Regional Clusters, Similarities, and Changes in Turkey’s Wood Production

This study aimed to separate the wood production in regions and provinces of Turkey into homogeneous groups based on similarities by using the country’s wood production figures for 2013 and 2018. Within this context, the hierarchical Ward’s and non-hierarchical K-means clustering methods were used comparatively. Clustering analyses of 2 to 6 in number were performed via both methods, and the same regions mostly fell into the same cluster groups, although in different cluster combinations. The results showed that some provinces with rich forest areas did not produce enough wood. It was observed that these provinces were in the same clusters with provinces having a low amount of forest areas and low wood production. Over the five-year period, very few provinces and regions differed in line with the previous development plans. The creation of a spatial database for wood raw material production using the findings obtained in this study will contribute to the development of operational inventory methods that can be included in long- and medium-term forestry plans.


INTRODUCTION 1. UVOD
Forests are important examples of sustainable natural resources. A great variety of products can be put on the market by cutting down trees, and the earned income provides good capital to produce more products for the coming years (Tietenberg, 1996;Koulelis, 2009).
Turkey is prominent among the countries with a high utilization rate in regard to rich forest resources and forest products (Istek et al., 2017). The latest report on Turkey's forest cover was published in 2015. According to this report, the forest area that amounted to 21.5 million ha in 2010 had reached 22.3 million ha in 2015. This number comprises 29 % of the country's surface area of approximately 78 million ha. According to the data of 2015, 57 % of the forest area (12.7 million ha) with canopy closure of above 10 % is classified, in terms of wood raw material production, as fertile forest, whereas the rest (43 % -9.6 million ha) with a canopy cover of less than 10 % is considered as infertile forest land, also referred to as unproductive or degraded forest (TAF, 2019). The official database of the General Directorate of Forestry (GDF) reported that by 2020 the fertile forest area reached 13 million ha, and the total forest cover was 22.7 million ha (GDF, 2020).
The wood production and marketing policies of Turkey are based on developments in the market and the raw material expectations of the forest industry. The GDF has increased production substantially when the developments in the economy, the growth potential of the construction sector, and the capacity of the industrial sector to expand are taken into consideration (GDF, 2016). About 75 % of Turkey's wood production is supplied by State forests, both legally (60 %) and "off the books" (15 %), 19 % by private sector production, and 6% by import (Ministry of Development, 2014; TAF, 2019). Figure 1 presents the changes in Turkey's wood production for the years 2013 and 2018. An examination of the figure draws attention to the increase in all production types except for thin poles and fuelwood. Overall, over the five-year period, log and fiber-chip wood display the highest amount of production, whereas telephone poles (118 %) and logs (54.5 %) have the highest rate of change.
Long-and medium-term forestry plans are prepared according to the principle of efficient use of the forest, and the production of wood raw materials is also regulated in accordance with this principle. When preparing these plans, attention must be given to the density and clustering of industrial plantations. It is important that density groups be technically and economically appropriate, environmentally tolerable, and socio-economically and institutionally acceptable. With these features in mind, this study separated Turkey's geography into homogeneous groups according to the similarities in wood production for the years 2013 and 2018. By analyzing the changes of homogeneous groups over the specified years, the aim was to determine to what extent the plans and arrangements made for the sustainability of wood production had been effective.  cover the last year of Turkey's five-year 9th Development Plan (2013) together with the last year of the 10th Development Plan (2018). In this way, the change in the forestry policies of the country could be followed more easily. Therefore, in order to promote the efficient use of forest resources, it would be possible to determine the necessary changes to be made in the distribution of industrial plantations according to regions. Thus, an effective forest products database could be established to render forestry activities more efficient and to facilitate the follow-up of these activities. This study presents the cluster analyses carried out using K-means and Ward's clustering methods, and the regional and provincial changes observed over the two different periods. Products Model (GFPM) based on forest products, including plywood, particleboard, paperboard, newsprint, printing and writing paper, and other paper and cardboard. Within this context, they used K-methods and silhouette clustering methods. Furthermore, Caridi et al. (2012) investigated the Italian furniture industry's supply chain preferences depending on product modularity and innovativeness. They compared supply chains of firms offering products with modularity and innovativeness at different levels using the K-means method and Pearson distance with factor analysis for clustering. The results revealed that both product features should be taken into account when designing a supply chain. In addition, Hitka et al. (2017) developed motivation programs for management and employee groups at a medium-sized wood-processing enterprise in Slovakia using hierarchical clustering analysis. In the study, in which three motivation-oriented clusters were determined for both groups, they indicated that the existing program of the enterprise was incorrectly designed and would have negative effects on personnel. They asserted that their own program would meet most personnel needs and increase the performance of the firm. Akyuz et al. (2019) also used hierarchical clustering and discriminant analysis to research the amount of industrial wood production in regional forest directorates in terms of similarities. According to the clustering analysis results, regional forest directo-rates could be divided into a maximum of six and a minimum of two groups. In another study, Fang et al. (2021), using hierarchical cluster analysis, classified poplar clones into different categories according to their growth performance, crown structure, and wood properties.
In addition, Keskin

MATERIJALI I METODE
The dataset of this study consisted of 81 wood production values in seven different geographical regions of Turkey. These data were obtained from the GDF and cover the years 2013 and 2018. Figure 1 gives Turkey's total output value for the specified years. In order to perform clustering analysis on the basis of regions and provinces, the production values of the provinces for logs, telephone poles, mine poles, industrial wood, pulpwood, fiber-chip wood, and thin poles were used. In the analysis phase, the years in question were evaluated independently from one another. Thus, the similarities and differences between 2013 and 2018 were observed. Within this context, first, it had to be determined whether or not the data showed normal distribution. The data regarding variables in the study did not comply with normal distribution, and a high level of positive skewness was found. Therefore, logarithmic transformation, which is used in cases of positive skewness, was applied to the data. As for fuelwood data, it was normalized by converting it to a range of 0-1. After logarithmic transformation, the skewness and kurtosis values of the data for wood production were between +1.5 and -1.5, which is considered normal distribution in the literature (Tabachnick et al., 2007; Eryilmaz and Kara, 2018) ( Table 1).
The two clustering methods used in the study were Ward's hierarchical method and the non-hierarchical Kmeans method. In the K-means method, the Silhouette Index was used in order to determine the number of clusters. The Silhouette Index values for wood production are presented in Table 2. The literature indicates that a Silhouette Index of more than 0.5 reveals that the clustering was successful within reason, and a value exceeding 0.7 indicates highly strong clustering (Ng and Han, 1994). Regarding the index values, two is the optimal number of clusters, with significant clustering achievable for up to six clusters. Consequently, a higher number of clusters would be better for observing and comparing regional changes. Therefore, the number of clusters in the study was selected as six.

Clustering analysis 2.1. Analiza klastera
Clustering analysis provides the categorization of units investigated in a study by grouping them based on their similarities, presenting their common features, and determining general definitions related to these categories (Kaufman and Rousseuw, 2009; Dinler, 2014). In parallel with discriminative analysis, it puts similar individuals in the same groups, and similar to factor analysis, it gathers similar variables in the same groups (Cakmak, 1999; Kizgin, 2009).
Clustering analysis methods are divided into two main categories: hierarchical and non-hierarchical clustering analysis. The hierarchical Ward's method is frequently used in clustering analysis and is considered to be a method that gives the best results (

K-Means method 2.2. Metoda K-prosjeka
The K-means clustering method (MacQueen, 1967) is widely used to divide a data cluster into k groups automatically (Wagstaff et al., 2001). The Kmeans method can be briefly described as creating various sections from a series of data and evaluating these sections via a specific standard (Tekin and Te-melli, 2020). In this method, the k value is identified beforehand, and random points are then selected as cluster centers. All the samples are assigned to the closest cluster center based on the normal Euclidian distance metric. After that, the center of samples in each cluster is calculated. Those centers are accepted as new center values for their own clusters. Finally, the whole process is repeated with the new cluster centers. Repetition continues until points are assigned to each cluster in successive clusters/tours, after which the cluster centers are fixed and remain the same forever (Kilic et al., 2020). The K-means assignment mechanism allows each data item to be assigned to only one cluster. Therefore, it is a strict clustering algorithm (Evans et al., 2005; Şen and Varürer, 2019).
In the K-means method, the objective function is minimized using Eq 1 given below (Tucker et  (1) Here, S j is the data point cluster, c j is the center of the S j cluster, x i is a data point belonging to the cluster, and k represents the number of clusters indicated by the user beforehand.
Although the K-means method has a great advantage in its ease of implementation, it also has some disadvantages. The quality of the final clustering outcomes depends on the arbitrary selection of the cluster centers at the beginning. Consequently, random selection of the centers at the beginning would give different results for different initial centers. Therefore, the first center should be selected meticulously and thus, the desired clustering should be provided. Moreover, computational complexity depending on the amount of data, the number of clusters, and the number of repetitions is another factor that must be taken into account when designing with K-means clustering (Yedla et al., 2010;Dhanachandra et al., 2015).

Ward's method 2.3. Wardova metoda
This method, also called the minimum variance method, was proposed by Joe Henry Ward (1963). In Here, x i is the score of the i th observation and n is the amount of data (Aldenderfer and Blashfield, 1984; Celik, 2013). As a result of the analysis via Ward's method, clusters are presented in a diagram called a "dendrogram" in which they come together successfully at different levels (Dibb, 1998;Ozturk, 2012). This method is quite effective and responsive to cross points; however, it tends to create small-scaled clusters (Sekerler, 2008). Table 3 gives the means for the final cluster centers at the end of the clustering analysis. High mean values here indicate the clusters where the wood production in question is intense, whereas low values represent the clusters where production is lower compared to other clusters. In addition, the data in the table give information about the reasons for the cluster differences of provinces in groups. For example, even though Clusters-1 and -2 had similar production means in 2013, the fact that the means for telephone pole and thin pole production in the provinces of Cluster-2 was close to 0 indicated that the specified products had not been produced in those provinces and thus, a different clustering was created. Similarly, the differences in Clusters-5 and -6 indicate that almost no production had been carried out in Cluster-6; however, fuelwood and fiber-chip wood were produced in Cluster-5. Table 4 presents the groups formed as a result of the clustering analysis related to wood production in 2013 and 2018. Clustering is observed to intensify in clusters-1 and -2 that include the provinces with high wood production. When the provinces showing cluster changes in the 5-year process are examined, the provinces of Hakkari, Van, and Mus are identified as producing only fuelwood in 2013 and producing nothing in 2018; in Mardin and Nevsehir, on the other hand, no wood was produced in 2013, but fuelwood and fiber-chip wood production began in 2018. Moreover, Bayburt, Nigde, and Elazig fell within Cluster-4 in 2018. They had produced only fiber-chip wood five years earlier, but in 2018 started to produce all wood products except for telephone and thin poles. In the provinces showing changes in Clusters -1, -2, and -3, in addition to the increase/decrease in their wood production, some product groups such as telephone and thin poles had never been produced or begun to be produced.

Rezultati klasteriranja metodom K-prosjeka
Furthermore, it can be stated that the aforementioned clusters also ranked in total wood production; however, the fact that the number of clusters was kept high and there were eight different types of products resulted in some provinces relinquishing one cluster based on their similarities.
The clusters are also presented as colored maps in order to demonstrate more clearly the regional changes in 2013 and 2018 ( Figure 2). An overall examination of the figure shows that Clusters-1 and -2 constitute some districts of the Black Sea, Marmara, Aegean, Mediterranean, and Central Anatolian regions where the forest areas are dense, and the provinces in the clusters apart from these two extend out to other regions of Turkey, with changes in the five-year process seen to occur more intensely in these provinces.

Rezultati klasteriranja Wardovom metodom
The dendrogram of clustering results obtained via the Ward's hierarchical clustering analysis is given in Table 5. When the dendrogram is examined, Turkey's wood production is shown divided into a maximum of four clusters in 2013 and five clusters in 2018.
In the most general categorization, wood production in both 2013 and 2018 is divided into two clusters. Almost all the provinces in the clusters obtained via the Ward's method are in the same cluster as the provinces in the clusters created via the K-means method. These similarities confirmed that Turkey's wood production had been categorized properly and successfully. As distinct from the K-means, the Ward's method reduced the number of clusters only by uniting some clusters. Indeed, the situation was indicated at the beginning of the study with the Silhouette Index values. It was also emphasized that clusters of two to six in number could be successfully categorized. According to Caglar (1990), the fact that the same regions remained within almost the same clusters at different cluster combinations can be considered as an important sign indicating the significance of the findings.

ZAKLJUČAK
This study utilized hierarchical and non-hierarchical clustering analysis methods to separate Turkey's forestry sector into homogenous clusters and investi-gated regional and provincial changes and similarities within a 5-year process.
In the study, the analyses made with both Kmeans and Ward's methods showed similarities. In other words, an increase or decrease in the number of homogenous clusters did not cause any provinces to be placed into different clusters. The provinces divided into six clusters via the K-means method were allocated to fewer clusters in the Ward's method solely to enable the change to be observed more clearly.
The evaluation revealed that during the five-year process, in some provinces such as Nigde, Elazig, Bayburt, Mardin, and Nevsehir, despite having less forest area, wood production had begun, whereas in Hakkari, Van, and Mus, wood production had completely stopped. Some provinces showed changes in their clusters from 2013 to 2018.
The aforementioned clustering results also give information about the effective use of the forest areas. The clustering results determined that some provinces with rich forest areas did not produce enough wood. It  was also observed that these provinces were located in the same clusters with provinces having a low amount of forest area and low wood production. This situation leads to inefficient use of forest resources and consequently, needs to be rectified by taking into account the factors included in the development plans, thus ensuring the sustainability of forest resources. When the clustering results for 2013 and 2018 were compared, the systems and training in wood production did not show sufficient development and therefore, the necessary professionalization could not be achieved. In addition, wood production not only varied according to the type of wood, but also varied according to the climatic differences among regions. In regions where different climatic circumstances are experienced, production activities are losing pace under aggravated working conditions. In order to meet the needs of the wood raw material market, it is important to establish regeneration and maintenance areas within the scope of economic management, taking into account production costs and silvicultural principles. This situation has been partially resolved in provinces where temporal and spatial arrangements have been made and is seen in the change between clusters.
The results of the study can contribute to the development of operational inventory methods by creat-ing a spatial database for wood raw material production. Therefore, it can provide economic and technical integration in terms of making annual applications and monitoring of long-term national forestry or development plans and medium-term forestry plans for management and silviculture.
Because Turkey is in the position of an importer with regard to forest products, it is essential to develop production in a more systematic and projectized way. Thus, imports will decrease, and production will increase in provinces that are rich in forest areas, with those that are unproductive in terms of wood products remaining in low-level clusters. Moreover, in the regions belonging to Clusters-5 and -6 where wood production is either very low or non-existent, building industrial plantations and planting the appropriate species and clones demanded by the country would be an effective way to end the deficit. C.; Üçüncü, T., 2019: Investigation of the similarities of industrial wood production statistics of regional directorates of forestry in Turkey using cluster and discriminant