City, Territory and Architecture

An interdisciplinary debate on project perspectives

City, Territory and Architecture Cover Image
Open Access

Rivers as integration devices in cities

City, Territory and ArchitectureAn interdisciplinary debate on project perspectives20163:1

DOI: 10.1186/s40410-016-0030-4

Received: 28 October 2015

Accepted: 8 January 2016

Published: 21 January 2016

Abstract

Background

As dynamic systems rivers and cities have been in interaction under changing relations over time, and the morphology of many cities has risen through a long and steady struggle between the city functions and the river system flowing inside. This makes river cities an interesting case to study how the presence of geographical features interacts with spatial morphology in the formation of cities.

Methods

The basis of this research is enabled by utilizing a novel model for cross-city comparison presented by Hillier in his Santiago keynote in 2012 called a “star model”. This is done on large samples of cities investigating concurrent configurations, as well as how the properties in this star model react to specific forms of disturbance.

Results

Results illustrate that the foreground network as identified through maximum choice values in cities are more vital to the structure of cities than the bridges. The overall syntactic structure tends to retain its character (degree of distributedness) and the location of its foreground network (which street segments constitute the foreground network) even when bridges are targeted. Furthermore, counter to the initial hypothesis, river cities tend to change less than non-river cities after targeted disturbance of the systems. Finally, the results show that while there is a statistical morphological difference between river cities and non-river cities, this difference is not directly explained through the bridges.

Conclusion

Integrating space syntax with statistical and geospatial analysis can throw light on the way in which the properties of city networks and urban structure reflect the relative effect of rivers on the morphology of river cities. The paper, finally, contributes through offering one piece of a better perception of the structure of river-cities that can support strategies of river-cities interaction as well as enhance our knowledge on the constraints and limits to that interaction.

Keywords

River-cities Morphology Space syntax Structured or distributed city Foreground network

Background

Rivers, as an important element of urban geographical environments influencing and confining the urban structure morphology, have been a vital part in the emergence of societies (Silva et al. 2006). Rivers characterize not only a primary circumstance for growth and development of many settlements but they also affect the artistic quality of urban forms as well as the functionality and size of cities (Čakarić 2010).This paper develops on the morphology of urban space, as a field of scientific discourse on the built environment. The Paper aims to elucidate the effects of the presence or absence of rivers on the morphological configuration of street networks. Thus rivers are considered as solid morphological phenomena with discerning boundaries from the surrounding areas. Therefore the physiography of rivers in their own right, although an important factor that is the focus of a large part of river studies (Kubat 1999; Ellis et al. 2007; Cengiz 2013; Jawarneh et al. 2015), is not examined in this paper.

It is further based on the consideration that analysis of the spatial patterns of settlements can further lead to knowledge about the social norms of societies. As this is said, it is important to clarify that we do not intend to postulate a direct link from a certain spatial pattern to a certain set of social norms; cities need to be understood as complex systems with inherent dynamic parameters, large numbers of discrete actors and a heterogeneous nature (Crooks 2012). This heterogeneity also concerns how a specific city population makes its relation to the spatial pattern through individual and collective processes—the current situation of a city is the result of continuous interactions and the interrelationships between people, space and objects. These processes take also part in structuring cities, as is cognitive, social and mental aspects with various degrees of direct relations to the physical environment (e.g. Portugali 2011, Batty 2013, Frank et al. 2001), meaning that ‘society and space’ affect one another rather than one depending on or emerging from the other. We acknowledge that given this complex nature of the city and its involving characters including spatial and socio-cultural objects, describing the behavior of such a complex dynamic system only with one analytical approach is insufficient. Therefore, as we discuss relations between spatial morphology and society in the coming, the limited methodological focus is seen as offering grounds to start reasoning around on how the analysis can contribute to a by necessity much larger and broader field addressing society-space relations. In the respect of realization and substantiation of the real evolution of the mentioned complex system many efforts have been made by researchers in recent decades leading to a bottom-up approach model called agent-based modeling (Batty 2005, 2008). Thus, even though it can be said that spatial patterns, or morphology of space, is one of the more consistent means through which society reifies itself, the extent to which this is done and what relations in society comes to be embedded, vary greatly (Hillier and Hanson 1984). This forms an important reason why the work of Hillier and others form the basis for the research presented in this paper, even if the specific relation between spatial and social aspects are not closely investigated.

In the extension of the research presented here, a fuller address of how the effects on spatial morphology found herein relates to other parts of socio-spatial relations and patterns is of importance as at this point, many of the social implications can only be inferred. Such an investigation should clearly more concretely take into account individual perceptual and cognitive capacities and its relation to legibility and imageability of space (e.g. Lynch 1960; Frank et al. 2001; Montello 2001; Peponis 2012). To further elucidate how individual cognition, dynamic social interaction processes and emergent complexity (e.g. Portugali 2011) would be affected, agent-based modeling (e.g. Ferber 1999; Batty 2013) would offer significant input to what in this paper remains a base reasoning built on findings on relations between specific aspects of behavioral patterns and syntactic properties in space syntax research. Our aim is simply to refine how we perceive the addition syntactical analysis can offer amongst all the technical, functional, social, spatial and other factors that would need to be studied to address ‘social’ or ‘societal’ effects of disturbances as tested here. Agent-based modeling is here especially of interest since it can account for how cities are complex and highly dynamic in both space and time (Crooks et al. 2014). In this light, we consider spatial configuration as offering a spatial interface (Hillier and Hanson 1984; Koch 2013) for individual and collective action and interaction, which has purchase on the ranges of probable and possible emergent patterns, but where the specific patterns are subject to complex processes reaching beyond what an analysis of spatial configuration can provide. However, as the investigation to be presented here is focused on syntactic properties, it is in order to more thoroughly review earlier related research in this field.

This paper focuses on the spatial properties specifically to advance theory and knowledge on spatial configuration, and to develop further how this can be worked with in large samples of cities. As examples of working with large samples of cities in the context of space syntax we can notice to some research done earlier although the objectives are different from this paper. Early work of Peponis et al. (1989) attempts to decompose cities by choice values to found a syntactic typology. Furthermore Shpuza in a series of papers Shpuza (2006, 2011, 2013, and 2014) study different syntactical measures of Adriatic and Ionian coastal cities prepared for three historical stages between 1800 and 2010. In a recent paper, titled algometry in the syntax of street networks (Shpuza 2014), he tried to explore generic principles of urban evolution and urban growth using quantitative analysis of algometric effect on space syntax properties such as depth, choice, and entropy.

The effects of river systems on the spatial form of cities have not been extensively studied in the field of urban morphology except for some brief and non-continuous research (Silva et al. 2006; Kubat et al. 2007; Mello and Holanda 2009; Asad and Ahsan 2012; Hossain and Bahauddin 2013). As an example, Silvia and their colleagues (Silva et al. 2006) tried to show the relationships between rivers and cities on both national and local scale. They studied the relations between specific parameters such as urban population and size, length of the riverfront, river width, surface of the water body and the number of crossings. The results showed that these parameters influence the structure of river-cities although the degree of influence can be different due to demographic and physical thresholds. In another paper, Mello and Holanda (2009) applied space syntax analysis on the two Brazilian cities in terms of physical and visual accessibility. They used integration values in global and local scale in order to study the degree of urbanity and the contribution of the river to these cities. Their findings show dissimilarity between the two cities in their interaction with river space as manifested in the high degree of urbanity in one city and the low degree of urbanity in another city. They point out the importance of social relationships and the interaction between people and the river side as the configurations that endorse the degree of urbanity in river-cities. Their work therefore points to aspects of river city morphology as analyzed through space syntax methods that complement the research of this paper, where we believe it is crucial to follow up in the future.

Space syntax as a set of theories and tools used for morphological analysis originally emerging in the field of architecture and analysis of settlements (Hillier and Hanson 1984; Hillier 1996; Hillier et al. 2012; Penn et al. 1998) can play a significant role in the research aimed to better understand the way a city’s structure and physical form relates to the presence of a river. One reason is that it allows gaining in-depth knowledge on a particular set of aspects of the morphology through focused analysis that concerns the material configuration of spaces into networks and systems. The interpretation of the configurational network morphology of river cities and the contribution of rivers to the development of their physical form will be of value to planners, urban designers, and architects.

This paper aims to benefit from statistical and geospatial analysis integrated with analysis of syntactical properties of cities, utilizing the former’s tradition to work with large samples of data to gain knowledge of general tendencies and common patterns, while the latter is often concerned with specific understanding of smaller samples or individual cities. Analysis of larger samples does exist and several methods used to enable the integration of geospatial analysis and space syntax in this paper come from this research (Hillier et al. 2012; Peponis 2012; Shpuza 2014). This research aims to –in an explorative way- extract patterns formed by the effect of rivers on urban spatial structures in river cities and compare them to the same morphological properties in non-river cities. In this analysis the river cities form the primary study and the non-river cities the control sample. One purpose is to methodologically and theoretically show how a large sample of river-cities can be analyzed and to begin to unearth if there are morphologically recognizable types or physical forms that coincide with the river-cities. Finally, the paper aims to study these properties under the hypothesis that the limitation to the morphology of urban space placed on cities by the need to bridge the water affects how they react to changes and strategically develop in relation to these particular acts of crossing. In this way we expect that cities affected by the rivers streaming through them show even slightly different patterns in comparison to cities without river.

Finally in this research we try to increase the understanding of relations between architectural and urban morphology and topography, and hereby contribute to the way we understand cities as responses to social, cultural, functional, natural, and technical conditions in a broader sense as outlined above. The aforementioned premises are the reasons why this research is focused on finding possible scientific-theoretical and practical elements of the effects of rivers on the physical form of cities.

The sample: data and area of research

Due to the comparative and explorative structure used in this paper and thanks to the pattern- recognition basis of the research, one of the big concerns in doing this research was collecting a good sample of cities from different sizes and in different countries, a time-consuming, tedious, and costly process that often needs adequate resources, appropriate funding, and sufficient time. Providentially, nowadays with coming into existence of volunteered geographic information, which presents tools for creating and assembling data provided voluntarily by people (Goodchild 2007), GIS researchers and users have access to a massive resource of geographic data called user generated content. In this paper we benefit from the open street map (OSM) database which is a project aiming to produce and freely distribute digitally editable geographic maps and data of the world in different formats through the World Wide Web (Haklay and Weber 2008). We prepared data in the format of shape files from Geofabrik’s free download server (2015) which contains data extracts from OSM project which are normally updated every day and in the different file formats (http://download.geofabrik.de). It should be noted that the data needs some editing to be ready for the analyses, and that this data is susceptible to the quality of user generated content. As the focus is on larger scale phenomena and a large sample of cities, their quality has been judged as adequate for the purpose of the research.

We concentrate on collecting river-cities from small to medium size from different countries in Europe. As shown in Table 1, 42 river-cities from 2000 segments in size to 80,000 segments were collected mostly from France, Germany, UK, Sweden and other countries. In order to further elucidate whether the knowledge we develop on river cities is specific to them or generic for cities, we decide to collect an additional sample of 21 non river-cities with the average size of the river-cities as a control sample for comparison. These cities are from mostly the same countries and a similar size distribution, and have been subjected to the same analysis (Table 1). Due to the emphasis of this paper on river-cities as well as time- consuming process of data collection and data preparation so that they meet our conditions (same size, same location, and reasonably without river or any other egregious morphological phenomena like mountain and seas), we decide to confine the number of non-river-cities to 21, which while a smaller sample it is still statistically significant.
Table 1

Collection of river-cities and non-river-cities studied in this paper

City (with river)

Country

No. of segments

City (with river)

Country

No. of segments

City (without river)

Country

No. of segments

Amiens

France

8910

Lohr Am Main

Germany

8823

Aix -En-Provence

France

27666

Angers

France

25,058

Maghull

UK

1618

Basingstoke

UK

19,825

Aschaffenburg

Germany

24,108

Montlucon

France

7771

Beaun

France

4130

Auxerre

France

4526

Leicester

UK

30,559

Beauvais

France

5486

Bergerac

France

6998

Oberhausen

Germany

26,349

Böblingen

Germany

9971

Blackburn

UK

6854

Örebru

Sweden

18,013

Braga

Portugal

23,529

České Budějovice

Czeck Republic

7057

Orleans

France

7410

Châteaubriant

France

3608

Český Krumlov

Czeck Republic

1953

Prerov

Czeck Republic

2430

Colmar

France

10,237

Charleroi

belgium

29,397

Saintes

France

4428

Coventry

UK

46,253

Châtellerault

Franceance

5078

Saint-Quentin

France

9999

Gloucester

UK

16,098

Cluj-Napoca

Romania

8789

Schweinfurt

Germany

22,085

Harrogate

UK

8500

Criel-Sur-Mer

Franceance

2938

Tours

France

20,561

Issoudun

France

2605

Dax

France

5100

Umeå

Sweden

33,895

Les Herbiers

France

4575

Dresden

Germany

86,828

Valenciennes

France

4304

Luton

UK

14,688

Fontenay-Sous-Bois

France

4434

Venlo

Netherlands

15,779

Mansfield

UK

7757

Haarlem

Netherlands

17,254

Villeneuve

France

6112

Modena

Italy

25,396

Helmond

Netherlands

24,550

Vilnius

Lithuania

43,366

Mönchengladbach

Germany

50,565

Laval

France

7592

Wolfsburg

Germany

38,617

Saint-Étienne

France

13,438

Le Mans

France

21,737

Würzburg

Germany

19,956

Sindelfingen

Germany

15,806

Lille

France

15,431

York

UK

13,179

Soest

Germany

4738

Limoges

France

21,854

Zlin

Czeck Republic

15,984

Wiesbaden

Germany

42,675

Methods

The measures: normalized angular choice, normalized angular integration, and the star model

The basis of this paper is calculating the normalized angular integration (NAIN) value and normalized angular choice (NACH) value. These two key properties in space syntax are comparable to closeness centrality (Bavelas 1950) and betweenness centrality in graph theory (Freeman 1977) (for more information see Hillier et al. (2012). Integration has been defined as the degree to which each segment is accessible from all others in a system. According to Hillier and his colleagues, integration value shows the “to movement” property of a space and reveals how easy and how often a route or segment would be passed through by pedestrians (Hillier et al. 2012; Hillier and Iida 2005). Choice has conceptually been more concretely defined in relation to movement, and is calculated as the number of shortest paths that passes through each segment of all shortest paths between all pairs of segments in a system. According to Hillier and his colleagues, this shows the “through movement” property of a space and explains how likely it is a pedestrian would pass through a segment on a trip or how probable (how many times) a segment would act like a bridge along the shortest path between all others (Hillier et al. 2012; Hillier 2009; Hillier and Iida 2005).

As time has gone by, space syntax research has developed new techniques which enable researchers to be more attentive to differentiations revealed in the structure and spatial configuration of different cities. Angular segment analysis (ASA) is one of the methods recently added in the segment analysis which has shown relatively better relationships to movement pattern than the same properties in axial line analysis (Hillier et al. 2012; Hillier and Iida 2005). Additionally, it has been demonstrated (Turner 2007; Dalton et al. 2003) that ASA properties are free of some axial line problems as highlighted by Ratti (2004). In the meanwhile, the choice value is also gaining popularity in the recent works of researchers. It shows a correlation to movement on par with, and in some cities better than, integration as a traditional measure of space syntax (Hillier and Iida 2005; Turner 2007).

Although the integration value for axial analysis originally was normalized, this initially was not done for the angular measures for methodological reasons as well as the way the measures were used; normalization becomes important primarily when the measures of different systems are compared which has not been the focus of most research using angular analysis. In order to avoid having large amount of values for choice which makes calculation and especially spatial comparison between cities in different sizes problematic, Hillier et al. (2012) suggest a method of normalization of angular choice and angular integration values. In this way not only the calculations are free of large spreads of values but there is also an opportunity to compare different cities with different sizes and even streets with cities. In addition, they show that the normalization in this way slightly improved the correlation between choice and the movement rates. Figures 1, 2 and 3 show a subsample of river-cities and non-river-cities and their extractions (river-cut, river-high-cut, and non-river-high-cut samples discussed in the next sections) illustrating NACH and NAIN properties.
Fig. 1

Illustration of NAIN value for a subsample of river-cities, river-cut and river-high-cut group. In all figures the color range varies from green to red representing the lowest and the highest values respectively

Fig. 2

Illustration of NAIN value for a subsample of non-river-cities and non-river-high-cut group. In all figures the color range varies from green to red representing the lowest and the highest values respectively

Fig. 3

Illustration of NACH value for a subsample of river-cities (first row) and non-river-cities (second row). We escape to show illustrations of the other groups due to non-visible changes between them and their origins in this scale. In all figures the color range varies from green to red representing the lowest and the highest values respectively

Presenting a way of comparison between cities based on normalized values, Hillier et al. establish a star model and try to compare and interpret the structures and spatial configurations of different cities in terms of structured and distributed cities or organic and planned cities. Before going through the characters of a star model, we concisely describe the properties of these two different types of spatial configurations of cities. A structured city is a city where there is a clear differentiation of the network role of different street segments; that is, where there are a smaller set of streets that are more central, and a high number of streets that rely on these central streets for their global connections. A distributed city, in comparison, is a city where the streets have a more similar network role, and where the dependency on specific streets for global connectivity is less in favor of a larger portion of the network operating as global connectors to an equal extent (Hillier et al. 2012).

Elaborating on the complexity of spatial configuration as spatial embedding of socio-cultural relations, a range of studies have shown its relevance and importance, but simultaneously the complexity with which configuration of space relates to social and cultural questions which necessitates inclusion of other factors in the analysis (Hillier and Hanson 1984; Markus Thomas 1993; Hillier 1996; Hanson 1998). As a specific example Karimi (1997) conducted space syntax research comparing six Iranian cities representing the Islamic tradition cities and six England cities representing the western medieval cities based on integration derived from axial lines. The results showed that although cities from the same category (genotype) tend to represent different spatial characters, they share common characters especially at local levels. He also shows that English cities are shallower and more connected than Iranian cities which were deeper and less connected. The integration core was denser and more compact in Iranian cities.

In another comparative study between a sample of 12 American and European cities, (Major 1997) using space syntax analysis tried to show the similarities and dissimilarities in the spatial structure between the two classes of cities based on axial line integration in local and global scale, depth and connectivity along with metric characters such as area and length of line and segment length. He found that American cities are well-structured and do not show typically a geometrical grid, but preserve their formal composition as they grow in size. European cities, similarly, also tend to preserve their spatial structure during the growing process, which comes from their core. Both thus show a tendency for preserving the primitive plan.

While there thus are tendencies on a generic level to be links between how structured and how planned a city is; this is not the concern of our research that focus on the question of structured versus distributed systems. The central methodological reference to large portions of this study is the research on comparison of the overall configurational properties of cities developed by Hillier and his colleagues (Hillier et al. 2012) called star model. The star model, as shown in Fig. 4, is based on four measures all extracted from normalized ASA properties placing one measure on each corner.
Fig. 4

A schematic of star model illustrated by Hillier et al. (2012). Mean choice and mean integration values are placed at the top and down of the model. Maximum choice and maximum integration values are set at the right and left side of the model respectively. Standardization is conducted on each set of measures varying from −3 at the center to +3 at the edge of star model, showing the minimum and maximum value for that measure respectively

Mean choice and mean integration values which meet the background properties of a city network are placed at the top and down of the model. Maximum choice and maximum integration values which match the foreground properties of a city network are set at the right and left side of the model respectively. To make these four measures comparable to each other, standardization is conducted on each set of measures varying from −3 at the center, showing the minimum value for that measure, to +3 at the edge of star model, showing the maximum value for that measure. Figure 5 shows the star models prepared for a subsample of cities illustrated in Figs. 13.
Fig. 5

Star models for the subsample of river-cities and non-river-cities and their extractions; first row blue color is for the original river-city (before cutting bridges or the highest choice values), red color is for river-cut sample, and orange is for river-high-cut sample. Second row blue color is for original non-river-city (before cutting the highest choice values) and red color is for non-river-high-cut sample

Integration measures are indices for accessibility and choice measures highlight the degree to which each system is structured or distributed. The higher value at the top and down of the model, the more distributed (geometrical) the city is and the higher value at the right and left side of the model, the more structured the city is (Hillier et al. 2012).

Coming back to our research specifically, we intend to clarify if and how the space syntax properties explore the effect of the presence of a river on the configurational form of cities and if and in addition by use of a control sample compare how rivers make the structure of cities different from non-river-cities.

In the next steps, data preparation for different analyses conducted on our sample is explained;

Bridges and cities with river banks

In geography a river bank is defined as the land beside the body of a river in which its slope can differ from vertical to horizontal (Leopold et al. 1964). Here we use this term in a wider range including the infrastructure of a city located in each river-bank connected to the opposite side of the river by crossing bridges. A bridge across a river is one of the most morphologically and a visually recognizable objects in river- cities. It plays a significant role not only from aesthetic point of view but also has a great effect on the connection and transportation between riverbanks. Bridges are currently the most common way, among all other ways like e.g. shipping or tunneling, to connect two river-banks together (Hölzer et al. 2010) and additionally convey economic and cultural meaning in cities as well (Chang and Choo 2009; Ioannidis 2011; Psarra 2012).

In this respect, giving a role to bridges to study the effect of rivers does not mean that we limit the effect of a river on the structure of a city to the bridges across it. We need to have in mind that bridges are only a part of the relation between city and river, and the bigger part might be the city itself. There are many cities settled in a side of rivers without any bridges while the river still shows its effect on development and expansion of the city. However, we here try to understand the effect of rivers on the spatial form of cities by using the role of the bridges in connecting the city, through correlation studies of systems with and without bridges.

An initial hypothesis here is that the cities are structured around the bridges, and that they constitute central, key structural parts of the city network. If this is true, removing the bridges should lead to dramatic changes of the configurative characters of the cities. Furthermore, river cities would be comparatively more sensitive to this kind of change than cities without rivers or similar geographical conditions that enforce key structural connectivity elements. To test this hypothesis, we simply cut all connecting bridges from the city and conduct a space syntax analysis on the remaining part of the city and extract the properties. We henceforth use the term of river-cut for this type of analysis. Figures 1 and 3 illustrates NACH and NAIN properties for a subsample of river-cut cities extracted from river-cities by cutting the bridge connections.

Cutting the highest value of choice measure

In order to understand the structure of river-cities, we compare the results to a control sample of non-river- cities in terms of structure and urban form consisting of 21 non-river-cities about the same size as the river-cities with a similar geographical as well as size distribution. In the first step, non-river-cities are analyzed in the same way as the river-cities based on angular properties, and then followed by the comparison to river-cities in their properties and their star models.

Working under the initial hypothesis of a central systemic role of bridges, there is a challenge to identify similar segments in non-river cities. Since we are looking after segments of central importance to structuring the city, this does not translate easily to ‘bottleneck segments’. We have chosen to consider the property best corresponding to the hypothesized role of bridges to be those included in what Hillier (2009) names the foreground network; i.e. the limited number of segments that tend to have a distinctly higher centrality than the other segments operating as connectors in the city’s configuration. It is worth mentioning that the number of segments, with the highest values of choice measurement, which were cut from each river-city, is the same as the number of segments which forms the bridges in that city. Also for the non-river-cities, it is the same as the number of segments cut in the river –cities with the same size. Our stress on the size and number of cutting segments is to establish comparability between the effect of bridges (in wider view rivers) and the effect of the highest values of choice (in a wider view foreground network). For reasons that will become clear, we also made a sample where the method of cutting segments was exactly the same in river cities as in non-river cities. In this respect we will have two types of results from a mathematically consistent method; river-high-cut (Figs. 1 and 3) and non-river-high-cut (Fig. 2) pointing out respectively to the river-cities and non-river-cities in which the segments including the highest values of choice measurement are cut from them.

At the end we need to note that the idea of cutting bridges or the highest choice values which to some extent evokes the foreground network is unique and not remarked in any similar research in space syntax. The only possible work which may show some similarities is Shpuza’s paper, (Shpuza 2013), on the effect of foreground network properties such as length and connectivity on the integration of street network in order to quantify this effect he defined the foreground network in four different thresholds including the highest 90, 80, 70, and 60 % values of choice. He showed that among them the mean length and the mean connectivity of the 80 % foreground network had the highest influence on the axial map integration. The method also has conceptual similarities to some of Hillier’s investigations in ‘the laws of the field’ (Hillier 1996b) where he identifies how within the confines of a grid, inserting walls have different distance and configurational effects depending on where in the field they are added—and consecutively where additional walls are added in relation to earlier additions. It also similarly relates to some investigations of office floor plate shapes and configuration by Shpuza (2006) that studies how interruptions in floor plan layouts by e.g. structural elements have different effects depending on both floor plate shape and organization of work places in offices.

Result and discussion

We intend to make the comparisons between our samples in two ways: a statistical part which includes correlation between different properties, and an anova test to verify if our data are significantly different from each other in their mean. Then in the main part, we figure out star models of cities for different groups of samples and analyze and interpret the different tendencies observed in different groups. We need to remark that due to a big sample of cities we are not going to thoroughly discuss the result for each unique city but the research is based on the general trend perceived in each group and comparison between them. Future research may further elucidate if and how differentiation within the samples is further related to other morphological properties of rivers or other geographical conditions.

Basic statistical analysis

As seen in Table 2, we have the correlation between the size of cities (NC) in a logarithmic scale and all normalized angular properties (NAP) including mean and maximum choice values and mean and maximum integration values for the both river-cities and non-river-cities. Based on the correlation values, there are no significant relationships between the size of cities and NAP suggesting that NAP is independent of size as Hillier et al. (2012) commented. It is worth nothing that the correlation values for the river cities are greater than the correlation for the non-river-cities in all cases and among them; maximum integration (and maximum choice show the higher correlation in both groups. Also based on Table 2 which shows the correlation between each pair of properties, the correlation between mean and maximum integration is the highest among all, as expected the mean is a well-dependable forecaster of the maximum in NAIN (Hillier et al. 2012).
Table 2

Pairwise correlations (R-value) between different properties for the river-cities and the non-river-cities in different situations

Correlation

River

River-cut

River-high-cut

Non-river

Non-river-high-cut

Mean and maxchoice

−0.324

−0.319

−0.316

−0.303

−0.005

Mean and max integration

0.964

0.788

0.98

0.978

0.984

Mean integration and mean choice

−0.132

0.008

−0.095

−0.064

0.213

Max choice and max integration

0.77

0.682

0.863

0.709

0.614

Mean choice and log NC

−0.321

−0.328

0.117

−0.402

−0.211

Max choice and log NC

0.429

0.255

−0.034

0.441

0.533

Mean integration and log NC

0.488

0.322

0.039

−0.024

0.071

Max integration and log NC

0.5

0.303

0.009

0.002

0.104

However, this value differs in different groups of cities as shown in Table 2; it is the highest for non-river- high-cut cities with a value of 0.968 and the lowest for the river-cut with 0.621. Bridges seem more related to the accessibility than the through movement(choice).The maximum changes among different groups of river-cities is for maximum integration in river-cut (0.631) while this change for mean integration is at the minimum level (0.135) for river-cut at the same time. This will be further explained later (Table 7) as it depends on additional methodological steps to be explained first. This might explain why the correlation between these two properties for the river-cut is the lowest among all especially when we can see that changes between maximum and mean integration for the other cases are very close to each other.

As the general patterns explored in almost all pairwise correlations, the pair of mean and maximum integration and the pair of maximum integration and maximum choice shows the highest correlations, conversely the correlations between the pair of mean choice and maximum choice and the pair of mean choice and mean integration are the lowest among all.

Doing further analysis is reasonable when we can verify that there is a significant difference among different groups of cities (five different groups) prepared for the analysis (Field 2007). Without performing the anova test we could not say for certain if rivers have any effect on the form of cities or not, since the anova test ensures that the samples are valid for comparative analysis. This literally means that space syntax analysis alone is not able to reveal the influence of morphology on cities. To do so, auteurs apply the anova test. Analysis of variance (ANOVA), developed by Fisher (1932), is the most accepted statistical methods used to analyze the differences among group means. The result of anova generally shows if the mean of samples in different groups are different. The null hypothesis which needs to be rejected, in this case, is that the samples of cities in different groups are derived from populations with the same mean values. To do so, the test of anova along with the test of homogeneity and the robust test of equality of means are conducted. Based on the test of homogeneity, shown in Table 3, which means that the population variances in each group are equal, the p value of Levene’s test for all variables are higher than the significance level (p < 0.05). This means that the null hypothesis assuming that variances of the populations from which different samples are drawn are equal is accepted and satisfied for doing the anova test. Table 4 shows the result of anova test for all properties in all different samples including river-cities, non-river-cities, and their extractions (five groups). As one can see there are significant effect of rivers on the form of cities as at the p < 0.01 level for the three conditions [F(5, 253) = 7.423, p = 0.000] for mean choice value, [F(5, 253) = 6.937, p = 0.000] for maximum choice value, [F(5, 253) = 6.986, p = 0.000] for mean integration value and [F(5, 253) = 12.174, p = 0.000] for maximum integration value. Furthermore based on Table 5 which shows robust tests of equality of means, p-value for all properties in different groups of samples are less than significance level of 0.01 and equal to 0.000. This again means that the null hypothesis of having equal means between groups is rejected and therefore the difference between means of groups is significant. This suggests that such differences in different groups are rooted in spatial and structural form of different groups. This is what we intend to scrutinize it in the next step by means of star models.
Table 3

Test of homogeneity of variances

 

Levene statistic

df1

df2

Sig.

Mean-choice

1.495

5

253

0.192

Max-choice

0.519

5

253

0.762

Mean-integration

1.792

5

253

0.115

Max-integration

0.814

5

253

0.541

Table 4

ANOVA test

 

Sum of squares

df

Mean square

F

Sig.

Mean-choice

 Between groups

0.030

5

0.006

7.423

0.000

 Within groups

0.205

253

0.001

  

 Total

0.235

258

   

 Max-choice

 Between groups

0.049

5

0.010

6.937

0.000

 Within groups

0.355

253

0.001

  

 Total

0.404

258

   

Mean-integration

 Between groups

0.378

5

0.076

6.986

0.000

 Within groups

2.737

253

0.011

  

 Total

3.115

258

   

Max-integration

 Between groups

1.893

5

0.379

12.174

0.000

 Within groups

7.868

253

0.031

  

 Total

9.761

258

   
Table 5

Robust tests of equality of means

 

Statistic a

df1

df2

Sig.

Mean-choice

 Welch

6.367

5

82.323

0.000

 Brown-Forsythe

7.116

5

141.480

0.000

Max-choice

 Welch

8.299

5

84.043

0.000

 Brown-Forsythe

7.193

5

189.387

0.000

Mean-integration

 Welch

5.737

5

82.329

0.000

 Brown-Forsythe

6.714

5

140.321

0.000

Max-integration

 Welch

12.167

5

83.104

0.000

 Brown-Forsythe

12.239

5

168.376

0.000

a Asymptotically F distributed

Star models

In this part, we draw out star models for each individual group in three levels including minimum, maximum, and mean which together replicate the general specifications of that group and try to compare the related results. Figure 6 shows the star model emerged from river-cities and Fig. 7 is the same extraction for non-river-cities. As one can see river-cities show a wider star on the horizontal range whereas the two models show similar distribution of values on the vertical range.
Fig. 6

Star model of 42 river-cities

Fig. 7

Star model of 21 non- river-cities

To reveal differences between these two groups the star models of river-cities and non-river-cities in average, minimum, and maximum level are shown in Fig. 8. Generally speaking, the average is indicating higher values for all measures in river-cities. According to Table 6 the maximum difference (0.243) is between averages of mean integration values for river-cities and non-river cities while averages of maximum choice are showing the minimum difference (0.042).
Fig. 8

Star models; (Blue) star models for river-cities at three different levels from minimum to mean and maximum. (Red) The same star models for non-river-cities

Table 6

Mean, Maximum and minimum values extracted from the star models of different groups

Value

Mean-choice

Max-choice

Mean-integration

Max-integration

 

Minimum

−2.473

−2.106

−1.735

−2.179

River-cities

Mean

−0.387

−0.066

0.035

−0.447

Maximum

1.743

2.059

2.120

1.685

Minimum

−2.919

−1.541

−2.099

−2.000

Non-river-cities

Mean

−0.571

−0.108

−0.208

−0.658

Maximum

1.743

1.424

1.705

0.973

Minimum

−3.000

−2.718

−3.000

−3.000

Non-river-cities-high-cut

Mean

−1.201

−0.959

−1.024

−1.519

Maximum

1.297

0.388

1.159

0.141

Minimum

−2.270

−1.329

−2.079

−1.266

River-cut

Mean

0.131

0.382

0.170

0.184

Maximum

2.432

3.000

2.019

3.000

Minimum

−2.878

−1.776

−2.707

−2.848

River-bank

Mean

0.204

−0.011

0.232

−0.299

Maximum

3.000

3.000

3.000

3.000

Minimum

−2.635

−3.000

−2.858

−2.989

River-high-cut

Mean

−0.552

−0.259

−0.506

−0.993

Maximum

1.784

2.059

1.745

1.435

Looking at Fig. 8 reveals that the minimum and maximum star models for these two groups show the different trends from the average star model. As seen in Table 6 mean choice values are equal (1.743) for both groups for the maximum star model and unlike the average star model maximum choice value in maximum star model shows a big difference (0.635) although the maximum difference stands for maximum integration value which is equal to 0.712. The situation in minimum star model is more surprising in which unlike the average star model maximum choice values show the maximum difference (0.565) and also unlike the maximum star model maximum integration values show the minimum difference (0.179).

Figure 9 show the average, minimum and maximum star models for the river-cut sample in comparison to the other samples including river-cities and non-river-cities. Based on Table 6 a dramatic increase in just about all measures is observed. This situation is more problematic when we see a substantial increase in the maximum choice values (−1.329, 0.382 and 3 respectively for minimum, average and maximum) and integration values (−1.266, 0.184 and 3 respectively for minimum, average and maximum) in comparison to river-cities and non-river-cities.
Fig. 9

Star models at three different levels; upper- left (minimum); upper- right (mean) and lower (maximum). River-cities (blue), river-cut- cities (green) and non-river cities (red)

This is part of the reason, aside ensuring comparability between samples, why we develop our research to investigate the effect of cutting the highest value of choice in river-cities and non-river-cities as well. Star models in different levels are shown in Fig. 10 for both river-high-cut and non-river-high-cut samples along with all other star models for a better comparison.
Fig. 10

Star models at three different levels; upper- left (minimum); upper- right (mean) and lower (Maximum); River-cities (blue), river-cut cities (green), non-river cities (red), river-high-cut cities (brown) and non-river-high-cut cities (orange). As seen in figure the model presented here for non-river-high-cut cities is in a line shape rather than a star shape due to showing the minimum values for some syntactical properties

At first look it can be seen that star models for both river-high-cut and non-river-cut show a dramatic decrease in almost all values (Table 6). The only exceptions are higher mean choice value (1.784) and equal maximum choice value (2.059) in maximum star model for river-high-cut in which compared to the same values for river-cities. Another trend is that river-high-cut contains higher values in almost all properties (the only exception is maximum choice value in minimum star models which is −3 for river-high-cut compared to −2.718 for non-river-high-cut).

This arguably means that the spatial configuration of river-cities is more distributed and more differentiated from the non-river-cities and at the same time it is more integrated as we can see in the model (we discuss it more in “Quantitative comparison of reaction to cutting the highest value and bridges” section.).

To go even deeper and test the emerging evidence and interpretations put forward above, we decided to apply the same analysis on each river-bank disjointedly and then compare the results to current outcomes. As seen in Table 6 the average values for the separated river-banks are bigger than the values for river-cities as their origins. However according to Table 6 the values measured for the maximum and minimum star models are more dispersed while the minimum star model seems to be smaller than minimum star model for non-river-cities, maximum star model is the biggest among all (the measured values for all the properties are equal to 3) and forms the climax of the models.

The reasonable answer for such pattern seems to be concealed in the physical shape and interaction between other characters of river cities, such as river dimension, size of each riverbank, landuse, and utility of each river-bank (Silva et al. 2006). As an example in Fig. 11, we show the effect of asymmetrical river-banks on the average star model for the river-banks of Zlin a river-city in Czech Republic. Here, the river crosses the city asymmetrically with a big river-bank in the south part (river-bank1 in Fig. 11), and a small one in the north part (river-bank2 in Fig. 11).
Fig. 11

A city (Zlin in Czech Republic) asymmetrically crossed by its river; forming a big river-bank in the south part (river-bank1 in figure), and a small one in the north part (river-bank2) in figure

The space syntax properties derived for both sides of the river-city and river-cut are brought for the comparison in Fig. 12. As we can see, the values for the south part of the city are the highest among all, showing a well-structured and well-distributed city, which is internally highly integrated. Inversely the north part is showing the lowest values among all suggesting an unstructured and internally segregated part of the city.
Fig. 12

Space syntax properties for a river- city (Zlin) in all different situations. As seen in diagram river-bank 2 shows the lowest values for almost all (except mean choice value) syntactical properties while river-bank 1 shows the highest values among all

In this case, one well-structured river-bank compensates not only the effect of a river in the city but also for the unstructured and segregated part of the city located in the other side of the river. On one side of the river, the city depends on its relation to the other side, whereas the other side of the river, the city acts largely independently from the other side—from the point of view of the structure of the street network. This implies the role of the city, itself, on the interaction with its river.

Quantitative comparison of reaction to cutting the highest value and bridges

To further develop on the usability of the star models in comparison between the river- cities and non-river cities in response to cutting the highest choice values, in Table 7 we calculate the absolute distance between values extracted from different star models of river-cities as original values to all other extraction of river-cities including river-cut, river-banks and river-high-cut. The same comparison is done for non-river-cities and their values are compared to Non-river-high-cut values. According to Table 7, it is seen that non-river-cities show a higher reaction to losing the highest choice value (foreground network) in comparison to the river-cities at the same situation as the total change for non-river-cities is equal to 3.158 while it is 1.732 and 1.445 for river-cut and river-high-cut cities respectively. Furthermore the property with the lowest change within groups and between groups is mean choice value in which its total change is 1.313. However, we need to mention that the impact of cutting bridges is additive but the impact of cutting the highest choice value is completely deductive. River-high-cut with a total change of 1.445 shows the minimum changes among all.
Table 7

Absolute distance between values extracted from different star models and their origins (river-cities and non-river-cities)

Distances to origin

Mean-choice

Max-choice

Mean-integration

Max-integration

SUM

River-cut

0.469

0.447

0.135

0.631

1.682

Non-river-high cut

0.629

0.850

0.816

0.861

3.157

River-high cut

0.214

0.193

0.541

0.546

1.494

SUM

1.313

1.491

1.492

2.038

 

Reflections on the statistical results

Before concluding, it seems in place to make some reflections on the results of these analyses as they have to an extent perplexed us. How come river cities seem more robust in their syntactic structure than non-river cities, when they clearly have a distinct set of connections that are crucial for their function? The results suggest that the foreground network is not directly dependent on how the geographic barrier in river-cities is crossed. One might speculate that this is because of the role and importance of the foreground network in transportation and movement, leading to that a river-city is so configured as to keep and preserve the best efficiency for the foreground network in reaching to and along the river as much as or more than to reach past it. However, the higher value of maximum integration means that generic accessibility in river-cities is higher than the non-river-cities as well. In this respect the average star model for the river-cut sample is asserting that the structure of river-cities is able to compensate the effect of disconnection and disintegration caused by the river running through it. From perspective of overall configurative properties, the bridges themselves are not structurally important. In relation to our initial hypothesis, this is a surprising result, since the initial hypothesis would suggest that the effect of removing bridges would significantly lower most of the values, since it would mean removing the hypothesized main focus of the connectivity of the city.

Cutting bridges from river-cities reveal that the river-cities are comparatively robust in their structures and less dependent on the crossing bridges. The latter is also supported by the big effect of cutting the highest choice values on the same sample of these cities and the bigger effect on the non-river-cities. Therefore we can assert that non-river-cities are more sensitive to changes in their spatial form.

While the results suggest there is something in the distinctly segregating effect of a river splitting a city into two parts that leads to some sort of compensation in the network as a whole, the reasons why this is and the processes leading to it remain to be established. It also needs to be further investigated what the effective separation of river banks mean, in relation to and regardless of how the syntactical properties are affected. Here, limitations to the approach and the overall methodology to understand the complex socio-spatial relationships of cities become apparent, and the benefits of expanding the analysis to include, for instance, agent-based modeling is, we argue, clear. Such analysis would allow to take better into account the effect on social relationships on individual and collective basis as structured in and through the city, and to what extent these are affected depending on a much wider range of factors, and from a more complete set of aspects. Arguably this becomes even more of interest as the analysis here show some results that were unexpected and raise questions regarding how to interpret and understand them that challenge some common explanations utilized in space syntax research. This would also further clarify the extent to what on syntactic level appears to remain similar or changed affect people’s ranges of possibilities in appropriating and making use of the city in other ways.

Conclusion

As a common element in the development of urban form and function throughout history, with significant effects on the physical structure of cities the aim of present research is to explore the relationship and influence of river systems on the physical form of cities. Using normalized properties of integration and choice values support us to realize some significant points of the interaction between rivers and cities as well as significant dissimilarity between river-cities and non-river-cities.

The first finding shows that although bridges are the most important way of connection between the two river-banks in river-cities, they are not as clearly important to manage the limits arisen from rivers in other respects. This was found by comparison between river cities and non-river cities, and how the networks in general, and ‘without’ bridges, are configured on a generic, averaged level. Analysis of the foreground and background network in a star model reveals that the river-cities are well-structured as if to compensate the discontinuity occurred by rivers, to the extent that as a general tendency, they are more integrated and better internally connected than non-river cities on each side of the river.

The second finding illustrates that the foreground network as identified through maximum choice values (c.f. Hillier 2009) in cities are more vital to the structure of cities than the bridges. This again emphasizes that the morphological adaption of a city to its river takes place in the structure of the city in general and not in the connection area across the river which includes the bridges. In other words, having the higher choice value in comparison to non-river-cities and showing more stability and control on the structure after cutting the highest value of choice shows that river-cities are more flexible in their structure than non-river-cities.

This suggests a kind of double-sided relation between rivers and cities on the overall urban morphological level. On the one hand, it is clear that rivers and bridges form important morphological elements, and they are clearly related to in various ways through history. Be it for transport, for views, for gathering places, or for other social, cultural, or functional reasons. On the other hand, it seems that the river cities tend to respond to the separation generated by a river by becoming more connected and integrated on a general network level. The social, cultural, functional, and other possible reasons for this remain to be investigated and it is premature to make any conclusions aside from these tendencies however tempting it is.

In this way, integrating space syntax with statistical and geospatial analysis can throw light on the way in which the properties of city networks and urban structure reflect the relative effect of rivers on the morphology of river cities. In this paper, some of these network properties have been investigated, contributing with a small piece to an understanding that needs to be taken further by other analytical approaches to contextualize and develop a broader understanding that at this stage remains inferred from a focused and limited study of the specific network aspect analyzed through syntax. The paper, finally, contributes through offering one piece of a better perception of the structure of river-cities that can support strategies, not only to improve river-cities interaction but to also enhance our knowledge on the constraints and limits to that interaction.

Declarations

Authors’ contributions

EA initiated the idea and conceived the study, performed experiments, analyzed data and wrote the paper. DK gave technical support and conceptual advice; helped interpret the result and edited the paper. Both authors read and approved the final manuscript.

Authors’ information

Ehsan Abshirini received the Master degree in remote sensing and GIS from Shahid Chamran University, Ahwaz, Iran in 2005. He is currently pursuing the PhD degree in Geoinformatics at the Royal Institute of Technology (KTH). He presently joined the Spatial Analysis and Design (SAD) research group at KTH School of Architecture that aims at the development of knowledge and theory on architectural space and form. His research interests are mainly in the area of GIS, Space Syntax and integration between them for the analysis of urban morphology. Daniel Koch is a researcher at the KTH School of Architecture, Vice Director for the research environment Architecture in the Making, and on the board for the national research school ResArc. He has contributed to space syntax research through a number of years, focusing largely on space syntax in relation to architecture theory, and is editor of The Journal of Space Syntax. In addition to papers, he has been the author of or contributor to a number of books, several of which focus on the relation between architecture and fashion. He is furthermore co-founder of and still active as a designer at Patch- work Architecture Laboratory operating primarily in Stockholm.

Competing interests

Both authors declare that they have no competing interests.

Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.

Authors’ Affiliations

(1)
Division of Geoinformatics, Department of Urban Planning and Built Environment, KTH (Royal Institute of Technology)
(2)
School of Architecture, KTH (Royal Institute of Technology)

References

  1. Asad R, Ahsan R (2012) Structuring urban sustainability with water: a case of kamrangir chor, Dhaka, Bangladesh. J Soc Dev Sci 3(8):293–303Google Scholar
  2. Batty M (2005) Cities and complexity: understanding cities with cellular automata, agentbased models, and fractals. MIT Press, CambridgeGoogle Scholar
  3. Batty M (2008) Fifty years of urban modelling: macro-statics to micro-dynamics. In: Albeverio S, Andrey D, Giordano P, Vancheri A (eds) The dynamics of complex urban systems: an interdisciplinary approach. Springer Physica-Verlag, New York, pp 1–20. doi:10.1007/978-3-7908-1937-3_1 View ArticleGoogle Scholar
  4. Batty M (2013) The new science of cities. Mit Press, Cambridge. doi:10.1111/gean.12042 Google Scholar
  5. Bavelas A (1950) Communication patterns in task-oriented groups. J Acoust Soc Am. doi:10.1121/1.1906679 Google Scholar
  6. Čakarić J (2010) Water phenomenon: urban morphology transformation. Facta Univ Ser Archit Civ Eng 8(4):375–388. doi:10.2298/FUACE1004375C View ArticleGoogle Scholar
  7. Cengiz B (2013) Urban river landscapes, advances in landscape architecture, book edited by Murat Özyavuz, InTech, pp 978–953
  8. Chang SP, Choo JF (2009) Values of bridge in the formation of cities. In: IABSE Symposium Report, International Association for Bridge and Structural Engineering 95(1):25–46 doi:10.2749/222137809796089142
  9. Crooks AT (2012) The Use of Agent-Based Modelling for Studying the Social and Physical Environment of Cities. In: De Roo G, Hiller J, Van Wezemael J (eds) Complexity and planning: systems, assemblages and simulations. Ashgate, Burlington, pp 385–408Google Scholar
  10. Crooks AT, Patel A, Wise S (2014) Multi-agent Systems for Urban Planning. In: Pinto NN, Tenedório J, Antunes AP, Roca J (eds) Technologies for urban and spatial planning: virtual cities and territories. IGI Global, Hershey, pp 29–56Google Scholar
  11. Dalton N, Peponis J, Dalton R (2003) To tame a TIGER one has to know its nature: extending weighted angular integration analysis to the description of GIS road-center line data for large scale urban analysis. Proceedings of the 4th International Space Syntax Symposium, London
  12. Ellis PA, Mackay R, Rivett MO (2007) Quantifying urban river–aquifer fluid exchange processes: a multi-scale problem. J Contam Hydrol 91(1):58–80. doi:10.1016/j.jconhyd.2006.08.014 View ArticleGoogle Scholar
  13. Ferber J (1999) Multi-agent systems: an introduction to distributed artificial intelligence. Addison-Wesley Longman, HarlowGoogle Scholar
  14. Field A (2007) Analysis of variance (ANOVA), in NJ Salkind (ed.) Encyclopedia of measurement and statistics, SAGE Publications, Inc., Thousand Oaks, CA, pp 33-36. doi:10.4135/9781412952644.n19
  15. Fisher R A (1932) Statistical methods for research workers, Oliver & Boyd, Edinburgh, 4th Edition
  16. Frank AU, Bittner S, and Raubal M (2001) ‘Spatial and cognitive simulation with multi-agent systems’. In spatial information theory—foundations of geographic information science (Int. Conference Cosit 2001, Morro Bay, USA, September 2001). (Montello DR, ed.), Lecture Notes in Computer Science, Vol. 2205, Berlin Heidelberg, Springer-Verlag, pp 124–139. doi:10.1007/3-540-45424-1_9
  17. Freeman LC (1977) A set of measures of centrality based on betweenness. Sociometry 40:35–41. doi:10.2307/3033543 View ArticleGoogle Scholar
  18. Geofabrik GmbH and openstreetmap contributors (2015). http://download.geofabrik.de. Accessed 22 Feb 2015
  19. Goodchild MF (2007) Citizens as sensors: the world of volunteered geography. GeoJournal 69(4):211–221. doi:10.1007/s10708-007-9111-y View ArticleGoogle Scholar
  20. Haklay M, Weber P (2008) Openstreetmap: user-generated street maps. Pervasive Comput IEEE 7(4):12–18. doi:10.1109/MPRV.2008.80 View ArticleGoogle Scholar
  21. Hanson J (1998) Decoding homes and houses. Cambridge University Press, CambridgeGoogle Scholar
  22. Hillier B (1996a) Space is the Machine. Cambridge University Press, CambridgeGoogle Scholar
  23. Hillier B (1996b) Space is the machine, part three: the laws of the field. Space is the machine: a configurational theory of architecture. Cambridge University Press, Cambridge, pp 273–368Google Scholar
  24. Hillier B (2009) Spatial sustainability in cities: Organic patterns and sustainable forms. In: Koch D, Marcus L and Steen J (ed) Proceedings of the 7th International Space Syntax Symposium, Royal Institute of Technology, Stockholm, p.K01.1–K01.20
  25. Hillier B, Hanson J (1984) The Social logic of space. Cambridge University Press, Cambridge. doi:10.1017/CBO9780511597237 View ArticleGoogle Scholar
  26. Hillier B, Iida S (2005) Network and psychological effects in urban movement. In: Cohn AG, Mark DM (eds) Spatial information theory. Springer, Berlin, Heidelberg, pp 475–490. doi:10.1007/11556114_30 View ArticleGoogle Scholar
  27. Hillier B, Yang T, Turner A (2012) Normalising least angle choice in depthmap-and how it opens up new perspectives on the global and local analysis of city space. J Space Syntax 3(2):155–193Google Scholar
  28. Hölzer C, Wiethüchter A, Räume M, Regionale (2010) Riverscapes: designing urban embankments. Birkhauser, BaselGoogle Scholar
  29. Hossain N, Bahauddin KM (2013) Integrated water resource management for mega city: a case study of Dhaka city, Bangladesh. J Water Land Dev 19(1):39–45. doi:10.2478/jwld-2013-0014 View ArticleGoogle Scholar
  30. Ioannidis K (2011) Designing the edge: an inquiry into the psychospatial nature of meaning in the architecture of the urban waterfront, Doctoral dissertation. KTH, StockholmGoogle Scholar
  31. Jawarneh RN, Julian JP, Lookingbill TR (2015) The influence of physiography on historical and future land development changes: a case study of central Arkansas (USA), 1857–2030. Landsc Urban Plan 143:76–89. doi:10.1016/j.landurbplan.2015.06.010 View ArticleGoogle Scholar
  32. Karimi K (1997) The spatial logic of organic cities in Iran and the United Kingdom. Proc. Space Syntax 1st Intl. Sym. Vol 1. Comparative Cities, London, pp 1–17
  33. Koch D (2013) The architectural interface inside-out. In: Kim YO, Park HT, Seo KW (eds) Proceedings to ninth international space syntax symposium. Sejong University Press, Seoul, pp 067:1–067:16Google Scholar
  34. Kubat AS (1999) The morphological history of Istanbul. Urban Morphol 3(1):28–41Google Scholar
  35. Kubat A S, Kaya H S, Sarı F, Güler G, Özer Ö (2007) The effects of proposed bridges on urban macroform of Istanbul: a syntactic evaluation. In Proceedings of 6th international space syntax symposium Istanbul, Turkey
  36. Leopold LB, Wolman MG, Miller JP (1964) Fluvial processes in geomorphology. W.H. Freeman and Co, San FranciscoGoogle Scholar
  37. Lynch K (1960) The Image of the city. MIT Press, CambridgeGoogle Scholar
  38. Major M D (1997) Are American cities different? Proc. Space Syntax 1st Intl. Sym. Vol 3, London, pp 1–14
  39. Markus Thomas A (1993) Buildings, power: freedom and control in the origin of modern building types. Routledge, LondonGoogle Scholar
  40. Mello S, Holanda F (2009) There is a City on the Riverside. In: Koch D, Marcus L and Steen J (ed) Proceedings of the 7th International Space Syntax Symposium. Royal Institute of Technology, Stockholm, Sweden,) 075(1-12
  41. Montello D (2001) Spatial cognition. In: Smelser NJ, Baltes PB (ed). International encyclopedia of the social and behaviorial sciences. Pergamon Press, Oxford, 14771–14775. doi:10.1016/b0-08-043076-7/02492-x
  42. Penn A, Hillier B, Banister D, Xu J (1998) Configurational modelling of urban movement networks. Environ Plan 25(1):59–84. doi:10.1068/b250059 View ArticleGoogle Scholar
  43. Peponis J (2012) Building layout as cognitive data: purview and perview interface. Cogn Crit 6:11–37Google Scholar
  44. Peponis J, Hadjinikolaov E, Livieratos C, Fatouros DA (1989) The spatial core of urban. Ekistics 56:43–55Google Scholar
  45. Portugali J (2011) Complexity, cognition and the city. Springer Science & Business Media
  46. Psarra S (2012) A Shapeless hospital, a floating theatre and an island with a hill: the invisible architecture of Venice. In: Greene M, Reyes J, Castro A (ed) Proceedings to the Eighth International Space Syntax Symposium, Santiago, 2012
  47. Ratti C (2004) Urban texture and space syntax: some inconsistencies. Environ Plan 31(4):487–499. doi:10.1068/b3019 View ArticleGoogle Scholar
  48. Shpuza E (2006) Floorplate shapes and office layouts: a model of the effect of floorplate shape on circulation integration, Doctoral dissertation. Georgia Institute of Technology, AtlantaGoogle Scholar
  49. Shpuza E (2011) A coupled configurational description of boundary shape based on distance and directionality artificial intelligence for engineering design. Anal Manufact 25:357–374. doi:10.1017/S0890060411000217 Google Scholar
  50. Shpuza E (2013) Foreground networks during urban evolution: the effect of length and connectivity on street network integration. In: Kim YO, Park HT, Seo KW (eds) Proceedings of the 9th International Space Syntax Symposium. Sejong University, SeoulGoogle Scholar
  51. Shpuza E (2014) Allometry in the syntax of street networks: evolution of Adriatic and Ionian coastal cities 1800–2010. Environ Plan 41(3):450–471. doi:10.1068/b39109 View ArticleGoogle Scholar
  52. Silva JB, Serdoura F, Pinto P (2006) Urban rivers as factors of urban (dis) integration. In 42nd IsoCaRP Congress, Istanbul, Turkey
  53. Turner A (2007) From axial to road-centre lines: a new representation for space syntax and a new model of route choice for transport network analysis. Environ Plan 34(3):539–555. doi:10.1068/b32067 View ArticleGoogle Scholar

Copyright

© Abshirini and Koch. 2016