Developing Subjectively Weighted Attractiveness Scenarios for Urban Allocation Modelling in the Brisbane-South East Queensland RegionChhetri P. CR-SURF, School of Geography Planning and Architecture University of Queensland Brisbane, Qld. p.chhetri@uq.edu.au Corcoran J. CR-SURF, School of Geography Planning and Architecture University of Queensland Brisbane, Qld. Stimson CR-SURF, School of Geography Planning and Architecture University of Queensland Brisbane, Qld. R, Bell M. QCPR, School of Geography Planning and Architecture University of Queensland Brisbane, Qld. Pullar D. CR-SURF, School of Geography Planning and Architecture University of Queensland Brisbane, Qld. Cooper J. QCPR, School of Geography Planning and Architecture University of Queensland Brisbane, Qld. Phone: 07 3365 3565, Fax: 07 3365 3561 ABSTRACT This paper presents a GIS-based methodology to integrate a measure of geographic attractiveness of localities in the process of allocating potential dwellings in the context of a large urban region. The methodology was developed for a study area in Brisbane-South East Queensland (SEQ), known as the Sunbelt Region, a rapidly growing region and a popular tourist destination in Australia. In this paper, we have used a multivariate technique to develop a set of parameterised linear equations to define underlying dimensions that drive residential location decision choices. Aesthetic and accessibility factors were identified in the factor analysis from data collected via a survey of Quality of Life. Spatial measures were based on a combination of network distance and kernel density estimation to calculate aesthetic and accessibility scenarios, which were then overlaid and multiplied by their subjective weights to create a third scenario. These scenarios were integrated as a set of criteria in the urban allocation model. The study found that the areas with high attractiveness scores are concentrated around the western parts of Brisbane and along the coast, while the suburbs in the inner city and along the major transport routes have scored high on the accessibility factor. 1. INTRODUCTION The factors that underlie the residential location decision choice process are complex, being multi-dimensional and multi-scaled in nature (Stimson, 1985; Orford, 1997). The effects that a neighbourhood and its characteristics have on the process of selecting a location for residential purpose are well documented (Stimson 1978; Minchege and Brown 1980; Maher and Saunders 1994; Dökmeci and Berköz 2000; Vogt and Marans 2004). Geographic characteristics of urban residential spaces, in particular accessibility to jobs and amenities, proximity to natural features and socio-cultural networks that people develop in a place are found to be important in the process of residential location selection. Modelling future urban growth, or its manifestations in terms of form and structure, therefore relies on the attractiveness offered by a locality to potential settlers. Most of the earlier attempts have used objective GIS data to control the allocation of dwellings simulated over a period of time across space. However, residential location decisions are far more complex as people exhibit different preferences. They tend to assign different weights to factors that they perceived to be significant for their needs or liking. Often, such factors are related to people’s perceived dwelling needs, their locational preferences, and their knowledge about the area. Those decisions are also related to life-cycle events (for example, children leaving home, marriage, separation and divorce, retirement, and so-on), and they are also embedded in socio-economic factors including housing as a status symbol, an investment vehicle, and a form of stored wealth (Golledge and Stimson, 1997). It is therefore critical for urban modellers to develop methods for embedding these preferences in the process of allocating potential dwellings to an area. Such preferences are often ascertained through social surveys. When these preferences in the form of weightings are integrated into the more objectively defined geographic data to create a ‘subjectively weighted attractiveness index’. This paper develops a weighting regime for factors that drive the future urban growth both through intensification and urban sprawl using survey data collected in the Brisbane-South East Queensland region. 2. URBAN ALLOCATION MODELLING: AN OVERVIEW This paper reports the preliminary outcomes of a project that aims to develop a Predictive Urban Population and Housing Allocation Model (PUPHAM). The PUPHAM is a housing allocation model that had its genesis in the SEQHUM model originally developed for DLGP by Bell and Cooper and the PUP model developed for Adelaide by Bell et al. (2000). The methodology proposed in this paper is an extension of the existing PUPHAM model, which explores the opportunity to enhance the allocation process through the use of a ‘Subjectively Weighted Attractiveness Index’. A key feature of POPHAM is the determination of a development envelop, computed as the difference between current and potential residential density, which defines the scope for future housing development in a given geographic locality such as a grid cell. Potential residential density is computed by reference to a planning scenario, which determines the amount of land available for housing and the dwelling density associated with each land class. Aggregation of total current and potential dwellings across all land parcels and land classes generates the current and potential residential density. The difference between these two reflects the ‘dwelling capacity’ of the grid cell to absorb additional dwellings. The task of the allocation model is to allocate regional housing growth across grid cells within the development envelope. It is envisaged that this allocation will be driven by reference to a number of forces, which shape the pattern of housing distribution. These may be primarily spatial in nature (e.g. distance from the CBD or transport routes), in addition to economic, social or physical criteria applied under different weighting regimes to model in housing patterns. This paper is an attempt to develop a set of scenarios offering different weighting regimes to incorporate the effects of these factors in the allocation model. 3. STUDY AREA The Brisbane - (SEQ) region has been experiencing rapid growth and socio-economic transformation over the last two to three decades. The region is characterised by a poli-centric urban structure, connecting the State capital Brisbane with two coastal growth corridors, south to the Gold Coast and north to the Sunshine Coast, and with a less radically growing western corridor through Ipswich, a long-established industrial and mining city. The region’s population has increased from 1.8 to 2.35 million between 1991 and 2001, and it is forecasted to reach 3.2 million by 2011. 4. METHODOLOGY This section describes the datasets that are used to generate a set of Subjectively Weighted Attractiveness (SWA) scenarios. It then outlines a statistical technique to identify the drivers of residential location decision choices and reports on the spatial techniques that determined a range of objective GIS datasets. It then develops a spatial approach for generating a number of scenarios and demonstrates their integration into the urban allocation model. 4.1 Data In 2003, a stratified probability sample survey of Quality of Life (QOL) in the Brisbane–SEQ region was conducted using the Computer Assisted Telephoning Interviewing (CATI) facility at the University of Queensland. A total of 776 survey participants aged 18 years and older were interviewed. Respondents were asked to indicate the relative importance of each of a series of items in their decision to move to their current place of residence. The relative importance of each of the thirteen items was ascertained on a five-point Likert scale where 1 represented ‘not at all important’ and 5 ‘very important’. The items included are: ‘close to work’; ‘convenience to shopping centres and schools’; spaciousness of the area’; close to natural areas (creeks, parks, beaches etc); close to public transport’; ‘attractive appearance of neighbourhood’; and ‘lots of recreational opportunities’. Digital data were acquired from MapInfo StreetPro (8.0.1) database, Department of Local Government and Planning and the Australian Bureau of Statistics 2001 Census. More specific sub-layers - such as shopping centres, schools and recreation facilities (Sport Facility, Cricket, Soccer, Hockey, Baseball, Entertainment Venues and the like) - were extracted as separate layers from the ‘Features’ layer in StreetPro database. The data from Queensland Department of Local Government and Planning were used to identify areas assigned as industrial and commercial zones. 4.2 Statistical Techniques Factor analysis tools are commonly used multivariate data reduction techniques (Chhetri and Arrowsmith 2003; Hair et al., 1995). PCA facilitates the exploration of manifest data to identify latent components from a set of interrelated variables (that is, the neighbourhood attributes in the survey questionnaire). A reduced number of new variables, known as components, are obtained from highly correlated variables. More than 70 percent of the correlations are found to be significant at the 0.05 level (at the 95 percent confidence level). This means that PCA can be considered an appropriate technique to be used on this dataset. In addition, the calculated Kaiser-Meyer-Olkin (KMO) of 0.712 clearly exceeds the 0.5 level considered to be acceptable for the use of PCA (Hair et al., 1995). The selected model generated by PCA with varimax rotation, extracted two components with eigenvalues greater than 1 (refer to Table 1). These components were also tested for internal reliability using Cronbach’s alpha. The first component accounts for the greatest proportion of variance (eigen value 2.1) and the four items defining have high loadings from 0.556 to 0.818. The second component has a lower eigen value (1.86) and the four items defining it have loadings from 0.585 to 0.820. These components are descriptively, not normatively labelled and together explain 56.85 percent of the variability in observations.  The items that load heavily on Component 1 are: openness or spaciousness; closeness to natural areas (bush, creeks, beaches etc.); attractive appearance of neighbourhood; many recreational opportunities and community size. This has been labelled the aesthetic factor. It accounts for 30.28 percent of the total variance. This factor could be understood by the fact that people perhaps perceive neighbourhoods in terms of their visual features or biophysical characteristics. The residential decision choices are made on the basis of the offered quality of those features. Therefore, the visual appearance and aesthetic appeal of the neighbourhood are the driving factors of residential location decisions for some residents. Component 2 is defined by: closeness to work; convenience to shopping centres and schools; and proximity to public transport. This component explains 26.57 percent of variability in the data and is named the accessibility factor as it is related to the relative availability of housing services and facilities. The importance of being adjacent to features of socio-economic importance, including workplaces, is represented by this factor. The significance of neighbourhood is assessed by its ability to facilitate easy access to amenities such as shopping centres and schools. It allows people to be connected and less restricted by the constraints imposed by space and time. It is important for people to know how certain locations are organised in relative space and how well they are connected to features of social and economic significance. 4.3 Spatial Techniques So far the data used in the analysis were subjective evaluations collected through the QOL survey. Such data hold no explicit geographic properties. In order to build a spatial model of attractiveness over a larger geographic space, objective datasets measuring characteristics of the environment itself were needed. Objective datasets were quantified or approximated from digital data. Some indicators were directly derived from the spatial and biophysical characteristics - such as distance, accessibility, and proportion of open space to total area - while others are more complicated (for example, slope diversity). Objective datasets for the items identified in the factor analysis were generated so that the geographic variability in each dimension could be mapped. These objective datasets were derived as follows: 4.3.1 The neighbourhood operation The Neighbourhood Operation routine is often used in a situation where relationships between locations or surrounding areas need to be evaluated. Spatial relations provide information about what, where and how features of socio-economic importance are spatially distributed, arranged and organised within a given neighbourhood. Using kernel density estimation (KDE) (Silverman, 1986) density surfaces were constructed, (see figure 1). In order to complete this task, three parameters need to be specified. These include: (i) one or more foci (parameter 1); (ii) the neighbourhood membership, for example a set of locations around each focus cell that are within a specified distance or direction (parameter 2); and finally (iii) a function to be performed on the cells within the defined neighbourhood (parameter 3).  Figure 1: The process of estimating density using a Kernel (Source: Bailey and Gatrell, 1995) Parameter 1 is a 1 x 1 km cell in the grid generated; parameter 2 is the 2.5 km radius assigned to define the scanning neighbourhood; and parameter 3 is a function. For example, using an average function, the mean value of the data points (respondents’ scores) present in the scanning neighbourhood of a 2.5 km area was calculated which was then assigned to the focal cell, situated at the centre of the moving window. The procedure is then replicated for each cell across the entire grid. The following equations were used to generate three grids representing three components: The neighbourhood operation was employed to calculate three objective indicators. These include ‘lots of recreational opportunity’, ‘slope diversity’ and ‘density of bus stops’. Using a digital terrain model, elevation diversity was calculated as the standard deviation of height within a 2.5 km radius of the spatial filter. The higher value indicates the greater elevation diversity. It was calculated as: The survey item ‘lots of recreational opportunity’ was measured as the number of recreation facilities available for contact or interaction from a given point or location within its surrounding neighbourhood. The radius for the neighbourhood operation is defined as a limit beyond which people are less likely to travel to visit a facility. A radius of 2.5 km, as people often drive to those facilities, was assumed as a commutable distance, which roughly corresponds to the average size of a neighbourhood, except in rural areas. Using the StreetPro database, neighbourhood operation was conducted on a point data layer of sporting facilities and a count for each cell in the output grid was calculated as: 4.3.2 Accessibility measures In this study two types of accessibility measures were utilised, Euclidean distance and a composite distance (an additive Euclidean and network distance measure). A composite measure of network and Euclidean distance is used here to evaluate the proximity of grid cells to various facilities. This bespoke distance measure was conceived out of the failure of either Euclidean or network distance to provide a good indication of proximity across the entire study area. In areas where natural features are present (such as a river) the use of Euclidean distance tends to misrepresent the true proximity of a grid cell. This is especially true where the road network is spare and the number of crossing opportunities across rivers is limited.  Figure 2: Distance measures In areas of high road network density the use of network distance provides good estimates of accessibility, however, where the road network is sparser this measure fails to capture the true distance from the origin to destination, through the exclusion of the distance from the origin to the road network (see figure 2). Here we have used a composite measure to overcome some of the disadvantages of both Euclidean and network distance measures. Euclidean distance was used to estimate accessibility to three survey items - ‘close to natural areas’, ‘distance to industrial and commercial areas’ and ‘proximity to coast’. Composite distance to the CBD and Euclidean distance to industrial and commercial areas were merged to derive the item ‘close to work’. While for proximity to public transport, network distance to the nearest railway station and density of bus stops were calculated. Similarly, the survey item ‘close to natural areas’ was also generated by amalgamating three layers: distance to creeks, distance to parks and reserves, and distance from the coast. Finally, two layers were used to measure attractive appearance of landscape. These include indices for open space and elevation diversity. The open space index was calculated as the proportion of open space to the total area for each cell.  The GIS layers created in the previous section were in different metrics. For example, distance cannot be readily comparable with the number of shopping centres or with dwelling density. Therefore, the data were standardised in order to remove the effect of measurement units. The data were converted to standard scores by subtracting the mean and dividing by the standard deviation for each variable (Hair et al. 1998). This converts raw data into a standardised value with a mean of 0 and a standard deviation of 1. 5. GENERATION AND INTEGRATION OF SCENARIOS IN THE ALLOCATION MODELLING The factors identified in the statistical analysis can be integrated as scenarios in the urban allocation model. A scenario refers to a set of possible events or situations which can be related to a given state or time period usually, but not always, the future (Ducot and Lubben 1980). In simple terms, a scenario can be described as a hypothetical form, for example, ‘if X then Y’ that can be developed to predict the growth, form and characteristics of a phenomenon. In the context of urban growth, a hypothetical form would be if everyone in the population would like to move to a place for a simple reason to be close to natural environments or the aesthetic quality of landscapes. It is important to know the likely outcome of the allocation model in such a ‘hypothetical situation’ where aesthetic characteristics are paramount. Using the factors that people consider important in their decision to move in SEQ region, three different ‘descriptive scenarios’ can be developed to reflect specific choices leading to specific inferences.
To develop a spatial model, it is important to translate subjective weightings ascertained from the QOL data into objective GIS datasets generated in section 4.3. One approach is to develop a set of linear parameterised equations using the regression-based factor-score coefficients matrix. The matrix is an array of coefficients that are similar to ‘regression coefficients’ and are used to generate factor scores from variables (Nei et al., 1975: 488; Tabachnick and Fidell 1996: 591). The use of factor scores for mapping is commonplace in the geographical literature (Burley and Brown 1995; Western and Larnach 1998; Kliskey 2000) the employment of factor score coefficients matrix as a set of linear equations is also supported (Nei et al., 1975; Tabachnick and Fidell 1996; Arrowsmith and Inbakaran 2001; Chhetri and Arrowsmith 2002). The objective indicator data for each factor were simply multiplied by their respective factor coefficient scores and were then added. There were two linear equations representing the factors, as a result two scenarios were produced. The following linear models were used: “Scenario One: Aesthetic factor” = .818v1 + .800v2 + .701v3 + .556v4 + .100v5 + .005v6 + .020v7 “Scenario Two: Accessibility factor” = -.071v1 + -.063v2 +.216v3 + .464v4 +.820v5 + .758v6 + -.585v7 Where, v1 = standardised value of close to natural areas v2 = standardised value of spaciousness v3 = standardised value of attractive appearance v4 = standardised value of lots of recreational opportunities v5 = standardised value of convenience to shopping centres v6 = standardised value of public transport v7 = standardised value of close to work So far the characteristics and spatial patterns of the two dimensions that underlie perceptions of neighbourhood attractiveness that influence the decision to move have been explored and analysed. However, those areas that are perceived to be attractive or poor on all these dimensions not have been yet identified or mapped. By amalgamating scores on the three dimensions of neighbourhood attractiveness, we can form the scenario, which indicates the combined effect of these two factors on the residential location decision choices of people. This will provide an indication of the variation across an urban area of the relative attractiveness of one suburb vis a vis others, and will permit the identification of the degree of spatial concentration and dispersal of places of high and low residential attractiveness across the totality of residential territorial space of the Brisbane-SEQ region. The two factors of attractiveness identified earlier however carry different levels of importance in terms of their ability to influence residential location decision choices. These were reflected by the eigenvalue of each of the factors derived from the PCA. Those eigenvalues may be used to weight survey respondent scores on the three dimensions of neighbourhood attractiveness and the scores of suburbs on those three factors. To derive the overall rating of areas on the third scenario, the two scenarios were simply multiplied by their respective weights (eigenvalues), and were added and then divided by the sum of total weights. The methodology used here is similar to the recreational suitability indices adopted by Levinsohn et al., (1987) and Kliskey (2000). Thus, the scenario 3 is calculated as: Scenario Third Subjectively Weighted Attractiveness = (2.1SRv1 + 1.86SRv2)/3.98 where, SRv1 =subjective rating for aesthetics SRv2 = subjective rating for accessibility On the regional level, the first map (refer to Figure 1) dealing with the aesthetic factor indicates that areas with highest scores are neatly aligned along the coast, indicating the effects of proximity to natural areas such as creeks, beaches and parks. Some interior parts of the Brisbane-SEQ region, despite their closeness to natural areas, have low to moderate scores, perhaps due to the presence of few recreational opportunities, while the coastal areas - particularly around the Gold Coast and the Sunshine Coast - seem to have the comparative advantage of considerable investment on recreation infrastructure in recent years. Broader areas, such as Ipswich in the western corridor of the region, the western part of Logan, Beaudesert, Caloundra (Hinterland) and Caboolture (Part B) along with other interior regions, had low scores on this factor. The second map (See Figure 2) identifying the accessibility factor also reveals a definitive pattern of contiguous areas. The areas identified in the map with high scores are potentially attractive to those people who consider accessibility and housing cost as important in their residential location decision choices. The overall pattern is CBD-centric and exhibits a somewhat concentric pattern with distortions along the transport arteries producing a wedge like structure.   Clear spatial differences were captured across the entire SEQ region. The data suggest that there are some areas which have more access to amenities, particularly around inner Brisbane and along Brisbane River and Motorways, leaving vast areas relatively less connected to job opportunities and regional activity centres. It was clear that these dimensions have large geographic variations, which perhaps could be attributed to the differences in the urban morphological landscape (for example, nature features, transportation networks and land use) or urban ecological patterning (for example, demographics, lifecycle stages, income status, and ethnic segregations). It can therefore be inferred that people perhaps adjust their residential location decision choices as a trade-off between their concurrent needs for accessing job availability, amenities, recreational opportunities and social events. This led to the construction of three scenarios that may allow urban modellers to develop a flexible interface in a spatial decision support system. These scenarios were generated as grids with scores assigned to every cell that are subjectively weighted on the basis of people’s evaluation of the reasons to move in their current place of residence. Significant differences were detected in the mapping outcomes where certain areas were identified to have greater advantage on aesthetic or accessibility criteria. REFERENCE
|