed with the K-means clusters to the variability of the station-observed precipitation data at respective sub-periods, the COV analysis was carried out. The resulting COV patterns were visually compared to the spatial shifts in the clusters for the 1980–2020 period and 5- and 10-year sub-periods. The aim here was to identify if any coherency exists between the COV patterns and the clustering configurations.

Considering the thickness of the silhouette plot, the maximum silhouette value, and the lowest negative silhouette number, the silhouette analysis yielded a total of five clusters as the optimum number of precipitation clusters for the 1980–2020 period. In Figure 4, the thickness of the silhouette plot (cluster size) is illustrated. Here, n_clusters equal to 5 yield the best average silhouette score of 0.287 along with all the clusters above the average, indicating that the resulting five clusters are reasonable. In Figure 3, the right panel indicates whether the clusters are located in the lower variance region or the higher variance region. More importantly, it indicates if the boundary among the clusters is well-defined or not. This description applies to all the silhouette plots given in the analysis.

Similarly, a silhouette analysis was performed for the four 10-year sub-periods in order to determine the number of K-means clustering for each sub-period. In doing so, all the silhouettes were visually examined based on different distance metrics as previously described in the analysis of the full-time period (1980–2020). Our analysis indicates that the cosine distance method gave the best result for the three sub-periods, including 1980–1989, 1990–1999, and 2010–2020. Only for the 2000–2009 period does the correlation distance method give the optimum number of clusters (Table 2).

The silhouette analysis was again performed to obtain an accurate number of K-means clustering for the eight 5-year sub-periods in a similar way used for the 10-year sub-periods and the full-time period. Our analysis indicates that the cosine distance method gave the best result for the six 5-year sub-periods, including 1980–1984, 1985–1989,1990–1994, 1995–1999, 2000–2004, and 2015–2020 (Table 3). For the 2005–2009 and 2010–2014 sub-periods, however, the correlation distance method identified the cluster numbers more accurately.

When we refer to the silhouette plots of the clusters at the respective sub-periods, it is concluded that the ‘cosine’ distance measure provided a better interpretation of the clustered data for the most respective sub-periods, except for a very few cases when the ‘correlation’ distance method was the resulting distant metric.

While the clustering structure exhibited some influence from the physiographic features of Turkey, the structure of the precipitation clusters does not exactly fit those features and exhibits some significant shifts (Figure 7). These physiographic regions (also called typical geographical regions of Turkey) are displayed in Figure 8. The regions are referred to in the study describing shifts in the precipitation clusters.

When the 10-year subsequent periods are taken into account, four clusters are obtained for the 1980–1989 period, six clusters for the 1990–1999 period, six clusters for the 2000–2009 period, and seven clusters for the 2010–2020 period, based on the silhouette values and the cosine distance method (Table 4). Overall, the 10-year cluster structures differ from the 41-year cluster pattern described in the previous section considering the entire 1981–2020 period. Both the number of clusters and their spatial coverage change slightly to moderately in the sub-periods. It should be noted that the last sub-period (2010–2020) consists of 11 years just to update the period until 2020 instead of the typical 10-year period

The overall number of the clusters for the 5-year sub-periods varies between three and eight with the cosine distance method resulting as the optimum for five sub-periods and the correlation distance method optimum for three sub-periods (Table 5).

The most noticeable change in the cluster structure during the 2005–2009 period is an enlargement of the Southeastern Anatolian cluster (Cluster label 3) extending further north and west, causing a reduction in the EA cluster size. During this period, Marmara, Aegean, and Mediterranean regions formed one single cluster (Cluster label 5). The coastal Black Sea appears as a unique cluster (Cluster label 4) and the central Anatolian cluster (Cluster label 2) nearly resembles the physical geographic boundaries of the region, except for a northward enlargement observed towards the western Black Sea region.

The 2010–2014 period is characterized by five clusters, and two of them indicate spatial enlargement beyond their typical geographic boundaries. One of them is the coastal Black Sea cluster (Cluster label 5) which extends further west, occupying large parts of the Marmara region. Similarly, the Southeastern Anatolian cluster (Cluster label 1) tends to expand further west occupying the coastal Mediterranean area. The last period encompassing the 2015–2020 sub-period portrays striking clustering features. This last sub-period consists of six years just to update the period until 2020 instead of the typical 5-year period. We should keep in mind that Turkey observed its highest annual mean temperature anomalies and decreased rainfall trends during this last 6-year period. The semi-arid central Anatolian cluster (Cluster label 2) expanded its boundaries southward, reflecting the warming temperatures and below-normal precipitation of the last decade. The moist north Black Sea cluster (Cluster label 3) stretched further westward, causing the retreat of the less moist Aegean climate zone (Cluster label 4) back to its typical geographic boundaries. Moreover, the continental form of the Mediterranean climate covering Southeastern Anatolia also occupied the coastal Mediterranean zone as was the case in the previous 5-year sub-period (Cluster label 1 in both sub-periods). Overall, the clustering structure during the last 6-year period changed remarkably with the impact of the warming and less precipitation than the normal observed.

Teleconnections are recurring and persistent disturbances in the atmosphere that cause weather and climate anomalies over many areas of the globe on both inter-annual and decadal timescales (Nigam & Baxter 2015; Feldstein & Franzke 2017). The AO, the NAO, and the ENSO are three main teleconnection patterns that have proven to influence the larger Mediterranean region, including Turkey. The Mediterranean Oscillation (MO) and the North Sea Caspian Sea Pattern (NCP) are also recognized to influence the Turkish climate, but they were not taken into consideration in this study as their data series were of insufficient length to be correlated with the precipitation data.

The teleconnection patterns affect temperature and precipitation over the Mediterranean region, especially by regulating the moisture transport from the oceans into the Mediterranean area (Mariotti et al. 2002; Lionello et al. 2006). Changes in storm tracks associated with teleconnection patterns can result in wet and dry conditions across Eurasia. Several studies argued that NAO profoundly influences the Mediterranean basin precipitation patterns (Adjez 2000; Jacobeit et al. 2001; Dünkeloh & Jacobeit 2003). During a positive phase of the NAO, Turkey is influenced by a cooler and drier climate (Cullen & deMenocal 2000). Another study also highlighted that drier than normal conditions prevail during a positive phase of NAO over parts of southern Europe, including the northern Mediterranean countries (Marshall et al. 2001). The link between the atmospheric teleconnections and the climate of Turkey has also been thoroughly investigated by several studies (Kahya & Karabörk 2001; Türkeş & Erlat 2003; Karabörk et al. 2005; Karabörk & Kahya 2009). Türkeş & Erlat (2009) questioned the link between the NAO and winter temperatures and found significant negative correlations. The AO is another teleconnection pattern that impacts the mid-latitude climates. A strongly negative phase of the AO is associated with warm weather in high latitudes while bringing cold and stormy weather to the more temperate regions (Thompson & Wallace 2001). During the positive AO phase, a surge of arctic cold air and enhancement of the polar front jet (PFJ) bring colder temperatures to Eurasia and North America, while northern Europe gets wetter and the Mediterranean area drier (Hurrell 1995; Thompson & Wallace 2000). A negative AO is distinguished by the southward displacement of arctic cold air and the southward shift of PFJ, causing low temperature and high precipitation in mid-latitude regions. The displacements of both are considered to be associated with the weakening of the polar vortex over the Arctic region (Kömüşcü & Oğuz 2021). The Mediterranean region receives more moisture during the negative AO phase with storms originating in the central Atlantic Ocean. Some studies identified links between AO and Turkish precipitation (Gong et al. 2001; Türkeş & Erlat 2008; Choi & Byun 2010). In a more recent study, Sezen & Partal (2019) demonstrated the presence of a significant correlation between AO and winter precipitation in western parts of Turkey and claimed that the Atlantic depressions advect more moisture to the Mediterranean region during the negative AO phase. Of the third teleconnection pattern examined, ENSO is believed to have some impact on precipitation anomalies over the Euro-Mediterranean region. The oscillating warming and cooling patterns of ENSO influence not only the rainfall distribution in the tropics but also the weather patterns across many parts of the world (Halpert & Ropelewski 1992; Ropelewski & Halpert 1996; Davey et al. 2014; Lin & Qian 2019). Several studies highlighted the ENSO-associated significant changes in the frequency and intensity of rainstorms and rainfall anomalies in the Euro-Mediterranean region (McPhaden et al. 2006; Brönnimann 2007; López-Parages & Rodriguez-Fonseca 2012). Kadıoğlu et al. (1999) studied the potential effects of the ENSO warm patterns on the Turkish monthly precipitation totals and argued that the high ENSO index has distinct impacts on the seasonal variability of Turkish precipitation. Kiladis & Diaz (1989) related the wetter or drier than normal conditions observed in the eastern and northeastern parts of Turkey to the different phases of El Niño. They noted that wetter than normal conditions over northeastern Turkey coincided with the El Niño warm phase, while drier than normal conditions occurred over eastern Turkey the year following an El Niño year. Lastly, Kahya & Karabörk (2001) found coherent and significant streamflow responses to the El Niño and La Niña events in Northwestern Anatolia (NWA) and EA.

Overall, we argue that in addition to the physiographic features, climate variability resulting from teleconnection patterns, such as AO and NAO, influences the precipitation patterns over Turkey. The change in the spatial patterns of the precipitation regions can be associated with different regional precipitation mechanisms that are controlled by the varying atmospheric circulation patterns, including teleconnection patterns operating at an inter-annual scale. Some previous studies indicate that NAO usually has ∼6–8 years of variability with the highest power for a periodicity of 7.7 years (Zhang et al. 2011; Deser et al. 2017; Rust et al. 2022). In an earlier study, Jevrejeva et al. (2003) showed that NAO and AO presented high coherency in the 2–3.5, 5.2–7.8, and 12–20 year bands. Zhang et al. (2011) concluded that the spectrum of the NAO presented different frequency domains seasonally, did not reveal the classical peaks at about 2.5 and 6–10 years in winter, and represented the enhanced power at periods 6–10 years. The same study also concluded that ENSO (Niño 3.4 index) represents statistically significant power for a band of ∼2–7 years, indicating quasi-periodic behaviour. Moreover, Pinault (2020) found that the Southern Oscillation Index (SOI) signal has an average period of about four years and a broad peak prolonging beyond 6.5 years based on the Fourier spectrum analysis. The variability of the teleconnection patterns mentioned here is thought to influence change in the spatial patterns of the precipitation clusters, which needs to be further detailed.

The COV analysis aimed to gain insights into how the identified cluster patterns relate to the dispersion in the observed station data in order to describe their spatio-temporal variability at varying time scales. We especially looked at shifts in the cluster patterns at subsequent time periods in relation to the variability of the COV patterns. It is our view that identifying such a relationship between the two will provide some understanding of changes in the precipitation regime at decadal and inter-decadal scales.

The 5-year sub-period analysis indicates that the precipitation over the Black Sea coastal areas is becoming more uniform throughout the year and lacking significant year-to-year variability, while the coastal Mediterranean region has more irregular precipitation during the year as it receives a great portion of its precipitation in the winter months. Overall, we can argue that no significant spatial shifts in the cluster patterns are observed in the lower COV areas (i.e. Black Sea region) where the monthly precipitation is more uniform throughout the year while larger inter-decadal variability is observed in the higher COV areas (i.e. Mediterranean region) where seasonality in precipitation is more pronounced. We find that the precipitation regime of a region with more irregular behaviour and a dominant seasonal component is more vulnerable to spatial shifts than a region that has a more uniform monthly precipitation distribution throughout the year.

In this study, the K-means clustering is applied to define precipitation regions of Turkey and identify variations in the clusters' spatial extensions in subsequent 5- and 10-year periods because of climate variability. Our results found that the number and the spatial domains of the precipitation regions changed considerably throughout the study period reflecting regional climate variability impacts. As the domain of the clusters expanded or contracted, some stations did not retain their original position within a cluster but rather exhibited spatially variable behaviour as they were classified under another cluster when the sub-periods changed. The main conclusion of the study is that the precipitation regions are not spatially coherent over time and considerable variability can be manifested even between different regions of the country at an inter-decadal time scale.

Five distinct precipitation clusters (Figure 7) are revealed when the entire 1980–2020 periods are taken into account, each reflecting different characteristics of Turkish precipitation. The identified clusters include:

• 1.

  SCM cluster

• 2.

  EA cluster

• 3.

  Subtropical semi-humid/humid coastal Mediterranean (SS/HCM) cluster

• 4.

  HBSC cluster and

• 5.

  SCCA cluster

The SCM cluster appeared to be a more continental form of a typical Mediterranean precipitation region with high summer temperatures. The SS/HCM cluster differed from the SCM exhibiting wetter conditions and extended north to cover the western Marmara region. The semi-arid central Anatolian precipitation cluster encompasses the largest precipitation cluster. The wetter Mediterranean precipitation region consisted of one single region with an extensive spatial coverage. The resulting cluster structure reveals the semi-arid character of the Turkish climate extending towards the north and the east which are more humid than the central Anatolian parts of the country.

Clustering of the 10-year subsequent periods resulted in the number of clusters varying between four and eight indicating considerable decadal variability. During the first 10-year period (1980–1989), Marmara, Aegean, and Mediterranean regions appeared in a cluster matching the general pattern except with less spatial extent. The last period including the 2010–2020 years displayed more diverse cluster structures with a total of seven clusters. During these 10 years, the Eastern Mediterranean region is characterized as a separate cluster, whereas the Aegean and Marmara regions formed two different clusters during the 2010–2020 period.

The number of clusters for the 5-year sub-periods varied between three and eight, portraying a more variable pattern than the 10-year sub-period patterns. The main clustering feature observed during the 5-year sub-periods was the entire Aegean, Marmara, and most parts of the Mediterranean regions that were classified as one single region in the first two sub-periods. Except for the first 5-year sub-period, the continental form of the Mediterranean rainfall regime dominates Southeastern Anatolia, indicating less variability in its spatial extension as compared to the other clusters. Occasionally the Humid Coastal Black Sea cluster, which usually has a well-defined coastal extension, expanded further west into the Marmara region forming one single cluster in some 5-year sub-periods. The 2015–2020 period portrayed the most striking cluster features of all 5-year sub-periods that are characterized by semi-arid features in Central Anatolia that expanded its boundaries reflecting warming temperatures and reduced precipitation. Moreover, the moist north Black Sea zone extended further west during the last 5-year sub-period occupying the Marmara region, causing the retreat of the less moist Aegean climate zone.

The variable precipitation patterns observed in the decadal and inter-decadal periods highlight the dynamic behaviour of Turkish monthly precipitation. Not only do the number of precipitation regions change, but also their spatial coverage enlarges or shrinks even during 5-year sub-periods. The variability of the clusters may also affect precipitation regimes in the future. For example, if the Humid Coastal Black Sea cluster keeps expanding westward in future periods, it can modify the precipitation regime of the Marmara region and cause more uniform precipitation distribution and less inter-annual variability. Similarly, penetration of the SCCA cluster northward may bring the more semi-humid and semi-arid character of the SCCA precipitation zone into the Black Sea region and alter its uniform precipitation distribution pattern. Such variability can have significant implications for water availability and water allocation for various water-reliant sectors in the country as the precipitation regime will change during inter-decadal and decadal periods. Future climate change may exert additional pressure on the country's water resources, and demand for water by different sectors might increase during the next 2–3 decades. Therefore, analysis of the regional precipitation variability using the observed station data can shed a light on future studies to project such variabilities with more accuracy and assess its implications on the country's water resources more efficiently.

It is our view that although the physiographic factors, such as topography and elevation, influence the general pattern of the precipitation regions, climate variability resulting from teleconnection patterns, such as AO and NAO, impacts the dynamics of the precipitation patterns over Turkey. The monthly precipitation totals were correlated with AO more significantly in the coastal regions except for the Black Sea region, while the NAO yielded a better correlation at Eastern and Southeastern Anatolian stations. All the stations in Southeastern Anatolia were significantly correlated with NAO, while the Marmara region exhibited the highest percentage of correlation with AO, namely 87%.

In order to investigate and compare the coherency of the observed station precipitation variability with the spatial changes of the clusters for the respective sub-periods, an analysis of the COV is performed. We attained two main conclusions from this analysis. First, the cluster structures tend to retain their spatial domain in the lower COV areas and do not exhibit a significant shift in subsequent sub-periods. Second, the areas characterized by the lower COV have a more uniform precipitation regime, while areas of high COV have more irregular precipitation during the year and receive a great portion of their precipitation in the winter months. In other words, COV values tend to be higher in regions (such as the Mediterranean region) where the marked seasonal regime and winter precipitation are dominant, while they decrease in areas (such as the Black Sea region) where precipitation is relatively uniform throughout the year. An important conclusion reached based on the COV values is that the cluster structures are more stable in areas where the lower COV values dominate, such as in the HBSC cluster, while the variability of the clusters in the subsequent sub-periods increases in the areas where the COV values are higher, such as in the Southeastern Anatolian and Mediterranean region clusters, identified as the SCM cluster. In summary, we find that the areas with highly variable precipitation regimes tend to shift more spatially. The analysis of the spatial distribution of the COV values can also be crucial for understanding the risk of extreme events. The regions identified with higher inter-decadal and decadal variability in precipitation may be more susceptible to drought events.

Another important conclusion reached in the study is the selection of different distance methods for determining the optimal number of classes for silhouette analysis using the K-means clustering technique. Our study found that distance measures may vary among the different sub-periods due to the inherent nature of temporal variability of precipitation, and therefore any assumption of choosing a single distance measure by default should be avoided and an appropriate distance measure should be searched. In this study, primarily the cosine method appeared to be an optimum distance method except for very few cases when the correlation distance method was the resulting distant metric. The optimum distance metrics were determined based on the optimum cluster number, considering the minimum negative values and the maximum average silhouette value for the respective sub-periods.

In order to thoroughly investigate the factors affecting the spatial clustering of the precipitation regions and, more specifically, how the clusters are structured at decadal and inter-decadal time scales, additional research is needed, especially concerning climate variability. The authors plan to focus future work on a more detailed investigation of the role of the teleconnection patterns and large-scale atmospheric patterns on the variability of precipitation clusters, given the results found in this study. Moreover, it is unclear how future changes in climate may affect the current clustering structure. Climate projections indicate decreasing rainfall throughout Turkey, particularly in the southern and inner parts of the country, and irregularities in rainfall regimes have already been noticed (Demircan et al. 2017). If such changes impact the amount of local and regional precipitation, the defined clusters may eventually shift and reorient. Therefore, further attention needs to be focused on how the projected climate change may redistribute the existing precipitation regions.

The spatial patterns in precipitation regions and their variability as defined in this study have significant implications for the water-related sectors in Turkey considering the irregularities and seasonality of the monthly precipitation. Hence, it is our view that understanding temporal and spatial changes in precipitation patterns and their groupings is a critically important aspect of planning and management of water resources including planning for drinking water supplies, groundwater, rivers, hydroelectric production, agricultural activities, and other water-related sectors in the country in future. Such information also provides invaluable input to authorities in taking adaptation and mitigation measures to minimize the impacts of extreme climatic events such as floods and droughts common in recent years. We expect that the results of this study will provide essential scientific settings when preparing the country's medium and long-term water resources, and agricultural and drought management plans. Last but not least, knowledge of the homogenous precipitation regions’ identification and their inter-decadal variability consideration can pave the way for a better scientific basis to understand the role of climate variability on the precipitation regime in different temporal scales over the country.

The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.

All authors contributed to the study's conception and design. Material preparation, data collection, and analysis were performed by A.Ü.K. and E.T. The first draft of the manuscript was written by A.Ü.K., and all authors commented on previous versions of the manuscript. T.D. edited the final version of the manuscript. All authors read and approved the final manuscript.

MATLAB codes for the K-means analysis are available upon request.

Data cannot be made publicly available; Data is available upon request of the authors.

The authors declare there is no conflict.