Sage Journals: Discover world-class research

Abstract

Against the backdrop of rapidly expanding urban structures, land policies in many countries have been adapted to contain and redirect growth to existing urban structures. However, obstacles remain to measure the effects of policies. In the meantime, geoinformation technologies have given rise to a wide range of approaches to measure and describe urban form. Nevertheless, its application for the assessment of land policy has a high, but not yet fully exploited, potential. It is thus the aim of this research to address and investigate the options of spatial analysis and machine learning in particular to analyse urban form from a land policy perspective. To do so, we develop urban metrics informed by urban planning and land readjustment policies of two countries describing urban form on different spatial levels. We therefore formulate hypotheses on causal relations between policy and form. Based on the metrics, we apply the random forest algorithm to classify the building stock of the region. We then extract the residential areas, those with single-family houses, as this is where the effects of the policy are considered most visible. In a next step, we use random forest to predict the nationality of a building. Through variable importance measures, we identify and discuss urban morphological differences between the two countries and test the hypotheses on effects of land policies. We develop and test the approach for the French-German city-region of Strasbourg using OpenStreetMap data. We identify significant differences in the building coverage ratios, which tend to be higher in Germany. This can be linked to differences in planning regulations. Furthermore, German residential areas appear to be more diverse in urban form. Differences in land readjustment policies have proven to be plausible here, as French policies favour strong actors that develop residential areas more uniformly. In Germany, policies favour fragmented ownership-oriented development of residential areas. The metrics and the applied algorithm for building classification have proven to be robust in terms of data heterogeneity and have shown high levels of accuracy. They could also be successfully used for tracing causal relations.

Keywords

Comparative land policy random forest urban analytics urban form

Introduction

Urban development is facing substantial challenges, as climate and environmental urgencies need to be addressed. Subsequently, much research has been conducted on identifying determinants of urbanisation (Colsaet et al., 2018; Siedentop, 2018) and on developing policies that promote more sustainable modes of urbanisation, such as reduction of land take and densification (Béchet et al., 2017; Gerber et al., 2018). In France and Germany in particular, demands for sustainable development have led to dynamics in land policies, while challenges of adapting complex and intertwined institutional settings that regulate land use have become apparent (Delattre and Napoléone, 2016; Jehling et al., 2018; Kretschmer et al., 2015). Here, institutionalist approaches to analyse these settings are of increasing interest, as they set urban development in the context of public policies and private property rights (Gerber et al., 2018). These approaches could establish first links between the interplay of public and private actors and resulting urban morphologies (Buitelaar and Segeren, 2011; Harbers and Tennekes, 2016). The influence of specific institutional settings on urban morphology of urban form is shown by Harbers and Tennekes (2016) through a comparative qualitative study across national systems. They compare differences in urban form at the building level to draw conclusions about the effects of land policies. They further focus on single-family house residential areas, as they are considered to reflect planning regulations clearer than historic and mixed use urban areas (Tennekes et al., 2015). Also research in planning theory explores the link between rationalities in planning systems and their impact on urban form, i.e. urban design (Hartmann and Jehling, 2019).

Against this backdrop, comparative quantitative approaches appear to be promising in order to gain further insights into the links between urban form and national land policies. Addressing urban form at the building level, where public policies and private interests interact and then analysing the “emergent” urban structures (Barthélemy, 2016), could open up new ways for land policy assessment. However, while quantitative studies exist that address the effects of specific land policies on urban form at the building level (Jehling et al., 2018, 2020; Koomen et al., 2018), their application to national comparisons is challenging due to different underlying official data sets. In particular, due to the different geometric and semantic representation schemes, the buildings are often not comparable. With regard to the comparative analysis of urban form and policy, various strands of research explore causes and effects through modelling approaches (Cortinovis et al., 2019; Li et al., 2013; Verstegen et al., 2016). These studies are, however, most often based on land use information and are therewith above the granular level of buildings and thus the level of urban morphology.

To fill this gap, data-driven approaches that quantitatively express morphological characteristics in urban form through urban metrics at the building level (Vanderhaegen and Canters, 2017) become of interest here. Buildings are like streets, urban blocks and plots the main elements of urban form that structure cities (Oliveira, 2016) and have a long-standing role in analysing urban morphology (Conzen, 1960). Urban metrics serve as the basis for mapping urban forms and functions. On one hand, they allow us to discover new urban typologies, whose individual entities (types) are characterised by certain morphological features by applying clustering techniques (Berghauser Pont et al., 2019; Gil et al., 2012). On the other hand, urban metrics can be used to distinguish between specific urban types by means of supervised classification approaches (Barr et al., 2004; Meinel et al., 2009; Steiniger et al., 2008; Wurm et al., 2016). With the increasing availability of data, machine learning techniques are becoming more widely used to process large areas very effectively. In particular, the random forest algorithm (Breiman, 2001) has proven to be suitable for classifying building footprints by geometrical, topological or spectral features (Droin et al., 2020; Du et al., 2015; Hecht et al., 2019; Lu et al., 2014). It further allows us to measure the variable importance for data exploration and understanding (Strobl et al., 2007; Verikas et al., 2011). However, using classification approaches beyond national systems for comparative research remains challenging. Trained classifiers, even within a country, are only transferable to a limited extent due to regional differences in the building cultures and histories (Hecht et al., 2015; Steiniger et al., 2008). National differences in morphology require training data that covers all areas considered. In addition, differences in the building data used, e.g. geometric detail or minimum mapping unit, can have an impact on the results. Additional context information on the level of building groups (Du et al., 2016; Wu et al., 2018) could be helpful here to develop a robust comparative approach.

It is thus the aim of this research to make use of the capabilities of machine learning to analyse urban form at the building level across national boundaries and explore their feasibility in identifying the effects of land policies through comparison. As argued above, this research will focus on residential areas to reduce the complexity of influential factors. To further account for the high complexity in urban development and the challenge to cope with indeterminacy in social systems (Rohlfing, 2012), we develop a theoretical basis that allows us to draw on casual effects between land policy and the development of residential areas. Taking the example of the Strasbourg city-region, which spans across the French-German border, this research aims at answering the following questions: Which land policies are considered key in developing residential areas in France and Germany and what is their expected effect on urban form? Which national characteristics of urban form become visible by applying a data-driven classification approach? Does the comparison of characteristics allow for tracing back effects of national land policies?

Land policies for developing residential areas in France and Germany

From a land policy perspective, urban form emerges through the interplay of public and private interests (Buitelaar and Segeren, 2011). Policies and their application by public actors through (planning) instruments are constraining or enabling the use of private property. Consequently, actors’ interests and societal norms also influence how the instruments are used (Gerber et al., 2018). In terms of causal inference (Rohlfing, 2012), urban form is considered to be the result of this interplay of interests (Jehling et al., 2020). We consider two instruments to be of particular importance. These are public land use planning, which is considered a key driver of urbanisation (Colsaet et al., 2018; Kretschmer et al., 2015; Siedentop, 2009) and land management, which is the interface between public interests and private property. It is applied to adapt property rights (parcels of land) to planning by acquiring land or land readjustment (Hartmann and Spit, 2015).

Instruments of land use planning

In France, the Plan Local d’Urbanisme (PLU) is key for defining urban form (Herrmann, 2017). Since its introduction in 2000, it is applied on municipal and inter-municipal level to steer urbanisation. Municipalities that are part of a communautés urbaines or métropoles mutually develop a PLU. The PLU sets the legal basis for granting building permits (Geppert, 2014). Its regulation is directly binding for public and private property owners. In accordance with higher tier planning, the plan divides the municipal area into zones of different land uses, which are defined by the building code (Code de l’Urbanisme, 2020). For current and planned urbanization, it sets the land use (préciser l’affectation des sols) and determines the urban form through regulations such as building coverage or limitations for buildable land on a plot (e.g. l’implantation du bâti and coefficient d’emprise au sols) (Direction générale des Finances publiques and Direction générale des Collectivités locales, 2017.

In Germany, municipalities are required to deploy land use plans. Land use planning consists of a two-tier system, which was established in the 1960s (Söfker, 2018). A preliminary land use plan for the whole municipal area sets the current and planned land use (Flächennutzungsplan). It is also applied for inter-municipal planning in urban agglomerations. With a second-tier local land use plan (Bebauungsplan), municipalities set detailed regulations. This tier is then binding for land owners and relevant for building permissions (Jehling et al., 2020). The local land use plan defines the use and the building coverage (Art und Maß der baulichen Nutzung) and designated buildable land. This is done based on the national building regulation (Baunutzungsverordnung), which defines land use categories with respective limits for the building coverage ratio (Grundflächenzahl). Within this limit, municipalities are free to redefine the building coverage ratio.

In comparison, land use planning instruments are applied by municipalities, who decide on urban form through zoning and standards for building coverage. As we look at a functional interlinked cross-border region with comparable socio-economic standards, we hypothesise that urban form subsequently does not show significant differences between Germany and France.

Instruments of land management

In France, residential development is dominated by public or private developers (Oxley et al., 2009). The most common instruments are joined development zones (zones d’aménagement collective, ZAC), which is a specific form of public-private partnership for land management. The public actor himself or – by concession – a further actor conducts the development. This includes the relevant studies, the construction works, the acquiring and reselling of land, and the public infrastructure. The parcellation of land (lotification) is a central part of the process (Olei, 2020; Oxley et al., 2009). The urban form is thus largely seen as a result of a negotiation between municipality and developer (lotissuer) (Herrmann, 2017). Instruments of owner-based land readjustment exist as well (association fonciere urbaine) but only play a minor role (Kerger, 2017). Owner-based piecemeal and small-scale development happens to a limited extent in rural and suburban contexts (Guelton, 2018).

In Germany, land management through the formal instrument of land readjustment (Umlegung) is the most prominent (Hartmann and Spit, 2015). This instrument allows for co-operation between land owners and the municipality in adapting plots to a planned development. The key characteristic is that the procedure retains the existing ownership structure and that municipalities are not required to buy and resell land. All plots of land of an area are included in the process, rearranged and then redistributed after land is taken for public needs (Erbguth and Wagner, 2005; Markstein, 2004). Thus, residential areas are developed by a heterogeneous group of actors. In practice, municipalities and private actors also cooperate, using an instrument for public-private partnership (Städtebaulicher Vertrag) through which land is acquired, subdivided and resold. In some cases municipalities also opt to develop land themselves (Erbguth and Wagner, 2005; Hartmann and Spit, 2015).

In comparison, practices of land management show a clear difference. We hypothesise that due to the resulting more heterogeneous structure of property rights during the development, also the urban form of residential areas is more heterogeneous in Germany than in France.

Method and data

The method consists of three main steps (Figure 1). The first step is to define and compute urban metrics for quantifying morphological characteristics based on pre-processed input data. Second, we identify residential single-family housing (SFH) areas by classifying buildings using a supervised learning approach, particularly the random forest algorithm. Third, random forest is used as an analytical tool to analyse the building stock within residential areas and to compare French and German urban form.

Figure 1.

Workflow implementing the method (steps in grey scale).

Urban metrics

To discuss the assumed effects of land use policies and planning on urban form, i.e. “the physical components of urban space” (Fleischmann et al., 2020: 3), we require morphological information on different spatial and contextual levels. Vanderhaegen and Canters (2017) describe patch-based, profile-based and building-based metrics for built-up areas. While the patch-based metrics are important for describing the urban context, their comparative application across national systems is often hampered by different mapping standards. To be able to represent the urban context, grouping of adjacent buildings through graph-based algorithms has proven to be helpful (Zhang et al., 2012). Based on this proximity relationship between buildings, a minimal spanning tree (MST) is generated that spatially groups neighbouring buildings (Caruso et al., 2017; Cetinkaya et al., 2015; Wu et al., 2018). Following Caruso et al. (2017), the centroids are retrieved from building polygons and then used to generate the MST. Constraints for generating building groups are streets so that each entity represents an urban block. Furthermore, a cut-off distance of 70 m between buildings is set. Finally, geometrical and topological properties are derived for different spatial levels (Figure 2).

Figure 2.

Urban form metrics on building, neighbourhood and group levels. The example highlights a detached house (building A) and a terraced house (building B). Neighbourhood is defined by first order direct neighbourhood of detached (A) or adjacent buildings (B) within a MST. All buildings within a MST form a building group. (source: authors' work; data © OSM contributors)

Inspired by further approaches (Hecht et al., 2015; Steiniger et al., 2008; Vanderhaegen and Canters, 2017), a number of characteristics are defined, which can relate to different spatial levels, particularly individual buildings, the (immediate) neighbourhood and the building groups (see Figure 2). The resulting urban metrics are listed in Table 1, which also shows how they express the land policy perspective (chapter 2). The variables for building style on neighbourhood level and the metrics on building group level (range and medium values) are considered helpful to cover structural heterogeneity.

Table 1.

Urban metrics as input for identifying single-family housing (SFH) areas and analysing national characteristics of urban form.

Name	Planning and policy relevance	Description	Spatial level	Input for identifying SFH areas	Input for analysing national characteristics
elongation	Building form	Ratio of length to width of minimal bounding rectangle of building polygon	building	x
perimeter	Building form	Length of the outer edge of the building polygon in meters	building	x
footprint	Building form	Building footprint size in sqm	building	x
build_coverage	Building coverage ratio	Share of all area covered by buildings in a defined 40 m buffer zone around the building centroid	neighbourhood	x
distance	Distance regulation for construction	Mean distance of building centroid to next centroids in 1st order neighbourhood in the MST in meters	neighbourhood	x
separation	Distance regulation for construction	Approximation for the distance separating two buildings based on ratio of width of bounding rectangle and distance calculated for 1st order neighbourhood in the MST in meters	neighbourhood	x
adjacency	Building style	Number of buildings whose polygons share border(s) with a building	neighbourhood	x
ratio_n1	Building style	Ratio between building footprint size of a building and that of a first adjacent building	neighbourhood	x
ratio_n2	Building style	Ratio between building footprint size of a building and that of a second adjacent building	neighbourhood	x
semi_d	Building style	Dummy for having one adjacent building with ratio_n1 (0.8-1.2) to rule out auxiliary buildings	neighbourhood	x
terraced	Building style	Dummy for having two adjacent buildings with ratio_n1 and ratio_n2 (0.8-1.2) to rule out auxiliary buildings	neighbourhood	x
m_[building or neighbourhood variable](median of values)	Overall planning regulation	Median of values for all buildings and neighbourhood metrics (except ratio and dummy variables)	building group	x	x
r_[building or neighbourhood variable](range of values)	Structural heterogeneity	Range of values for all buildings and neighbourhood metrics (except ratio and dummy variables)	building group	x	x
evenness	Structural heterogeneity	After classification: Shannon's Evenness Index, defined by the Shannon diversity index divided by the maximum diversity (building types per group)	building group		x

Identifying single-family housing areas

The aim of this step is to derive residential areas with single-family housing (SFH areas). The starting point for this is the classification of the building footprints that allows for identifying residential buildings, i.e. single-family houses. To realize this, we apply a supervised classification approach, particularly the random forest (RF) algorithm. RF is an ensemble classiﬁer developed by Breiman (2001). In the context of building footprint classification, RF outperforms other machine learning classifiers in terms of the generalization ability and computational efﬁciency (see Hecht, 2014). Based on a given training sample, a large of number of individual trees (so-called random forest) are constructed by a random bootstrap selection of the training data. We use the RF implementation in KNIME Analytics Platform version 3.5.2 (Berthold et al., 2008; Fillbrunn et al., 2017), which uses a C4.5 algorithm to generate decision trees capable of using numerical and nominal attributes (Wu et al., 2008).

Supervised building classification requires a training data set of buildings with known type information. In order to collect this training data for the French–German region, a set of building groups is selected across the region (Figure S1, supplementary material). According to a defined building typology (Table S2, supplementary material), we assign each sample building polygon to a type by visual interpretation of aerial photographs and street views. The used typology differentiates between detached, semi-detached and terraced houses and further urban structures (Hecht et al., 2015; Jehling et al., 2018). In the process, the typology is adapted to encompass the variety of buildings in both countries.

In a next step, a random forest model is trained and then applied to the total building stock (number of trees is 500, split criterion is information gain ratio (Wu et al., 2008)). Metrics on all spatial levels are applied as variables (refer to Table 1). A validation is carried out through 10-fold spatial block cross-validation (Valavi et al., 2019) to minimize spatial autocorrelation (Brenning, 2012), while ensuring that the folds contain all types of buildings. After validation, the regional building stock is then classified by applying a RF model using the entire training data set. All building groups (see ‘Urban metrics’ section) with a predominant share of single-family houses, which we set as a share of over 90% by number, are then considered to represent residential areas, i.e. single-family housing areas (SFH areas).

Analysing national characteristics in urban form

To analyse the characteristics of urban form of the obtained SFH areas, their structural heterogeneity is of interest. We add Shannon's Evenness Index to the set of urban metrics (Table 1). The index has its origin in ecology but is often used for spatial pattern analysis for quantifying landscape structures (McGarigal and Marks, 1995; Martellozzo and Clarke, 2011). It expresses the observed level of diversity. Maximum diversity is achieved when all types are evenly distributed.

Based on the extended set of metrics, we again use the capabilities of the RF algorithm. We define the nationality as a binary classification problem with target classes representing nationality and train a RF model using urban metrics on group level (Table 1) to predict whether a building is German or French. Through the variable importance measure provided by the RF algorithm itself, we identify those variables that are most relevant for explaining differences between the German and French part of the region. We use a frequency-based measure (number of times a variable is used to split a tree) (Strobl et al., 2007). For the validation of the RF model, we perform a 10-fold spatial cross-validation using blocks. We use Welch’s t-test to determine whether the obtained mean values of the variables differ significantly between the two countries.

For validation, the classification results are spatially visualised to assess the model with regard to spatial patterns or variation of accuracy. Based on the causal relations assumed in ‘Land policies for developing residential areas in France and Germany’ section, we then discuss effects between land policy and urban form. To do so, we interpret the variables according to their importance.

Study area, data and pre-processing

We test the approach for the French–German city-region of Strasbourg. The cross-border region offers the opportunity to analyse urban form across national boundaries, while limiting topographic influences due to its location in the plains of the upper Rhine valley. To cover a wide range of urban densities and patterns within a functional region, a delimitation of 40 minutes car travel time isochrones around the city centre of Strasbourg is applied.

The data for the regional building stock and the street network (Table S1) are retrieved using the OpenStreetMap Overpass API (Padgham et al., 2017). While the data have the advantage of being available and freely accessible beyond national borders, data quality aspects need to be considered. Studies have shown that the data quality of OSM features such as streets (Barrington-Leigh and Millard-Ball, 2017) and buildings (Fan et al., 2014) is very good for cities but may be different in rural areas in terms of completeness (Hecht et al., 2013). Since we are investigating a large region with urban and rural areas, some geometrical details in OSM may be missing in some areas. It is therefore particularly important to take the aspect of heterogeneous modelling of the buildings into account by applying a pre-processing step. For this purpose, minor individual structures of ≤20 sqm, such as small building annexes, are eliminated. Separately represented building parts of ≤40 sqm are then merged with the polygon of the main building.

Results

In the following, we present the results of the classification of the regional building stock and the identification of single-family housing (SFH) areas as well as the results of the comparative analysis of residential urban form in France and Germany.

Identification of the single-family housing areas

Building footprint classification

For describing the classification accuracy, we use the confusion matrix with producer’s and user’s accuracy, as well as total overall accuracy and Cohen’s kappa coefficient. Table 2 shows high accuracies for the classes: Rural houses (33) and Detached SFH (31) and only minor classification errors for the classes Semi-detached SFH (32), Terraced SFH (34) and Linked SFH (36). The overall accuracy of 0.894 with a standard deviation (σ) of 0.016 and a Cohen’s Kappa of 0.843 (σ = 0.027) shows the high quality of the model. The class-specific assessment also shows high values for producer’s and user’s accuracy, keeping in mind that we reduce spatial autocorrelation through block cross-validation. We conclude that this classification model for identifying residential buildings is suitable as a basis for further comparative analysis.

Table 2.

Confusion matrix and accuracy measures of building classification.

		Reference type
		10	31	32	33	34	36	total	User’s accuracy
Predicted type	Other urban (10)	4039	456	13	24	21	8	4561	0.886
	Detached SFH (31)	256	5624	28	32	61	8	6009	0.936
	Semi-detached SFH (32)	100	29	381	0	13	4	527	0.723
	Rural houses (33)	128	103	5	2009	2	0	2247	0.894
	Terraced SFH (34)	24	83	9	3	246	2	367	0.670
	Linked SFH (36)	21	14	15	0	5	138	193	0.715
	total	4568	6309	451	2068	348	160	13904
	Producer’s accuracy	0.884	0.891	0.845	0.971	0.707	0.863
						Overall accuracy (σ)			0.894 (0.016)
						Cohen’s Kappa (σ)			0.843 (0.027)

Single-family housing areas for cross-border analysis

The classified building stock is used to determine the residential areas (SFH areas) by analysing the class composition for each building group (see ‘Identifying single-family housing areas’ section). Figure S2 gives examples of these derived SFH areas in France and Germany. Statistics for the residential building stock and residential SFH areas for the entire region (Table S3) reveal that the proportions do not show large differences. Detached SFH (31) are the most frequently occurring class in both France (41%) and Germany (45%), followed by other urban buildings (10) and rural housing (33). While Germany has a higher proportion of semi-detached SFH (32) with a share of 3% compared to France with 1%, terraced SFH are more common in France.

French and German characteristics

The RF model predicting the national context of the SFH areas has the following results. The overall accuracy of 0.837 can be considered good, while a Cohen’s kappa coefficient of 0.583 shows a moderate agreement. Figure 3 gives insights into the spatial variation of the accuracy of the model, indicating if a building was predicted correctly as German or French. The block-based mapping of overall accuracy shows differences between the countries France and Germany suggesting that German buildings tend to be more often predicted as French buildings than vice versa (see confusion matrix, Table S4). Differences within countries are less clearly noticeable, but overall accuracy tends to be higher in rural areas with smaller residential areas (e.g. in the north-west) and lower in urban areas (centre). Steeper terrain in the German south-east of the region probably affects the accuracy of the model.

Figure 3.

Overall accuracy of predicted national context of buildings within residential SFH areas in the Strasbourg region (source: authors’ work; data © OSM contributors).

Figure 4 gives the distribution of the variables sorted by their importance for the classification by the RF model (see Table S5). Only variables on building group level (m_ and r_ suffixes as well as the evenness index) are considered (Table S4 for the ranking of all variables). The medium building footprint size (m_footprint) and medium building coverage ratio (m_build_coverage) are the most important for the model. The distribution of the two variables shows a higher interquartile range and median for the residential single-family houses in Germany. The third most important variable is r_separation, which gives the range of how far buildings are separated within a building group. The range of the building footprint size within the building groups (r_footprint) is higher in Germany. Evenness, i.e. diversity in building types within groups, is of a low importance. While the median values are close to each other here, German building groups show a higher interquartile range compared to the French ones. Applying Welch's t-tests shows that the distributions of the standardized scores of all variables differ significantly (each with p < 0.01).

Figure 4.

Distribution of standardized scores of variables that explain national differences, ranked by importance (from low to high) (source: authors’ work).

Discussion

Tracing effects of land policies

The analytical results and tests showed that the urban form of SFH areas in the German and French parts reveal the national context. However, the moderate values of the validation results also made clear that other effects on urban form need to be taken into account. Visualization of accuracies showed that location within the region could have an impact. Here, additional metrics describing the spatial location of the SFH areas within the region could be helpful to control for rural and urban differences in the building structure (Jehling et al., 2018), such as different densities.

Based on assumed relations between national land policies and urban form, the results allow for a further interpretation. With regard to the effects of land use planning, we hypothesised that a similar institutional setting and regulatory framework for zoning lead to similarities in urban form. However, differences in urban form become apparent. The mean building footprint size (m_footprint) and the mean building coverage ratio (m_build_coverage) of single-family houses are substantially higher in Germany. The larger interquartile ranges of the two measures for Germany also suggest a higher variability between SFH areas in Germany compared to France. German building regulation, thus, seems to give municipalities more range in decisions on urban form, while also requiring higher densities. French municipalities, in contrast, appear to apply stricter regulations, but also plan for lower urban densities.

To further scrutinize these differences, more knowledge on the agency of municipalities in applying the instrument is required. A possible path for further analysis could be the role of regional planning in Germany for the high variability that requires municipalities to plan in accordance with assigned functional roles. In France, the role of inter-municipal competition in the context of a private developer-driven residential market (Herrmann, 2017) asks for further research on the lower densities.

For the effects of the instrument of land management, we hypothesised that according to differences in the instruments for land management and the subsequent more heterogeneous structure of actors, the resulting urban form within SFH areas is more heterogeneous in Germany than in France. The range of distances between buildings (r_separation) indicates that the inner structure of French residential areas is slightly more homogeneous (lower median) than German residential areas. In terms of the range, building footprint sizes and building coverage, French residential areas are clearly more homogeneous. The diversity of building types (evenness) played only a minor role, although German residential areas are characterised by a higher diversity. This suggests that practices of land management in France and Germany (Guelton, 2018; Hartmann and Spit, 2015) lead to different outcomes in urban form. However, more empirical knowledge on the application of these practices is required to better sustain this argument and establish causal relations.

Reflections on the approach

We made use of OpenStreetMap data, as it is a highly detailed and cross-border data set beyond country-specific mapping conventions of national mapping and cadastral agencies. Moreover, these data are available free of charge worldwide. The data may have limitations, e.g. incompleteness of buildings in rural areas (Hecht et al., 2013), but recent studies show that data accuracy has improved over time (Brovelli and Zamboni, 2018). It is the nature of OpenStreetMap that no data model is predefined and no strict mapping rules (only recommendations) are given by the community. Accordingly, buildings should be mapped as individual buildings, the outline of building blocks or other complex arrangements of properties. However, this freedom means that the building representation is not the same everywhere, depending on which form of modelling the contributor prefers, and can even vary within a city or region (Hecht et al., 2013). While we considered this in the selection of metrics, the sample data for this particular cross-border region showed no substantial inconsistencies in building representation.

Regarding the use of a data-driven approach for building classification (Hecht et al., 2019; Jehling et al., 2018), we could demonstrate its successful application to enrich comparative research on land use policies. We could show that it proved to be valuable for discussing assumed effects of land policies on urban form (Tennekes et al., 2015; Gerber et al., 2018). The applied supervised machine learning approach requires a sufficiently large amount of training data; however, these data can be collected with little manual effort. The application for identifying German and French characteristics allowed us to identify the most relevant variables. The results encourage us to further test the approach in other contexts for which no explicit boundaries are given, such as for distinguishing between areas of similar planning cultures or topographically caused influences on urban form.

After testing the approach for a cross-border region, the options for transferability of our approach to cover national levels, such as France and Germany as a whole, should be seen with caution. While the focus on residential areas is meant to limit influencing factors beyond land policies, the variety of urban form and policies within a country and across different landscapes might be still too high. Transferring the approach to other cross-border regions in Europe appears more promising, as the chosen input data guarantees good applicability.

With our approach, we are able to confirm hypotheses about a specific urban form, which we have derived from a priori knowledge on available instruments of land policies. However, it is not a proof of causality, since we only consider the outcome (urban form) and not the processes of the policy instruments in a model, as for example Verstegen et al. (2016) show. Taken the complexity of the application of instruments and the challenges to infer causal relations, we argue that further research should address actors’ agency and interests empirically (Jehling et al., 2019). This should also be sustained by a more elaborated theoretical framework (Rohlfing, 2012).

Conclusion and outlook

We presented a data-driven approach to identify residential areas as a basis for analysing effects of national land policies on urban form and tested it for the French–German city-region of Strasbourg. The approach enables a classification of a cross-border building stock using OpenStreetMap data according to a predefined typology, which serves as a consistent basis for comparative analysis. We take a land policy perspective to establish a theoretical basis that allows us to develop hypotheses on the causal links between policy and urban form. We identified instruments of land use planning and land management as being key for developing residential areas. Due to their characteristics, we expected that land use planning does not lead to significant differences between residential areas in Germany and France. Land management, however, should lead to residential form that is more heterogeneous in Germany than in France. To test these links, we analysed urban form characteristics using machine learning (random forest) for classification. First, we were able to satisfactorily identify residential single-family housing areas across national borders. Second, we identified significant differences in urban form between the two countries. The mean building footprint size as well as the mean building coverage ratio showed the highest importance in explaining these. A comparison of the statistical distributions of the variables thus contradicts our hypothesis on land use planning and supports the hypothesis on land management, that residential form is more heterogeneous in Germany compared to France. While this allowed for developing a consistent argumentation from land policy to urban form, the results also revealed the need to further combine quantitative urban analytics and qualitative empirical research on instruments to establish more robust causal relations.

Supplemental Material

sj-pdf-1-epb-10.1177_2399808321995818 - Supplemental material for Do land policies make a difference? A data-driven approach to trace effects on urban form in France and Germany

Supplemental material, sj-pdf-1-epb-10.1177_2399808321995818 for Do land policies make a difference? A data-driven approach to trace effects on urban form in France and Germany by Mathias Jehling and Robert Hecht in EPB: Urban Analytics and City Science

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

Mathias Jehling

Supplemental material

Supplemental material for this article is available online.

Mathias Jehling is senior researcher at IOER, focusing on geographic information in the planning context. He works on urban densification and institutionalist approaches to planning and land policies.

Robert Hecht is senior researcher at IOER. His research focuses on applied spatial data analysis, pattern recognition, machine learning, and geovisualization in the context of settlement structure analysis.

References

Barr

Barnsley

Steel

(2004) On the separability of urban land-use categories in fine spatial scale land-cover data using structural pattern recognition. Environment and Planning B: Planning and Design 31(3): 397–418.

Barrington-Leigh

Millard-Ball

(2017) The world’s user-generated road map is more than 80% complete ( Ali

(ed.). Plos One 12(8): e0180698.

Barthélemy

(2016) The Structure and Dynamics of Cities: Urban Data Analysis and Theoretical Modeling. New York: Cambridge University Press.

Béchet

Le Bissonnais

Ruas

(2017) Sols artificialisés et processus d’artificialisation des sols, déterminants, impacts et leviers d’action. INRA DEPE. Available at: http://temis.documentation.developpement-durable.gouv.fr/docs/Temis/0087/Temis-0087278/23737_1.pdf (accessed 10 February 2021).

Berghauser Pont

Stavroulaki

Marcus

(2019) Development of urban types based on network centrality, built density and their impact on pedestrian movement. Environment and Planning B: Urban Analytics and City Science 46(8): 1549–1564.

Berthold

Cebron

Dill

, et al. (2008) KNIME: The Konstanz Information Miner. In: Preisach

Burkhardt

Schmidt-Thieme

, et al. (eds) Data Analysis, Machine Learning and Applications Berlin, Heidelberg: Springer Berlin Heidelberg, 2008, pp.319–326.

Breiman

(2001) Random forests. Machine Learning 45(1): 5–32.

Brenning

(2012) Spatial Cross-Validation and Bootstrap for the Assessment of Prediction Rules in Remote Sensing: The R Package Sperrorest. 2012 IEEE International Geoscience and Remote Sensing Symposium, Munich, Germany, pp. 5372–5375.

Brovelli

Zamboni

(2018) A new method for the assessment of spatial accuracy and completeness of OpenStreetMap building footprints. ISPRS International Journal of Geo-Information 7(8): 289.

10.

Buitelaar

Segeren

(2011) Urban structures and land. The morphological effects of dealing with property rights. Housing Studies 26(5): 661–679.

11.

Caruso

Hilal

Thomas

(2017) Measuring urban forms from inter-building distances: Combining MST graphs with a Local Index of Spatial Association. Landscape and Urban Planning 163: 80–89.

12.

Cetinkaya

Basaraner

Burghardt

(2015) Proximity-based grouping of buildings in urban blocks: a comparison of four algorithms. Geocarto International 30(6): 618–632.

13.

Code de l’Urbanisme (2020) Available at: www.legifrance.gouv.fr/affichCode.do;jsessionid=A73B5F80F9B35657A8EFF335E4EFC149.tplgfr41s_1?cidTexte=LEGITEXT000006074075&dateTexte=20200331 (accessed 10 February 2021).

14.

Colsaet

Laurans

Levrel

(2018) What drives land take and urban land expansion? A systematic review. Land Use Policy 79: 339–349.

15.

Conzen

MRG

(1960) Alnwick, Northumberland: A Study in Town-Plan Analysis. London: George Philip&Son.

16.

Cortinovis

Haase

Zanon

, et al. (2019) Is urban spatial development on the right track? Comparing strategies and trends in the European Union. Landscape and Urban Planning 181: 22–37.

17.

Delattre

Napoléone

(2016) Écologiser les documents d’urbanisme pour protéger les terres agricoles et les espaces naturels. Courrier de l’environnement de l’INRA 60(mai): 63–71.

18.

Direction générale des Finances publiques and Direction générale des Collectivités locales (2017) Les documents d’urbanisme et les règles générales d’urbanisme. Available at: www.collectivites-locales.gouv.fr/documents-durbanisme-et-regles-generales-durbanisme#__RefHeading__112_1390438360 (accessed 19 November 2019).

19.

Droin

Wurm

Sulzer

(2020) Semantic labelling of building types. A comparison of two approaches using Random Forest and Deep Learning. In: Publikationen der DGPF, Stuttgart, 2020. Available at: www.dgpf.de/src/tagung/jt2020/proceedings/proceedings/papers/76_KKNP_DGPF2020_Droin_et_al.pdf (accessed 30 April 2020).

20.

Luo

Cao

, et al. (2016) Extracting building patterns with multilevel graph partition and building grouping. ISPRS Journal of Photogrammetry and Remote Sensing 122: 81–96.

21.

Zhang

(2015) Semantic classification of urban buildings combining VHR image and GIS data: An improved random forest approach. ISPRS Journal of Photogrammetry and Remote Sensing 105: 107–119.

22.

Erbguth

Wagner

(2005) Grundzüge Des Öffentlichen Baurechts. Vol. 4. Aufl. München: Beck.

23.

Fan

Zipf

, et al. (2014) Quality assessment for building footprints data on OpenStreetMap. International Journal of Geographical Information Science 28(4): 700–719.

24.

Fillbrunn

Dietz

Pfeuffer

, et al. (2017) KNIME for reproducible cross-domain analysis of life science data. Journal of Biotechnology 261: 149–156.

25.

Fleischmann

Romice

Porta

(2020) Measuring urban form: Overcoming terminological inconsistencies for a quantitative and comprehensive morphologic analysis of cities. Environment and Planning B: Urban Analytics and City Science 239980832091044.

26.

Geppert

(2014) France, drifting away from the ‘regional economic’ approach. In: Reimer

Getimēs

Blotevogel

, et al. (eds) Spatial Planning Systems and Practices in Europe: A Comparative Perspective on Continuity and Changes. 1st ed. London: Routledge, pp.109–126.

27.

Gerber

J-D

Hartmann

Hengstermann

(2018) Instruments of Land Policy: Dealing with Scarcity of Land. Urban planning and environment. London New York: Routledge, Taylor & Francis Group.

28.

Gerber

J-D

Hengstermann

Viallon

F-X

(2018) Land policy: How to deal with scarcity of land. In: Gerber

J-D

Hartmann

Hengstermann

(eds) Instruments of Land Policy – Dealing with Scarcity of Land. Urban planning and environment. London New York: Routledge, Taylor & Francis Group, pp.8–26.

29.

Gil

Beirão

Montenegro

, et al. (2012) On the discovery of urban typologies: Data mining the many dimensions of urban form. Urban Morphology 16: 27–40.

30.

Guelton

(2018) The end of the French model of land development? Town Planning Review 89(6): 553–556.

31.

Harbers

Tennekes

(2016) The role of institutions in the morphology of housing areas. Proceedings of the Institution of Civil Engineers – Urban Design and Planning 169(5): 230–243.

32.

Hartmann

Jehling

(2019) From diversity to justice – Unraveling pluralistic rationalities in urban design. Cities 91: 58–63.

33.

Hartmann

Spit

(2015) Dilemmas of involvement in land management – Comparing an active (Dutch) and a passive (German) approach. Land Use Policy 42: 729–737.

34.

Hecht

(2014) Automatische Klassifizierung von Gebäudegrundrissen: ein Beitrag zur kleinräumigen Beschreibung der Siedlungsstruktur. Germany: Rhombos-Verl.

35.

Hecht

Kunze

Hahmann

(2013) Measuring completeness of building footprints in OpenStreetMap over Space and Time. ISPRS International Journal of Geo-Information 2(4): 1066–1091.

36.

Hecht

Meinel

Buchroithner

(2015) Automatic identification of building types based on topographic databases – A comparison of different data sources. International Journal of Cartography 1(1): 18–31.

37.

Hecht

Herold

Behnisch

, et al. (2019) Mapping long-term dynamics of population and wellings based on a multi-temporal analysis of urban morphologies. ISPRS International Journal of Geo-Information 8(1): 2.

38.

Herrmann

(2017) Fabriquer la ville avec les lotissements: Une qualification possible de la production ordinaire des espaces urbains contemporains? Lyon, Lausanne: Université Lumière Lyon 2, Université de Lausanne.

39.

Jehling

Hecht

Herold

(2018) Assessing urban containment policies within a suburban context – An approach to enable a regional perspective. Land Use Policy 77: 846–858.

40.

Jehling

Hitzeroth

Brueckner

(2019) Applying institutional theory to the analysis of energy transitions: From local agency to multi-scale configurations in Australia and Germany. Energy Research & Social Science 53: 110–120.

41.

Jehling

Schorcht

Hartmann

(2020) Densification in suburban Germany – Approaching spatial processes through concepts of justice. Town Planning Review 91(3): 217–237.

42.

Kerger

(2017) Les associations foncières urbaines:Quand le foncier permet de financer son aménagement. Foncier en debate 17(mai-juin): 3.

43.

Koomen

Dekkers

Broitman

(2018) Analyzing and simulating urban density exploring the difference between policy ambitions and actual trends in the Netherlands. In: Thill

J-C

(ed.) Spatial Analysis and Location Modeling in Urban and Regional Systems. Berlin, Heidelberg: Springer Berlin Heidelberg, pp.145–171.

44.

Kretschmer

Ultsch

Behnisch

(2015) Towards an understanding of land consumption in Germany – Outline of influential factors as a basis for multidimensional analyses. Erdkunde 69(3): 267–279.

45.

Zhou

Ouyang

(2013) Forty years of urban expansion in Beijing: What is the relative importance of physical, socioeconomic, and neighborhood factors? Applied Geography 38: 1–10.

46.

Rhee

, et al. (2014) Building type classification using spatial and landscape attributes derived from LiDAR remote sensing data. Landscape and Urban Planning 130: 134–148.

47.

McGarigal

Marks

(1995) FRAGSTATS: Spatial Pattern Analysis Program for Quantifying Landscape Structure. PNW-GTR-351. Portland, OR: U.S. Department of Agriculture, Forest Service, Pacific Northwest Research Station.

48.

Markstein

(2004) Instrumente und Strategien zur Baulandentwicklung und Baulandmobilisierung in Deutschland, Österreich und der Schweiz ein methodischer Vergleich mit Entwicklungsvorschlägen für das Instrumentarium zur Baulandentwicklung in Deutschland. Universität München. Available at: https://d-nb.info/972068392/34 (accessed 10 February 2021).

49.

Martellozzo

Clarke

(2011) Measuring urban sprawl, coalescence, and dispersal: A case study of Pordenone, Italy. Environment and Planning B: Planning and Design 38(6): 1085–1104.

50.

Meinel

Hecht

Herold

(2009) Analyzing building stock using topographic maps and GIS. Building Research & Information 37(5–6): 468–482.

51.

Olei

(2020) La zone d’aménagement concerté (ZAC). CEREMA. Available at: http://outil2amenagement.cerema.fr/la-zone-d-amenagement-concerte-zac-r311.html (accessed 10 February 2021).

52.

Oliveira

(2016) Urban Morphology: An Introduction to the Study of Physical Form of Cities. The urban book series. Switzerland: Springer.

53.

Oxley

Brown

Nadin

(2009) Review of European Planning Systems. Centre for Comparative Housing Research De Montfort University. Available at: https://www.dora.dmu.ac.uk/bitstream/handle/2086/7536/NHPAU%20Planning.pdf?sequence=1 (accessed 10 February 2021).

54.

Padgham

Lovelace

Salmon

, et al. (2017) Osmdata. The Journal of Open Source Software 2(14): 305.

55.

Rohlfing

(2012) Case Studies and Causal Inference: An Integrative Framework. Research methods series. Basingstoke: Palgrave Macmillan.

56.

Siedentop

(2009) Einflussfaktoren Der Neuinanspruchnahme Von Flächen. Forschungen. Bonn: BMVBS.

57.

Siedentop

(2018) Ursachen der Flächeninanspruchnahme in Deutschland – Eine Zwischenbilanz. In: Behnisch

Kretschmer

Meinel

(eds) Flächeninanspruchnahme in Deutschland: auf dem Wege zu einem besseren Verständnis der Siedlungs- und Verkehrsflächenentwicklung. Berlin: Springer Spektrum, pp.45–55.

58.

Söfker

(ed.) (2018) Baugesetzbuch. 50. Auflage, Stand: 1. Januar 2018, Sonderausgabe. Dtv Beck-Texte im dtv 5018. München: dtv.

59.

Steiniger

Lange

Burghardt

, et al. (2008) An approach for the classification of urban building structures based on discriminant analysis techniques. Transactions in GIS 12(1): 31–59.

60.

Strobl

Boulesteix

A-L

Zeileis

, et al. (2007) Bias in random forest variable importance measures: Illustrations, sources and a solution. BMC Bioinformatics 8(1): 25.

61.

Tennekes

Harbers

Buitelaar

(2015) Institutional arrangements and the morphology of residential development in the Netherlands, Flanders and North Rhine-Westphalia. European Planning Studies 23(11): 2165–2183.

62.

Valavi

Elith

Lahoz-Monfort

, et al. (2019) blockCV: An R package for generating spatially or environmentally separated folds for k-fold cross-validation of species distribution models. Methods in Ecology and Evolution 10(2): 225–232.

63.

Vanderhaegen

Canters

(2017) Mapping urban form and function at city block level using spatial metrics. Landscape and Urban Planning 167: 399–409.

64.

Verikas

Gelzinis

Bacauskiene

(2011) Mining data with random forests: A survey and results of new tests. Pattern Recognition 44(2): 330–349.

65.

Verstegen

Karssenberg

van der Hilst

, et al. (2016) Detecting systemic change in a land use system by Bayesian data assimilation. Environmental Modelling & Software 75: 424–438.

66.

, et al. (2018) An extended minimum spanning tree method for characterizing local urban patterns. International Journal of Geographical Information Science 32(3): 450–475.

67.

Kumar

Ross Quinlan

, et al. (2008) Top 10 algorithms in data mining. Knowledge and Information Systems 14(1): 1–37.

68.

Wurm

Schmitt

Taubenbock

(2016) Building types’ classification using Shape-Based features and linear discriminant functions. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 9(5): 1901–1912.

69.

Zhang

X, AT

Stoter

, (2012) Characterization and detection of building patterns in cartographic data: Two algorithms. In: Yeh

AGO

Shi

Leung

, et al. (eds) Advances in Spatial Data Handling and GIS. Berlin, Heidelberg: Springer Berlin Heidelberg, pp. 93–107.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.49 MB