Turkey OpenStreetMap Dataset - Spatial Analysis of Development and Growth Proxies

1.1 OSM in Turkey

No key study of Turkey OSM dataset regarding (a) spatial evolution with time, (b) different socio-economic factors governing it, (c) mappers involvement within the country, (d) dataset quality is available online. Current study, therefore, becomes essential to answer key elements of the ecosystem on spatial and temporal ground as has already been done by other researchers for developing nations but at a much coarser level. For example, [13], [19] and [48] have done similar spatial and temporal analysis for street networks of Beijing, Kenya and China, respectively. Turkey OSM provides a rich dataset with nearly 17 million points, 1.3 million edges and 0.4 million polygons [49]. In this study, the term edge is used for any line and polyline feature. Authors hope that this will add-on some key findings of developing nations into the OSM literature.

1.2 Five aspects of dataset analysis

The studied five aspects of the analysis are as follows:

1. Time-series spatial evolution of OSM dataset - The spatial growth of the dataset within the region has been studied on an eight year of time-span (2007-2015). This advances the work of [13, 47, 48] for Ireland and China where authors have observed heavy-tailed pattern both for a certain time period and across time. It explains if the data density and inherent mapping activity by volunteered mappers is spatially biased or not.

2. Effect of socio-economic factors on OSM spatial evolution - Five socio-economic factors governing any VGI activity, as discussed by other researchers as well [21, 31], have been compared with corresponding OSM feature density for different regions. These factors are Population Density, Literacy Level, Tourism Activity, Internet Usage, and Human Development Index (HDI). It has been speculated that Internet Usage and Tourism Activity plays a vital role in VGI mapping in any region.

3. Processes governing the evolution of road networks within the country - The street network growth with time has been studied within the whole country by considering the Exploration and Densification mechanism of [50] which argues that any urban street mapping first follows an exploration of unmapped streets and then follows a densification of all nearby secondary and tertiary streets. This mechanism is observed in Ireland and Beijing as well [46, 47].

4. Mapping behaviour of OSM mappers - The density of the number of mappers has been measured to identify the maturity of the dataset of a region. This kind of proxy study has already been done by other researchers as well for different resolutions and regions, like in Germany by [30, 31], on a global scale by [3, 17], in England by [20], in UK and Ireland by [28], in Beijing, China by [13] and in USA by [16]. It has been reported that the number of active mappers, excluding bulk importers, show a strong correlation with data quality and density.

5. Quality of the dataset - Mappers’ participation inequality and various data sources have been identified to provide a general commentary on the Turkey OSM dataset quality and accuracy. Participation inequality has been used as quality proxy by past researchers as well, for example for China [13], for major world cities [21] and for whole Planet OSM dump file [22], where they have reported that it is inversely proportional to the dataset quality.

2.1 Source and format of the dataset used

Another reason for OSM popularity is its various data sources (Full Planet dump file [49], Geofabrik downloads [51], Overpass API [52]) and formats (ESRI-Shapefiles *.shp, Extensible Markup Language (XML) *.osm, Protocolbuffer Binary Format *.pbf), thus, causing high data interoperability. A switch from Creative Commons Attribution-ShareAlike 2.0 to Open Database License (OdbL) in 2012 [53] allowed users and developers to share, modify and use the dataset freely and per use-case [54]. Technological advancements and user requirements, further, caused the upgrade of its Editing API (latest v0.6) by bringing in many state-of-the-art editing functionalities [55]. The object history feature was introduced by Editing API v0.5 [55] and, therefore, one will not find any edits in Planet dump file prior to 2007 (section3.1). The 2012 event to change its license caused the loss of 1% dataset because of conflicting users’ interest [53]. Nonetheless, the dump files are proven to be the best source of OSM dataset to study any time-series evolution [16, 27, 32], since other sources only reflect its latest snapshots for specialized regions.

Full Planet dump file (1.5 TB uncompressed) from September 2015 is downloaded from[49] for this study. The file is in a human readable XML format with three primitive data features, i.e. nodes or points; ways or polylines or polygons; and relations (logical combination of nodes, ways and/or other relations). These features are annotated by tags in a key-value structure of free format text fields [56]. Provincial data of five stated socio-economic factors of the region is downloaded from TUIK (Turkish Statistical Institute) website [57]. Specific downloaded columns from TUIK website are as follows: “Number of students for vocational training school and undergraduate programs of higher education institutions: Graduates / Total – Education – Higher Education” for Literacy Level (fig.3a), “Annual growth rate and population density of provinces – Population and migration – General Population Censuses” for Population Density (fig.3b), “Number of arriving foreigners by province of border gate and mode of transport : Air way – Tourism” for Tourism Activity (fig.3c), “The proportion of individuals regularly using the Internet – Transportation and Communication” as Internet Usage (fig.3e), and “Per capita gross value added (GVA) : Per capita GVA ($) – National Accounts” for HDI (fig.3d). Since the ill-famed Syrian refugee crisis has caused a heavy influx of foreigners into the country via roadways (South-East Anatolia, fig.3e), only airways mode of transport is considered for Tourism Activity (fig.3c). Data to create the road density map of the country (fig.3f) is downloaded from the General Directorate of Highways [58] (section3.1).

Figure 1

Different geometrical features along with corresponding nodes count.

Figure 2

A plot between normalized features density and time. Maps represent density for three different time-slices (2009, 2012 and 2015).

Figure 3

Maps showing the level of different socio-economic factors in the region. Color legends are meant for within map comparison and denotes normalized factors by area. The Road Density map shows a uniform distribution of road network within the country.

2.2 Steps in data processing

OSM XML file could be processed by a variety of command line tools, like osmosis (Java application for reading/writing databases [59]), osmium (multipurpose tool for data interoperability and time-series analysis [60]), osm2pgsql (tool to convert XML data to PostGIS-enabled PostgreSQL databases [61]), osm2postgresql (to simplify rendering with QGIS and other GIS/web servers [62]), osm2pgrouting (to import data file into pgRouting databases [63]) etc. For this study, an osmium based tool, called osm-history-splitter [64], is used to split the Planet dump file using the bounding-box of Turkey. Subsequently, the ESRI’s shapefile of the country is used to further divide the dataset into 81 provinces. Finally, the data was loaded into a PostGIS enabled PostgreSQL database.

It is observed that only 2% of all data points of the region have some sort of tags associated with them. Therefore, to study the spatial evolution of dataset it is divided into the following four categories: (a) Points(all) (all data-points from the dump file), (b) Points(tagged) (data-points with tags associated with them), (c) Edges (all highway features), and (d) Polygons (all building, landuse, and natural features). All developed scripts used to process the Planet dump file can be downloaded from [? ], along with a comprehensive README file.

2.3 Data adjustment and normalization

Researchers, in the past, have already argued that feature count in a region can not be used to measure its density, especially for lines and polygons [18, 50]. In order to compare two elementary geometrical features, i.e. points, lines, or polygons, it is advised to count and compare the number of nodes constituting them [13] (fig.1). Same approach is used in this study to compare different features as well. Furthermore, each node count is divided by the area of the province to normalize it. In order to keep the point, line and polygon features comparable to each other the dataset has not been normalized by the demography of the province. However, researchers like [50] and [47] have argued that to compare only the line features the dataset could be normalized by demography as well.

Bulk import in OSM is defined as follows: Bulk import means more than a few hundred of nodes imported in a short period of time for a large area like a whole country [66]. This definition of bulk import is quite subjective and researchers have used their own definitions of threshold to remove it from dataset before any study [36]. The need to remove this kind of sudden import of dataset is to avoid unexpected data spikes or observations [27, 36]. In present study, authors have defined it as “25,000 nodes contributed by single mapper in a week” and have kept them out of the analysis. 10 mappers causing bulk import in the region are identified by this way and there contributions are removed for any spatial analysis.

In graph theory, the degree k_i of an i^th node is the number of nodes adjacent to it, i.e.

in terms of adjacency matrix [67]. In street network, the degree of a junction is the number of road segments originating from or terminating to it. Thus, frequency of different degree of junctions in a street network represents its topological structure. It explains how densely or sparsely the road segments are connected to each other. Degree distribution or fraction, which is defined as P(k) = N(k)/N where N(k) is the number of nodes with degree k and N is the total number of nodes in the network, is necessary to understand how evolved a network is at any given point in time. In other words, it explains how much the region has been explored and densified by mappers. High fraction of low degree junction in a city explains its Exploration phase as new roads are being identified and mapped, whereas, high fraction of high degree junction explains its Densification phase [13]. These are two elementary mechanisms that govern the evolution of any street network [50].

Finally, per capita GVA (Gross Value Added) ($) value of the region is used for HDI (Human Development Index) calculation. Previous researchers [21] have used per capita GNP (Gross National Product) value of a region to explain HDI. A detailed report of how to calculate HDI of a region can be found in the United Nations documentation [68, 69].

3.1 Time-series spatial evolution of OSM dataset

Fig. 2 shows the time-series (2007-2015) evolution of different features, i.e. their normalized feature density, for 81 provinces of the country. It can be seen that all curves merge into the origin at around April 2007. It is because no object history feature was present in Editing APIs older than v0.5 which was introduced in 2007 [55], meaning no history data before that. Between 2007 and 2012 the slope of curves are mostly gentle (fig.2a,c,d). This is because of limited editing flexibility by old OSM license [53]. However, a sudden spike in slope can be seen from 2012 onwards because of the change of its license to ODbl [13]. This change caused sudden boost in data usage by different web-services, thus, encouraging developers and users to come back to the project to improve it [16]. High slope values for few provinces is because of the large number of active mappers in those regions [13]. It can be further seen that almost all the curves are exponential-step curves, rising each year at around summer time. It is believed that this pattern is because of the increased tourism and outdoor activity in the region during summer time, thus, causing mappers to plot what they observe and collect in the field.

Fig. 2 also shows feature density map of the country for three different time-slices (2009, 2012 and 2015). In general, it can be seen that the eastern and south-eastern parts of the country are less populated than the western and south-western parts. [48] has also reported similar spatial distribution for street networks in China-OSM. Since similar pattern, in general, can be seen for different socioeconomic factors of the country fig.3, they are believed to be the driving forces for mapping activity. Fig. 3 shows the Literacy Level, Population Density, Tourism Activity, HDI and Internet Usage in the country for the year 2014. Fig. 3f is a road density map of the country obtained from the General Directorate of Highways [58]. It can be stated that street network density is quite uniform within the country and whatever spatial bias one is observing in fig.2c is because of the non uniform mapping behaviour by OSM mappers [13, 46, 47]. Few provinces in the eastern and southeastern parts are relatively dense. This is because of few senior mappers actively engaged in mapping activities in those regions. One such province is Batman (red box in fig.2) with its senior mapper from a university, table1. This mapper is responsible for the country’s 3-4% of all geodata upload. Since this mapper is currently a university student, only spike in the year 2015 can be seen.

3.2 Effect of socio-economic factors on OSM spatial evolution

Fig. 4 is a plot between Population Density and features density for three different time-slices (2008, 2011 and 2014). High R² value represents a strong correlation between the two axes. When comparing values based on human activities, values greater than 0.6 is considered to be highly correlated [45]. It can be seen that R² value is high for all features, especially for 2014. Similar observation is drawn for Literacy Level too, although it is not plotted to avoid redundancy and plot cluttering. No strong correlation between Tourism Activity – features density and Internet Usage – features density is found, in contrast to what [21] have reported. However, moderate correlation between HDI and features density is observed, similar to what has been reported by [21]. Overall, the Population Density and Literacy Level of the region are found to be the two strong proxies for features density in OSM, followed by HDI.

Figure 4

Graph between the Population Density and features density for different provinces at three different time-slices (2008, 2011 and 2014) showing an increasing correlation with time.

The R² value of Population Density for all different features increases with time (fig.4). For Population Density vs Points(all) the R² value for the year 2007 is high because of the scattered points lying close to the origin. It can be said that with time, overall, the correlation between socioeconomic factors and OSM features density of a region gets stronger.Most of the data points for earlier years (2008) fall exactly on y-axis. This is because although the population of the region was high, people were unaware of the project and, therefore, mapping activity was quite limited.

3.3 Processes governing the evolution of road networks within the country

Elementary processes governing the evolution of any street network is explained by [50] by devising two processes, i.e. Exploration and Densification. According to this, any evolution first follows an exploration phase where new roads get discovered, followed by a densification phase where roads get connected to each other. This model is also observed in VGI projects [13]. Y-axis of fig.5 is the slope of the curve between degree distribution and time (section2.3). X-axis represents different provinces plotted from the westto east of the country. They are grouped together, according to fig.3e, to reduce plot cluttering. Note that by plotting data points on x-axis in this way, whatever trend we observe from left to right will also explain the west to east trend within the country. Authors have intentionally kept the degree of distribution to be maximum 6 as higher degree of street junctions are not possible in real world scenarios [13]. It can be seen that for 1 (blue line) and 3 (grey line) degree junctions the slope of the plot is positive and negative, respectively. This shows a high slope value between the degree distribution and time curve for western provinces and low slope for eastern provinces. These high slopes explain two events in western provinces: (1)Heavymapping in short time interval, (2) Shift from exploration to densification phase. For eastern provinces the opposite is true. Only 1 and 3 degree junctions are studied as only they exist in majority in developing countries [70] and represent organic street layout [71]. Turkey OSM has followed the Exploration and Densification mechanism of evolution in given time span and this behaviour could be used to predict its future trend. [13] and [47] have reported similar trend for Beijing, China and Ireland, respectively.

Figure 5

Graph showing the Exploration and Densification mechanism as followed by the street network dataset of the country.

3.4 Mapping behaviour of OSM mappers

Fig. 6 is a normalized plot between the total number of contributors in a region and its feature density for three different time-slices (2009, 2012 and 2015). The R² value, again, is a measure of the correlation between the two axes. It can be said that with time the correlation between the number of contributors and features density within the country is

Figure 6

Graphs showing a good correlation between the total number of mappers in a region and its features density.

getting stronger. This is because of the recent popularity of OSM within geospatial community and change of its license to Obdl. For the year 2015, the correlation is quite strong. This proxy has also been discussed by other researcherswho have reported a direct relationship between the total number of points and registered users in a region [21, 30]. It should be noted that bulk importers (section2.3) are not considered in this plot to keep the observation free from any individual bias. The R² value for Polygons for the year 2009 is quite high because of the cluttering of data points close to origin. This is because very limited contributors/mappers and features were there in OSM project in 2009 in Turkey. It is observed that although mappers do contribute few features as soon as they register to the project, usually they take a while to get started with serious contribution.

3.5 Quality of the dataset

Paticipation inequality by mappers is used as a proxy for dataset quality checkup [20]. Researchers have reported that bulk import by mappers make the dataset unsuitable for specialized use cases [21, 28]. Fig. 7 shows the fraction of bulk imports in Turkey-OSM dataset for different features. It can be seen that almost 75% of all geo-data upload corresponds to bulk imports by few mappers, by the end of 2015. There are reported 37 and 12 bulk importers for point/edge and polygon features, respectively. It should be noted that for points(tagged) no bulk importer is identified. This is because tagged geo-data is hard to generate or find open-source. The observed data quality of Turkey-OSM is poor and data correction and quality analysis are necessary before any specific use case.

Figure 7

Pie charts showing the participation inequality in OSM data upload within the country, along with few bulk importers and their respective percentage contribution.

Table 1 contains some basic information of top five bulk importers in the country. Nationality of these mappers gives an idea of their ground knowledge, Technical Background is important to know if they are versed

with information and geospatial technology and Geo-data Sources to know their source of data. These information are collected by in-person communication with them. It should be noted that all of them are using satellite images of proprietary vendors, i.e. Google, Bing and Mapbox, for basemaps. They are also using geo-data from other VGI projects, like wikiloc and geofabrik, to upload into OSM. It shows an inter-dependency of different VGI projects for geo-data sources. It is, therefore, important to upload accurate data to VGI projects as it gets disseminated to

other projects as well,sometimes automatically. Two of the bulk mappers have used a more authentic geo-data source from local municipalities (Konya and Denizli Municipality). These two mappersare responsible for few data spikes in south-eastern province of the country, fig.2. In spite of having high participation inequality in the dataset, it is believed to be suitable for general use cases considering the varied data sources of these mappers.