Land Use Change Simulation and Prediction Using Integrated Cellular Automata Models: a Case Study of Ipswich City, Australia

Land use change simulation and prediction using integrated cellular automata models: A case study of Ipswich City, Australia

Yi Lu

A thesis in fulfilment of the requirements for the degree of

Doctor of Philosophy

School of Biological, Earth and Environmental Sciences Faculty of Science The University of New South Wales

December 2019

Thesis/Dissertation Sheet

Surname/Family Name : Lu Given Name/s : Yi Abbreviation for degree as give in the University calendar : PhD Faculty : Science School : School of Biological, Earth and Environmental Sciences Land use change simulation and prediction using integrated cellular automata Thesis Title : models: A case study of Ipswich City, Australia

Abstract 350 words maximum: (PLEASE TYPE)

Land use change models are important tools for understanding geographical phenomena and their dynamic interactions with other systems. They play a key role in the fields of land resource management, spatial science and urban planning. Among all land use change models, cellular automata (CA) have been widely explored and used in the literature, with considerable contributions to theories, methods and corresponding applications all around the world. An important issue with CA models is the unavoidable precision-loss with traditional raster-based CA and the demand of high-accuracy modelling tools.

In this thesis, a vector-based CA model is proposed, implemented as prototype, and then integrated with artificial intelligence algorithms and a planning support system. Taking Ipswich City, South East Queensland (SEQ) Region, Australia as the study region, the thesis provides a comprehensive exploration of vector-based CA modelling, with macro and micro spatial variables, qualitative and quantitative evaluation. Key topics assessed are the comparison, simulation and evaluation of vector and raster-based CA models; the effects of spatial heterogeneity and partitioned transition rules at multi-level spatial scales; the integration of vector-based CA model with planning support system for predicting future scenario developments. In general, the thesis has explored multiple research points of vector-based CA modelling, with a mixture of data sources. The characteristics and mechanisms of historical and future residential expansion, as well as its connection with government policies and social-economic factors, have been analysed and illustrated.

In conclusion, vector-based CA model could reduce the sensitivity of cell size and computation time effectively, identify the nonlinear connections between spatial variables and land use patterns, as well as forecasting the future development in a logical and coherent way.

Declaration relating to disposition of project thesis/dissertation

I hereby grant to the University of New South Wales or its agents the right to archive and to make available my thesis or dissertation in whole or in part in the University libraries in all forms of media, now or here after known, subject to the provisions of the Copyright Act 1968. I retain all property rights, such as patent rights. I also retain the right to use in future works (such as articles or books) all or part of this thesis or dissertation.

I also authorise University Microfilms to use the 350 word abstract of my thesis in Dissertation Abstracts International (this is applicable to doctoral theses only).

…………………………………………………………… ……………………………………..……………… ……….……………………...…….… Signature Witness Signature Date The University recognises that there may be exceptional circumstances requiring restrictions on copying or conditions on use. Requests for restriction for a period of up to 2 years must be made in writing. Requests for a longer period of restriction may be considered in exceptional circumstances and require the approval of the Dean of Graduate Research.

FOR OFFICE USE ONLY Date of completion of requirements for Award:

ORIGINALITY STATEMENT

‘I hereby declare that this submission is my own work and to the best of my knowledge it contains no materials previously published or written by another person, or substantial proportions of material which have been accepted for the award of any other degree or diploma at UNSW or any other educational institution, except where due acknowledgement is made in the thesis. Any contribution made to the research by others, with whom I have worked at UNSW or elsewhere, is explicitly acknowledged in the thesis. I also declare that the intellectual content of this thesis is the product of my own work, except to the extent that assistance from others in the project's design and conception or in style, presentation and linguistic expression is acknowledged.’

Signed ……………………………………………......

Date ……………………………………………......

‘I hereby grant the University of New South Wales or its agents a non- exclusive licence to archive and to make available (including to members of the public) my thesis or dissertation in whole or part in the University libraries in all forms of media, now or here after known. I acknowledge that I retain all intellectual property rights which subsist in my thesis or dissertation, such as copyright and patent rights, subject to applicable law. I also retain the right to use all or part of my thesis or dissertation in future works (such as articles or books).’

‘For any substantial portions of copyright material used in this thesis, written permission for use has been obtained, or the copyright material is removed from the final public version of the thesis.’

Signed ……………………………………………......

Date ……………………………………………......

AUTHENTICITY STATEMENT

‘I certify that the Library deposit digital copy is a direct equivalent of the final officially approved version of my thesis.’

Signed ……………………………………………......

Date ……………………………………………......

INCLUSION OF PUBLICATIONS STATEMENT

UNSW is supportive of candidates publishing their research results during their candidature as detailed in the UNSW Thesis Examination Procedure.

Publications can be used in their thesis in lieu of a Chapter if: • The student contributed greater than 50% of the content in the publication and is the “primary author”, ie. the student was responsible primarily for the planning, execution and preparation of the work for publication • The student has approval to include the publication in their thesis in lieu of a Chapter from their supervisor and Postgraduate Coordinator. • The publication is not subject to any obligations or contractual agreements with a third party that would constrain its inclusion in the thesis

Please indicate whether this thesis contains published material or not.

☐ This thesis contains no publications, either published or submitted for publication

Some of the work described in this thesis has been published and it has been ☒ documented in the relevant Chapters with acknowledgement

This thesis has publications (either published or submitted for publication) ☐ incorporated into it in lieu of a chapter and the details are presented below

CANDIDATE’S DECLARATION I declare that: • I have complied with the Thesis Examination Procedure • where I have used a publication in lieu of a Chapter, the listed publication(s) below meet(s) the requirements to be included in the thesis. Name Signature Date (dd/mm/yy) Yi Lu

Abstract

In this thesis, a vector-based CA model is proposed, implemented as a prototype, and then integrated with artificial intelligence algorithms and a planning support system.

Taking Ipswich City, South East Queensland (SEQ) Region, Australia as the study region, the thesis provides a comprehensive exploration of vector-based CA modelling, with macro and micro-scale spatial variables, qualitative and quantitative evaluation. Key topics assessed are the comparison, simulation and evaluation of vector and raster- based CA models; the effects of spatial heterogeneity and partitioned transition rules at multi-level spatial scales; the integration of vector-based CA model with planning support system for predicting future scenario developments. In general, the thesis has explored multiple research points of vector-based CA modelling, with a mixture of data sources. The characteristics and mechanisms of historical and future residential expansion, as well as its connection with government policies and social-economic factors, have been analysed and illustrated.

Acknowledgements

First and foremost, the excellent and persistent guidance from my primary supervisor

Prof. Shawn Laffan can never be exaggerated. He is not only a GIS expert but also a philosopher, and boosts my outlook on life. Academic writing, modelling and analysis knowledge, communication and presentation skills, it is difficult to list everything he has ever taught me. Nevertheless, it is Shawn’s encouragement that assists me to conquer the difficulties during my PhD study. As he suggests, there is always a solution to the problem. Similarly, there is always one more thing I wish to learn from him.

Besides, I would like to thank my co-supervisor Prof. Chris Pettit for his valuable instruction during my PhD study. As the co-author of Chapters 2, 4 and 5, Chris has improved their qualities in a further step. In addition, it is beneficial from the tutorial and research opportunities he has provided, which lays a solid foundation of my career development.

I wish to extend my gratitude to other contributors of my PhD thesis and research projects. My panel members: Prof. Graciela Metternicht, Prof. Michael Letnic, A/Prof.

Suhelen Egan and A/Prof. Wendy Shaw. They have made efforts to make sure I'm stay on the track, and set the milestone for me in advance. My colleagues and friends: Dr. Hugh

Burley, Dr. Adrian Fisher, Dr. Anthea Mitchell, Ms. Lauren Del Rosario, Dr. Xianglin Zheng,

Dr. Jin Zhu, Ms. Jiasui Li and Ms. Jingyi Ding. They have provided valuable feedback on my research and thesis. Mr. Jonathan Russell, the Administrative Officer of my school who is always helpful. A/Prof Min Cao (Nanjing Normal University), the co-author of

Chapter 2.

iii

It is acknowledged that the spatial data of this thesis are mainly provided by the

Australian Bureau of Statistics (ABS) and Queensland Spatial Catalogue (QSpatial).

Financial support from UNSW-CSC joint scholarship is also appreciated.

Finally, I'd like to express my most sincere thanks to all my family members. You raise me up and respect the decision I have made to be a PhD candidate. You are the source of energy, which supports me to keep on going during the memorable 4-year PhD study.

Table of Contents

Abstract ...... i

Acknowledgements ...... iii

Table of Contents ...... v

List of Figures...... ix

List of Tables ...... xi

Chapter 1 Introduction ...... 1

1.1 Overview ...... 1

1.2 A review of land use change models ...... 2

1.2.1 System dynamic (SD) model ...... 3

1.2.2 Markov model ...... 6

1.2.3 Conversion of Land Use and its Effects at small regional extent (CLUE-S) model ... 8

1.2.4 Cellular automata (CA) model ...... 12

1.2.5 Agent-based model (ABM) ...... 18

1.3 Structure of the thesis ...... 22

1.3.1 General objective ...... 22

1.3.2 Thesis outline ...... 23

1.4 Research significance ...... 25

1.4.1 A systematic exploration of vector CA modelling...... 26

1.4.2 The integration of CA and planning support system ...... 26

1.4.3 The comprehensive evaluation of modelling outcomes ...... 27

Chapter 2 Land use change simulation and analysis using a vector-based cellular automata (CA) model ...... 28

Abstract ...... 28

2.1 Introduction...... 29

2.2 Structure of CA models ...... 32

2.2.1 The differences between vector and raster CAs ...... 32

2.2.2 Workflow of CA modelling ...... 33

2.2.3 Transition rule discovery with artificial neural networks ...... 34

2.2.4 Accuracy assessment ...... 35

2.3 Application ...... 37

2.3.1 Study area and data processing ...... 37

2.3.2 Driving factors ...... 40

2.3.3 Model implementation ...... 43

2.3.4 Simulation results ...... 44

2.4 Discussion ...... 48

2.5 Conclusions ...... 51

Chapter 3 The integration of vector-based cellular automata and partitioned rules for simulation of land use change...... 52

Abstract ...... 52

3.1 Introduction...... 53

3.2 Methodology ...... 54

3.2.1 Using PSO for the discovery of transition rules ...... 54

3.2.2 Validation method ...... 56

3.2.3 Model implementation ...... 57

3.3 Case study ...... 58

3.3.1 Study area and data processing ...... 58

3.3.2 Simulation process and result ...... 60

3.4 Discussion and conclusions ...... 65

Chapter 4 A geographically partitioned cellular automata model for the expansion of residential areas ...... 67

Abstract ...... 67

4.1 Introduction...... 68

4.2 Methodology ...... 72

4.2.1 The structure of vector CA model ...... 72

4.2.2 Calibration of transition rules ...... 74

4.2.3 Model validation ...... 76

4.2.4 Workflow of general and partitioned CA models ...... 77

4.3 Case study ...... 79

4.3.1 Study area and data processing ...... 79

4.3.2 Model implementation ...... 87

4.3.3 Simulation results and assessment ...... 89

4.4 Discussion ...... 95

4.5 Conclusions ...... 97

Chapter 5 The integration of cellular automata and What If? for scenario planning:

Future residential expansion in Ipswich City ...... 99

Abstract ...... 99

5.1 Introduction...... 100

5.2 Methodology ...... 103 vii

5.2.1 General workflow of the CA – What If? model ...... 103

5.2.2 Population projection and scenario construction ...... 106

5.2.3 CA-based allocation process...... 109

5.2.4 Scenario evaluation ...... 110

5.3 Case study ...... 112

5.3.1 Study area description and data sources ...... 112

5.3.2 Future scenario of Ipswich ...... 113

5.3.3 Land use allocation by CA models ...... 117

5.3.4 Evaluation of scenario development ...... 118

5.4 Discussion ...... 121

5.5 Conclusions ...... 123

Chapter 6 Conclusions and future research ...... 125

6.1 Main conclusions ...... 126

6.2 Implications for future research ...... 132

6.2.1 The generality of vector CA modelling ...... 132

6.2.2 Sensitivity analysis of vector CA modelling in different aspects ...... 132

References ...... 134

viii

List of Figures

Figure 1.1 The framework for calculating land use demand, based on an SD model (He et al.,

2008)...... 4

Figure 1.2 The workflow of a Markov land use change model, modified from Nasiri et al. (2019).

...... 7

Figure 1.3 The workflow of a CLUE-S land use change model (Verburg and Overmars, 2007). . 11

Figure 1.4 The basic structure of a CA model, modified from Lu et al. (2019)...... 13

Figure 1.5 The framework of an agent-based land use model (Schwarz et al., 2012)...... 19

Figure 1.6 The diagram of CA models in Chapters 2 - 5...... 23

Figure 2.1 Example neighbourhood configuration of vector CA using a buffer of fixed radius. . 33

Figure 2.2 Workflow of CAs for city growth modelling...... 34

Figure 2.3 The study area is Ipswich City in Queensland, Australia. Source of satellite image:

ESRI, downloaded: 06 Jan 2018...... 38

Figure 2.4 Population density of the study area in ABS meshblocks (2006 and 2016)...... 41

Figure 2.5 Simulation process of land evolution using raster and vector CAs. Locations are

labelled in Figure 2.3. Source of satellite image: ESRI, imagery downloaded: 18 April 2018.

...... 46

Figure 2.6 Cumulative producer’s spatial accuracy of CA models...... 47

Figure 2.7 Distribution of misclassified cells in raster (A) and vector (B) CA models. Source of

base map: ESRI...... 48

Figure 3.1 Diagram of partitioned PSO-CA...... 57

Figure 3.2 The study area in Ipswich, Queensland, Australia...... 59

Figure 3.3 Simulation processes in entire/sub region(s)...... 62

Figure 3.4 Ratios of misclassified cells with different frequencies (%)...... 63

Figure 4.1 Spatial heterogeneity of a sample study region...... 71

Figure 4.2 Structure of vector CA model...... 73

Figure 4.3 Workflow of general and partitioned PSO-CAs...... 78 ix

Figure 4.4 Primary land uses of Ipswich City, Australia (year 2016)...... 79

Figure 4.5 The partition of sub-regions in study area. Noting the differing levels of influences

from variables...... 86

Figure 4.6 The fitness curves of PSO training in general and partitioned CAs...... 89

Figure 4.7 Simulated distribution of new residential cells (1999 to 2016) using general and

spatially partitioned PSO-CA models. The majority of changes are observed between four

suburbs (see panel G)...... 93

Figure 4.8 Rates of misclassified cells with different frequencies (%)...... 94

Figure 5.1 The implementation of integrated CA – What If? model...... 105

Figure 5.2 The location of Ipswich City...... 112

Figure 5.3 Suitability maps of scenario residential development...... 117

Figure 5.4 Predicted scenario developments of new residential areas (2016-2031)...... 118

Figure 5.5 Classification of parcel size (by area)...... 119

Figure 5.6 Urban areas for priority new residential developments...... 121

Figure 6.1 The main spatial variables of sub-regions with positive/negative contribution,

Chapter 3...... 128

Figure 6.2 The main spatial variables of sub-regions with positive/negative contribution,

Chapter 4...... 129

List of Tables

Table 1.1 Evolution of the urban, rural and total population of the world, by geographic regions,

1950–2050 (United Nations, 2018)...... 3

Table 2.1 Data sources used in the case study...... 39

Table 2.2 Driving factors of land use change in the study area...... 40

Table 2.3 Overall error statistics...... 47

Table 2.4 Misclassification frequency (%)...... 48

Table 3.1 The transformed and stable cells of study area during 1999 to 2016...... 60

Table 3.2 Comparison of average weights...... 61

Table 3.3 The PSA and CPSA values of sub and entire study areas...... 64

Table 4.1 Data sources used in the case study...... 80

Table 4.2 Land use statistics of study area(ha)...... 81

Table 4.3 Spatial variables of land use development...... 82

Table 4.4 Pre-test result of neighbourhood sensitivity ...... 85

Table 4.5 Comparison of average weights in general and partitioned PSO-CAs...... 90

Table 4.6 Average values of CPSA and FoM for residential cells (year 2016)...... 94

Table 5.1 The factors and categorized values of suitability sub-scenario...... 113

Table 5.2 Demographic trends in Ipswich (years 2006, 2011 and 2016)...... 114

Table 5.3 Detailed parameters of constructed scenario...... 116

Table 5.4 Suitability categories of land parcels...... 116

Table 5.5 Mean value and standard deviation of patch size...... 118

Table 5.6 Areas of new residential parcels (years 2016 - 2031) within planned suburbs...... 120

Chapter 1 Introduction

1.1 Overview

We are in the era of global urbanization (Turner et al., 2004; Coward, 2012), with cities increasing in both number and size across the globe (Henderson and Wang, 2007). A city is defined as a large settlement with high population densities (Goodall, 1987; Kuper, 2004). Cities mainly grow through a bottom-up process, as their size and shape follow the well-defined scaling laws that result from intense competition for space (Batty, 2008). As a complex system, a city has extensive subsystems – such as housing, infrastructure, land use, and transportation systems – in different sectors. Among all these subsystems, land use indicates how our land resources are applied, including the production of goods and services (Australian Bureau of Agricultural and Resource Economics and Sciences, 2019). Therefore, land use is regarded as the foundation of urban planning and management since it specifies the permitted or prohibited development in different land use zones (New South Wales Government, 2019).

In 1995, the Land Use and Land Cover Change (LUCC) International Project, which explores the interactions between geo-environmental and production systems, was proposed by the International Geosphere-Biosphere Program (Seitzinger et al., 2015) and the International Human Dimension Program (Janssen et al., 2006). One of its key aims is to develop robust and regionally sensitive global models of land use change with improved capacities to predict its evolution (Turner et al., 1995). Land use change models are defined as tools for supporting the analysis of the causes and consequences of land use dynamics (Verburg et al., 2004b). The theories, methods, and applications of land use change modelling have been a frontier research area of geography: it assists in land-use planning and policy making by exploring possible scenarios and alternative land use configurations, as well as the stability of linked social and ecological systems (Veldkamp and Lambin, 2001; Couclelis, 2005; Verburg, 2006).

According to the World Urbanization Prospects (2018 revision), there is an obvious increase in the world’s urban population ratio, which has ranged from 30% to 55% since the middle of the last century. It is also projected that the ratio will keep increasing in the following decades, and could be as much as 68% by the end of the year 2050 (United Nations, 2018). As Table 1.1 illustrates, the predicted differences in urban population growth will accentuate the redistribution of the urban population that occurred between 1950 and 2018. Along with research on urban population growth and migration flows at a global-scale, it is anticipated that there will be continued research of land use change modelling in terms of its integration and connection with biodiversity (Haines-Young, 2009; Laffan et al., 2012), climate change (Dale, 1997; Sertel et al., 2011; Bright et al., 2017), economics (Newburn et al., 2005; Kassie et al., 2010) and especially urban sprawl (Herold et al., 2003; Heppenstall et al., 2011; Rahman, 2016). The potential regulations of land use change could be identified from the urban system in terms of its dynamic, flexible and nonlinear characteristics (Batty, 2009a). On the other hand, for the science community, land change models are important in their ability to test theories and concepts of land change and its connections to human-environment relationships, as well as to explore how these dynamics will change future land systems without real- world observation (Brown, 2014).

1.2 A review of land use change models

In the field of land-use change modelling, five categories of models have been summarized on the basis of their characteristics, features and operational mechanisms: system dynamics (SD) model, Markov model, conversion of land use and its effects at small regional extent (CLUE-S) model, multi-agent systems (MAS) model, and cellular automata (CA) model (Yang et al., 2016). Among all of these models, CA is considered an advanced and intelligent tool since its “simplified” operation rules on micro cells can reflect the dynamic land use variations of the entire study area, which is in accordance

with the essence of complexity science: a complex system is derived from the interactions of simple subsystems (Mizraji, 2006; Li et al., 2007). Table 1.1 Evolution of the urban, rural and total population of the world, by geographic regions, 1950–2050 (United Nations, 2018).

1.2.1 System dynamic (SD) model The SD model was derived from differential equations in the early 1970s (Forrester, 1970), and used for understanding the nonlinear behaviours of complex systems over time using stocks, flows, internal feedback loops, table functions and time delays (Rahim et al., 2017). A typical structure of SD model is presented in Figure 1.1. The SD model has been applied mainly in the areas of agriculture, ecology and rural and urban land use change. In the first field, researchers have developed System Dynamic Simulation (GAPSIM)(Saysel et al., 2002), Agricultural-Institutional-Social-Eco-logical-Economic (AISEEM)(Shi and Gill, 2005), Agriculture Effect Policy System Dynamics (AEP-SD)(Li et al., 3

2012) and Volta River Basin System Dynamics (VRB-SD)(Kotir et al., 2016) models for assessing agricultural development under different scenarios and policies, as well as the environmental, economic and social effects of such development. It is indicated by these modelling works that the SD model helps policy-makers achieve an objective assessment of expected system responses and impacts, which is helpful to identify the optimal land use strategy for specific regions and work out detailed programs to achieve the target of sustainable development.

Figure 1.1 The framework for calculating land use demand, based on an SD model (He et al., 2008).

In terms of transportation modelling, Abbas and Bell (1994) reviewed the appropriateness and suitability of the SD model in this field. According to Shepherd (2014), transportation SD models can be classified into five types: modelling the uptake of alternative fuel vehicles (Ulli‐Beer et al., 2010; Pasaoglu et al., 2016), supply chain management with transportation (Tako and Robinson, 2012; Fera et al., 2017), highway maintenance/construction (Fallah‐Fini et al., 2010; Guevara et al., 2017), strategic policy at different spatial scales (Pfaffenbichler et al., 2010; Wang et al., 2015), as well as aircraft industry (Suryani et al., 2010; Langroodi and Amiri, 2016). The SD approach is

able to provide a holistic system model which deals with feedbacks and delays between system actors. Therefore, SD-based transportation models are more easily linked with other sectors by taking multi-scale time delays and feedbacks into consideration.

Aside from conducting agricultural and transportation research, researchers also pay close attention to land use change simulation in the application of SD models. Shen et al. (2009) constructed an SD model with five subsystems (population, economic, housing, transport and land), and analysed the impacts of different urban development strategies in Hong Kong, China. Based on land resource, population growth and social-economic factors, Yu et al. (2011) created an SD model for the simulation of future LUCC (a mixture of agricultural, grassland, forests, wetlands, unused and built-up) under different policies. Another case study involves three rural counties of the northern United States (Hyde, South Dakota; McHenry, North Dakota; Custer, Nebraska) for predicting future land transformation scenarios of these counties by integrating ecological, economic, and social components (Turner et al., 2016). These simulation results indicate that SD is a methodological simulation tool which is appropriate to analyse, comprehend and manage the development of complex feedback systems through causal loop and stock- and-flow diagrams (Tsolakis and Anthopoulos, 2015).

Nevertheless, as a top-down model (He et al., 2005), SD lacks the capability to reflect land use change characteristics on a micro-scale. In order to fill this gap, some researchers have begun to combine the SD method with CA models to depict the microscopic changes, such as self-organization and mutability in study regions. Several integrated SD-CA models have been proposed, with case studies utilizing such models at different regional-scales: Beijing and Shanghai metropolitan areas (He et al., 2006; Han et al., 2009), Yongding River Basin (Wang et al., 2011b) and 13 northern provinces of China (Huang et al., 2014). Simulation outputs from these studies demonstrate that SD- CA models can reflect both driving forces influencing urban expansion on a macro-scale,

and land use change processes in micro-scale. Therefore, the integration of models at various scales is considered an important trend for SD-based land use change models.

1.2.2 Markov model The Markov model, which was proposed by Cox and Miller (1965), has been recognised as the representation of spatial statistical models in research on land use change. It is constructed using the land-use distributions at the beginning and end of the simulation period and abstracted as a transition matrix. Values of matrix cells are derived from topographic maps and expressed as proportions. Under the assumption that a sample is representative of the study region, these proportional changes are taken as probabilities of land use change over the entire sample area, and form the transition matrices of the Markov model (Muller and Middleton, 1994). Another general assumption of the Markov model is stationarity: the transition matrix does not change over time (Howard et al., 1995). Figure 1.2 illustrates the workflow of land use change based on a Markov chain analysis. The Markov chain is the most commonly used type of Markov model and includes either discrete time or state space, which represents a system of elements moving from one state to another over time (Shamshad et al., 2005; Asmussen, 2008).

Figure 1.2 The workflow of a Markov land use change model, modified from Nasiri et al. (2019).

Under the above-proposed Markov chain modelling framework, López et al. (2001) explored the relationships between urban growth and landscape changes in Morelia City, Mexico at a 35-year temporal scale. In addition, Weng (2002) modelled the land use change processes of Zhujiang Delta, China from years 1989 to 1997, where urban/built- up land and horticulture farms had obviously increased while cropland decreased. Another case study utilizing the Markov chain model was conducted in Beijing, China, which indicates a notable and uneven urban growth between 1986 and 2001, accompanying the loss of cropland (Wu et al., 2006). Using a series of historical data spanning the years (1985–2005), Iacono et al. (2012) tested the Markov chain model’s ability to describe long-term land use process with simple transition probabilities.

Generally speaking, the Markov chain based land use change has been proven to describe simple trend projections for land use and land cover change. However, it is also restricted by the assumptions of relatively ideal and stable development trend of the study area (Guan et al., 2008; Sang et al., 2011), which limits its potential to reflect the study areas with dynamic and non-linear transformations.

To address the aforementioned issue, there are various studies on the integration of Markov chain and other models for the purpose of accuracy improvement. One of the most commonly applied methods is to integrate the Markov chain with CA models. Several case studies on land use development and urban sprawl have been reported: Bindura, District, Zimbabwe (Kamusoko et al., 2009); Harbin City, China (Gong et al., 2015), Jhapa District, Nepal (Rimal et al., 2017); Mumbai, India (Moghadam and Helbich, 2013); Norman City, USA (Myint and Wang, 2006); Setúbal and Sesimbra, Portugal (Araya and Cabral, 2010); Tehran, Iran (Arsanjani et al., 2013) and Tripoli metropolitan area, Libya (Al-sharif and Pradhan, 2014). According to these case studies, the integrated CA- Markov model serves as a vital tool for the understanding of spatial-temporal pattern in land use/cover and urban growth. It does not only depict the process of land use change in study areas with multiple transitions among categories (Pontius and Malanson, 2005), but also explains both the amount and constraints of urban growth (Puertas et al., 2014). Nevertheless, the projecting power of CA-Markov model will be reduced with the variation of major transformations in history. Furthermore, socio-economic (e.g. population growth and owners’ willingness) and unpredictable factors (e.g. climate change, government policy) are not accommodated into the model (Tang et al., 2007; Subedi et al., 2013), which also sets a limitation on its prediction capability.

1.2.3 Conversion of Land Use and its Effects at small regional extent (CLUE-S) model Veldkamp and Fresco (1996) introduced the CLUE model into the field of land use change modelling. The CLUE model is comprised of three modules: regional biophysical update,

regional land use objectives and local land use allocation. With the incorporation of these modules, CLUE is recognised as a tool through which the various biophysical and human land use drivers can be combined and interact in determining land use within a region. Similar to the Markov chain model, the CLUE model has been built on certain assumptions – the main one being that agriculture is the main employment and income generator in the study region. Unfortunately, the model ignores the spatial heterogeneity of different regions as well as the contribution from other natural and economic conditions and policies. These assumptions might lead to uncertain outputs during the modelling process, with subsequent impacts on its accuracy.

A modified version of CLUE was proposed by Verburg et al. (1999). This version has four interconnected modules (allocation, demand, population and yield) and two spatially explicit scales (coarse and fine). Taking China as the study region, Verburg et al. (1999) confirmed that this new version of the CLUE model can simultaneously capture the influences from both top-down and bottom-up driving factors. However, this model is still built on some ideal settings, such as the land use change could fulfil all land use requirements, which could not be produced in the final simulated output. Two years later, Verburg and Veldkamp (2001) presented a further extension of this CLUE framework, allowing the nested simulation of different crop types in addition to land-cover types, which makes it easier to communicate, discuss and compare results with economists and agronomists. Nevertheless, the implementation of this extended CLUE model acknowledges that finer scale data is required in order to achieve a deeper understanding of scalar dynamics by integration the presented (macro-level) CLUE framework with local (micro-level) studies.

Due to the above-mentioned gap between the macro-scale of CLUE model and other regional data features, Verburg et al. (2002) proposed an optimized version called the Conversion of Land Use and its Effects at Small regional extent (CLUE-S). The structure of

the CLUE-S model (Figure 1.3) clearly represents the hierarchical organization of land- use systems, allowing for a continuous iteration between regional level demands and local-level land suitability. The CLUE-S model has been widely used for the prediction of land use scenarios, with case studies conducted in Beijing City, China (Hu et al., 2013); Centre County, Pennsylvania, USA (Batisani and Yarnal, 2009); Changsha-Zhuzhou- Xiangtan urban agglomeration, China (Jiang et al., 2015); Eastern part of Netherland (Verburg and Overmars, 2007); Huai Thap Salao Watershed area, Thailand (Waiyasusri et al., 2016); Island Luzon, Philippines (Overmars and Verburg, 2007). Generally speaking, the CLUE-S model has demonstrated the capability to simulate multiple land use types simultaneously through the dynamic simulation of competition between land use types. In addition, the model gradually assigns land use change to spatial grids by iterative calculations. It displays simulation results accurately and intuitively and in the spatial location by considering both constraints and transformation rules comprehensively (Wang et al., 2010).

Figure 1.3 The workflow of a CLUE-S land use change model (Verburg and Overmars, 2007).

Recently, further expansion of the CLUE/CLUE-S models has been reported, including the integration with external approaches. To address the cross-scale interactions in land use modelling, Verburg and Overmars (2009) proposed the Dyna-CLUE model, which combines the large-scale, top-down allocation of land use change to grid cells with a bottom-up determination of conversions for specific land use transitions. Similarly, Britz et al. (2011) combined CLUE with Capri-Spat, an integrated element of the CAPRI (Common Agricultural Policy Regional Impact) model in order to obtain insights into the interactions between economic and geographic modelling. By integrating CLUE-S with SWAT (Soil and Water Assessment Tool) model, the impacts of land use planning on 11

hydrological fluxes (Zhou et al., 2013) and contamination loads (Wang et al., 2018) are quantified under different scenarios. Taking Naban River watershed reserve as the study area, a business-as-usual scenario has been provided with the combination of CLUE-S and VFHM (Village Farm Household Model)(Gibreel et al., 2014). It is illustrated that the integrated models can address land management issues since they do not only create explicit land use change scenarios at a regional scale, but also consider the socio- economic driving factors that influence land management processes. In general, a more comprehensive understanding of land use systems can be made by combining CLUE/CLUE-S model with models in other research fields. However, the linkage of different models over a wide range of scales inherently involves uncertainties and potential for error propagation (Verburg et al., 2008). Furthermore, the predictive accuracy of the CLUE-S model could be influenced by the incorporation of auto covariates due to the uncertainty of the input parameters (Jiang et al., 2015). Therefore, the above-mentioned uncertainty associated with integrated CLUE-S models need to be tested in a further step.

1.2.4 Cellular automata (CA) model According to Wolfram (1983), Cellular automata is defined as “mathematical idealizations of physical systems in which space and time are discrete, and physical quantities take on a finite set of discrete values”. As Figure 1.4 indicates, grid space, grid cell states, neighbourhood configuration, transition rules and discrete time steps are the five components of traditional raster-based CA (White and Engelen, 2000). During the past two decades, CA modelling has been put into practical use in a variety of research fields, including land use change and urban development (Xie, 1996; Wu, 1998; Batty et al., 1999; Li and Yeh, 2002a; Santé et al., 2010), transportation condition and vehicle interactions (Nagel and Schreckenberg, 1992; Barlovic et al., 1998; Jia et al., 2005), and forest fire spreading (Karafyllidis and Thanailakis, 1997; Berjak and Hearne, 2002; Karafyllidis, 2004; Encinas et al., 2007b; Encinas et al., 2007a). Among these fields, the

simulation of land use change and urban growth is the most significant research area. Couclelis (1985; 1989) first demonstrated the framework of urban CA model, after which broad-scale research on CA models arose in the 1990s. White and Engelen (1993) depicted the key fractal structures during the evolutionary processes of four US cities. Deadman (1993) modelled the rural residential development of Puslinch Township, Canada under different policy-based scenarios. Following the work of Couclelis (1985), Batty and Xie (1994; 1999; 1996) presented a general class of CA models for urban expansion, which was later evolved into the Dynamic Urban Evolutionary Model (DUEM). By integrating the Monte Carlo method with CA, Clarke (1997; 1998) produced a set of typical urban development scenarios for the San Francisco Bay and Washington areas by using his SLEUTH model. These models demonstrated CA’s ability to depict and simulate land use change initially, especially in the context of urban growth and development.

Figure 1.4 The basic structure of a CA model, modified from Lu et al. (2019).

Transition rules are of key importance in CA models, thus the definition of transition rules has received more attention than other parameters (Batty and Xie, 1994; Clarke et 13

al., 1997; Li et al., 2007). Many different calibration techniques have been proposed to determine the transition rules, including visual comparison and test (Clarke et al., 1997), analytic hierarchy process (AHP)(Wu, 1998; Mohammad et al., 2013), multiple criteria evaluation (MCE)(Wu and Webster, 1998; Mitsova et al., 2011), and logistic regression (Wu, 2002; Munshi et al., 2014). However, spatial variables are usually correlated with each other, and traditional methods (such as MCE) are not able to correct weights of these correlated variables (Li and Yeh, 2002a). On the other hand, linear transition rules cannot adequately accommodate the nonlinear characteristics of complex land and urban systems. Thus, there is an urgency for the discovery of nonlinear transition rules (Yang et al., 2008).

In order to solve the above-mentioned problems, artificial intelligence and machine learning methods have been applied in the discovery and optimization process of transition rules. In the last decade, researchers have applied a wide range of techniques for the derivation and optimization of transition rules. These include ant colony optimization (ACO)(Liu et al., 2008; Yang et al., 2012; Thilak and Amuthan, 2018), artificial immune system (AIS)(Liu et al., 2010; Curtis et al., 2016), artificial neural network (ANN)(Li and Yeh, 2002a; Almeida et al., 2008b; Omrani et al., 2017), bee colony optimization (BCO)(Yang et al., 2013; Naghibi and Delavar, 2016), cuckoo search (CS)(Cao et al., 2015), genetic algorithm (GA)(Li et al., 2008; Liu et al., 2014), particle swarm optimization (PSO)(Feng et al., 2011; Feng et al., 2018), simulated annealing (SA)(Feng and Liu, 2013), and support vector machine (SVM)(Yang et al., 2008; Huang et al., 2009; Feng et al., 2015; Mustafa et al., 2018b). It is accepted that such advanced computational methods offer a capacity of dealing with the large number of calibration parameters, resulting in a set of optimized coefficients and transition rules for improved accuracy of CA in urban modelling (Feng et al., 2011; Pinto et al., 2017; Feng, 2017). Nonetheless, the conventional raster-based CA model approach has defects in the expressive ability of geometric entities by using raster cells (grids). These defects become increasingly

obvious with the refinement of the simulation scale and the complexity of the simulation system (Pinto and Antunes, 2010; Barreira-González et al., 2015). Besides, when a geographical feature with irregular shape is being converted into raster-based cells, there will be a loss of precision. In general, the vast majority of aforementioned research is based on raster format data, and little attention has been given to derive transition rules from vector format data, another important source of GIS and CA modelling.

Neighbourhood configuration, which determines the range of influence a cell obtains from the surrounding environment, is another critical parameter of CA. It is suggested by Couclelis (1985) that the neighbourhood of a specific cell refers to a set of cells which have an impact on its transformation. In raster-based CA models, neighbour cells are usually being described through a single and fixed configuration, such as Moore or Von Neumann neighbours (Kocabas and Dragicevic, 2006). Nonetheless, the regular cell formation with single and fixed neighbourhood inhibits raster-based CA’s ability to accurately simulate geographical phenomena in the real world (Lu et al., 2015). Because of these shortcomings in raster CA modelling, some researchers have begun to take advantage of irregular polygons rather than regular cells to depict the real world (Semboloni, 2000; Flache and Hegselmann, 2001). Afterwards, the topological relation, distance adjacent functions and favourable states have been applied to the definition of neighbourhood configuration in vector CA (Benenson et al., 2002; Stevens et al., 2007; Moreno et al., 2009). Compared with raster-based CA, the vector data format reduces the sensitivity of spatial resolution by providing a landscape representation based on meaningful entities (Moreno et al., 2009; Wang and Marceau, 2013).

With the progress of vector-based CA modelling, cadastral land parcels have been widely used for the definition of cells in vector CA. In the past decade, the AIIA(Dahal and Chow, 2014), DLPS-VCA (Yao et al., 2017), iCity and iCity-3D (Stevens et al., 2007; Koziatek and Dragićević, 2017), MUGICA (Barreira-González et al., 2017), V-BUDEM (Long and Shen,

2015) and VecGCA (Moreno et al., 2009) models have been developed at the land-parcel scale, and applied to study regions in China, Canada, Portugal, Spain and USA. For the purposes of land use modelling, land parcels, cadastral lots and blocks are a better representation of the real world than raster cells, which do not directly correspond to actual geographical entities (Long and Wu, 2017). Nonetheless, the aforementioned reports on vector-based CA modelling are still relatively small in number compared to the large amount of achievements using raster-based land use change models. Considering this, utilizing advanced computational algorithms with both vector format dataset in the simulation and analysis of land use change and urban development, is an area that deserves further exploration.

Apart from current success using vector CA models, the automated subdivision of parcels and streets has been considered as a further component of land use change models (Wickramasuriya et al., 2013). As one of the pioneers in this field, Chen and Jiang (2000) proposed a computer-supported collaborative work (CSCW) land subdivision system using an event-based approach and GIS technologies. Vector-format data is the main source of land subdivision, although there do exist reports using raster-based methods (Ko et al., 2006; Morgan and O’Sullivan, 2009). Vanegas et al. (2009) developed a set of automatic algorithms for the generation of urban layouts, which was used for community visualisation, planning and policy analysis. Later, the algorithms were optimized and presented using an interactive approach in order to partition city blocks of varying areas, aspect ratios and irregularity (Vanegas et al., 2012). An automatic land subdivision tool Land Subdivision Simulator (Wickramasuriya et al., 2011) has been reported to enhance the credibility and reality of land use change models with differentiated patterns of divided parcels. On the basis of existing subdivision research, Wiseman and Patterson (2016) have made a comprehensive assessment of oriented bounding box (OBB), straight skeleton (SS) and generalized parcel divider 1 (GPD1) algorithms in a case study using Montreal, Canada. It was concluded that each algorithm

had its own advantage for the subdivision of different types of blocks, but none of them was particularly better than the others.

With the development of land subdivision research, such approaches have been integrated with land use change, especially CA models. In the parcel-based cellular automata (ParCA) model, Abolhasani et al. (2016) applied an automatic division mechanism of undeveloped space with minimum bounding rectangle (MBR) algorithm. It was confirmed that this model has the capability to create a parcel-based space for presenting the future urban growth and increases the automation of the model. The ParCA model was later updated to an asynchronous parcel-based cellular automata (AParCA) to explore the relationship between temporal scenarios, time concepts and simulation of urban growth (Abolhasani and Taleai, 2020). Using Shenzhen City, one of the largest Chinese metropolitan areas, as the study region, Yao et al. (2017) established another vector-based cellular automata model with dynamic land parcel subdivision (DLPS-VCA). In comparison with other patch-based CA models without a subdivision component, the DLPS-VCA model has achieved the highest simulation accuracy and similarity, which demonstrates the impact of reasonable land subdivision on land use change and urban expansion simulations at a fine scale. Therefore, the mechanism of land subdivision would ideally be incorporated in the current framework of CA modelling.

In addition to the aforementioned work, some researchers have also proposed novel CA models. Hewitt and Diaz-Pacheco (2017) implemented a systematic analysis on how scale-related components, specifically the cell resolution and neighbourhood configuration, affect the ability of CA models to simulate urban systems. While little impact on simulation outputs were observed by changing these parameters under a strong constraint, the authors cast doubt on its potential to explore different model scenarios with greater uncertainties. Newland et al. (2018) presented a generic and multi-objective optimisation framework which enabled the automatic calibration of CA

models with multiple dynamic land-use classes. Using Shannon's entropy (Shannon, 2001), Roodposhti et al. (2019) have introduced a dictionary of transition rules, DoTRules, to quantify the uncertainty of transition rules. This dictionary was applied for the calculation of transition rules and transition potential maps in land use change simulation with CA models. Similarly, using the maximum entropy method (Phillips et al., 2006), Wang et al. (2019) provided an algorithm to differentially control the intensity of the stochastic perturbation component, which led to an increase of both accuracy and stability in their urban CA model. In general, these works have further extended CA research, which broaden the topic and idea of land use change modelling methodologies, especially from an interdisciplinary perspective.

1.2.5 Agent-based model (ABM) An ABM (Figure 1.5) is composed of a series of heterogeneous agents, which make land use decisions on a portfolio of cells in a raster-based environment. ABM can be utilized to study the forms of organizations and interdisciplinary interactions among different levels in a deeper and more effective way (Bousquet and Le Page, 2004; Evans and Kelley, 2004). As the basic unit, an agent is defined as the object that is able to act with flexibility, which implies that agents are goal-directed and capable of interaction with other agents and a common environment (Weiss, 1999). With the development of ABM theories, a systematic overview of various correlated terms in agent-based simulation has been made, with a further category-based taxonomy: Simple and abstract, mid-range regionally or locally specific, and highly detailed models (Hare and Deadman, 2004; O'Sullivan, 2008). Besides, typologies of empirical approaches and methods have been identified by researchers for the purpose of informing and allocating agents in the study of land use change (Robinson et al., 2007; Valbuena et al., 2010; Grimm et al., 2014). Accordingly, ABM presents advantages in modelling human decision-making with rationality and heterogeneity, taking the interactions between them, and linking these micro-scale decisions to the framework of modelling dynamics of complex socio-

ecological systems at macro-scale. (Matthews et al., 2007; Sun and Müller, 2013; Elsawah et al., 2015).

Figure 1.5 The framework of an agent-based land use model (Schwarz et al., 2012).

As the pioneers of ABM-based land use change and urban development, SIMPOP (Sanders et al., 1997), STREETS (Schelhorn et al., 1999) and PEDFLOW (Kerridge et al., 2001) models were proposed to explore the evolution pattern of settlements and activities of pedestrian. On the basis of observation data, Batty et al. (2003) simulated the movement of pedestrians in specific small-scale event with a four-class ABM. Taking Altamira City, Brazil as the study area, Deadman et al. (2004) explored the land use decision making behaviours of colonist households with their Land Use Change In The Amazon (LUCITA) model. These initial applications have laid a solid foundation for the ABM models in theory and practice.

ABMs have been widely used in different research areas to model the behaviours of individuals and groups, as well as their mutual interactions with other agents and the entire system environment by assessing the possible response of on-going policies and strategies. In agricultural research, the Agricultural Policy Simulator (AgriPolis), which was originally derived from farm-based regional model (Balmann, 1997), has been used in 11 study regions (Sahrbacher et al., 2005). It is concluded that AgriPolis created a cohesive spatial framework for simulating the potential consequences of agricultural policy changes for farmers’ land use decisions and concomitant impacts on biodiversity and ecosystem services (Brady et al., 2012). In terms of ecological modelling field, both the “Overview, design, concepts and details” (ODD) protocol (Grimm et al., 2010) and “Transparent and comprehensive ecological modelling” (TRACE) documentation (Schmolke et al., 2010; Grimm et al., 2014; Schulze et al., 2017) have been developed for standardization of model parameters (Cartwright et al., 2016).

Concerning the land use change and urban development research, Le et al. (2008) proposed the Land use Dynamics Simulator (LUDAS), which contains a framework for integrating several micro dynamic models into the structures of household and landscape agents. Zhang et al. (2010) developed a multi-objective spatial optimization (MOMO) model for land use allocation based on multi-agent genetic algorithm. Schreinemachers and Berger (2011) describe the Mathematical Programming-based Multi Agent Systems (MP-MAS), which includes a comparison of eight separate simulators (both agent and non-agent based models) and explains the complex relationship of agriculture-market-environment-policy interactions. With the development of ABM methods, there is an upward trend of agent type by taking different individuals of groups into the modelling framework. One example is the Risks and Hedonics in Empirical Agent-based land market (RHEA) model, which can capture the natural hazard risks and environmental amenities of land markets through hedonic

analysis. This model integrates adaptive economic behaviour into the spatial landscape by using urban economics theory and traditional data sources (Filatova, 2015). Nevertheless, some ABMs are operated with simplified landscape and natural features, as well as the conflict between complexity and reality of agents (Magliocca et al., 2011). In addition, with its focus on the representation and realism of model structure, agent- based modelling is likely to be more vulnerable to solving problems with ontological inconsistency, which would not be solved by adjusting parameters of fitted functions (Gary Polhill et al., 2011).

In order to overcome the aforementioned shortcoming of ABM, many researchers have used hybrid modelling approaches that combine ABM with other models, especially CA in order to introduce evolutionary process between human activities and natural environment into land-use change simulations (Liu et al., 2020). Ligtenberg et al. (2001; 2004) proposed a conceptual framework by combining ABM and CA, with a case study in eastern Netherlands. The study was an initial trial of dynamic models that use both explicit spatial processes and actor interactions. With the integrated agent-based and CA models, simulated the urban growth process of Guangzhou City (Liu et al., 2006; Li and Liu, 2008), Tianjin City (Tian et al., 2016) and Changsha-Zhuzhou-Xiangtan urban agglomeration (Zeng et al., 2018), three Chinese metropolitan areas under various planning scenarios and agent settings, separately. Another case study utilized a hybrid urban expansion model (HUEM), which integrates logistic regression (Logit), CA and ABM approaches to simulate urban expansion of Wallonia, Belgium, from 1990 to 2000 (Mustafa et al., 2017). This study confirmed that the HUEM, as a representative of ABM- CA hybrid model, could capture the complexity of urban system better than traditional urban expansion. Nevertheless, it is believed by researchers that there is still potential space for hybrid models and their structures deserved to be fully explored, which enabled us to better understand the systems under study, as well as reconciled the

advantages of one modelling approach with the limitations of another (McNamara and Keeler, 2013; O’Sullivan et al., 2016).

1.3 Structure of the thesis

1.3.1 General objective The general objective of this PhD thesis is to contribute to the science of land use modelling and spatial analysis: the establishment of vector-based CA modelling framework, discovery of transition rules with different algorithms, simulation of land use development, and quantitative-qualitative combined assessment methods (Figure 1.6). It is anticipated that state government and urban planners could have a better understanding of the city futures with these simulation and prediction outcomes.

In Chapter 2, the effects of cell formats (vector and raster) are examined using artificial neural network algorithm. Afterwards, the detailed weights for all spatial factors and variables are calculated by the particle swarm optimization, with two case studies conducted in Chapters 3 (suburb-level) and 4 (city-level), separately. Assessing the simulation outcomes by means of digital reference maps and the landscape pattern indices, will be the final step in each chapter. Based on aforementioned work, detailed analysis can illustrate how the format of data and definition of transition rules influence the simulation processes of integrated vector-CA models. Finally, the integrated CA – What If? model is implemented for the purpose of scenario planning in Chapter 5. The model predicts and assesses the potential urban development pattern of Ipswich City from years 2016 to 2031. In this thesis, each chapter contains a discussion and conclusions section, and the key findings across all chapters will be summarized in Chapter 6 of the thesis.

Figure 1.6 The diagram of CA models in Chapters 2 - 5.

1.3.2 Thesis outline The titles and key findings of Chapters 2-5 are listed hereafter:

Chapter 2: Land use change simulation and analysis using a vector-based cellular automata (CA) model. In Chapter 2, vector-based and raster-based CA models are compared and analysed with derived transition rules in artificial neural network. Collingwood Park and Redbank Plains, two suburbs of Ipswich City, Queensland State have been used as the study area to 23

simulate its land use change from 1999 to 2016. It is demonstrated by the simulation outcomes that vector-based CA model could generate more authentic output on the urbanization process of the study area, which confirms the meaning of vector format data in CA theories and applications.

Chapter 3: The integration of vector-based cellular automata and partitioned rules for simulation of land use change. In Chapter 3, the vector-based CA model is integrated with two forms of transition rules (general and partitioned), which are discovered using the particle swarm optimization method. Concerning the selected study region (Bellbird Park - Brookwater and Redbank Plains, Ipswich City), general transition rules are applied to the entire study area, while partitioned transition rules are applied to each sub-region. The findings of Chapter 3 confirm that different simulation results will be produced while partitioned transition rules are applied, which improves the spatial accuracy of specific sub-region.

Chapter 4: A geographically partitioned cellular automata model for the expansion of residential areas. For the test of generality in partitioned PSO-CA modelling, the entire Ipswich City has been applied as the study area of this chapter. The entire Ipswich City is divided into two sub-regions according to administrative boundaries and population distribution. Both general and partitioned transition rules of CA models are defined on the basis of spatial variables, neighbourhood influence, as well as the South East Queensland Planning Scheme (2009 - 2031). It is concluded from Chapter 4 that partitioned transition rules could capture the dynamic relationship between variables and residential development of the study area with relatively obvious spatial heterogeneity, and significantly improve the overall accuracy of modelling results.

Chapter 5: The integration of cellular automata and What If? for scenario planning: Future expansion of residential areas in Ipswich City In Chapters 2 – 4, CA models with different methodologies have been proposed for the simulation of land use change during the past period (years 1999 to 2016). Nevertheless, people might have a stronger interest in what will heppen in the uncertain future. After identifying the regulations of historical land use change, Chapter 5 integrates the CA modelling framework with What If? planning support system (PSS) to predict its future development under the potential scenario. This integrated CA – What If? model is operated across macro and micro level scales. Specifically, the future land use demand of Ipswich City (years 2016 to 2031) is estimated in online What If? (Pettit et al., 2013), where detailed land use allocations are produced by CA model. Both current residential distribution (year 2016) and South East Queensland Planning Scheme (2009 – 2031) (Queensland Government, 2017) have been applied to assess the consistency and reasonableness of aforementioned development scenario. With the case study in this chapter, the integrated CA – What If? model is shown to be a refined tool for both land use demand prediction and allocation simulation, and is a more practical aid for the decision-making and planning aid for state government.

In this thesis, Chapter 2 is published in the journal of Environment and Planning B: Urban Analytics and City Science (Lu et al., 2019). Chapter 3 has been published in the 15th International Conference of GeoComputation as a peer-reviewed conference paper, with an oral presentation made by the author on 20 Sep 2019. Chapters 4 and 5 are prepared as journal articles. It is anticipated that they will be submitted to selected international journals in the following stage.

1.4 Research significance

The main research significance is summarized in the following.

1.4.1 A systematic exploration of vector CA modelling As described in Section 1.2, raster grids are still the dominant spatial format used to represent the real world and cross-system spatial interactions. Nevertheless, the regular grid is not in accordance with the nature of our irregular world. Furthermore, despite current achievements in vector-based CA research, there are questions to be further explored and discussed. These questions are related to the accuracy difference between vector and raster formats; the discovery of transition rules; the definition of neighbourhood configuration, as well as its complexity, sensitivity and uncertainty. In Chapters 2-4, the difference between cell forms (vector and raster), transition rules (artificial neural network and particle swarm optimization), and spatial scales (suburb and city) have been analysed in a systematic way. Therefore, the enriched vector CA theory and methodology is considered as the first contribution of this thesis.

1.4.2 The integration of CA and planning support system In recent years, GIS has been widely used in many fields where spatial scientists are required, such as land resource management and urban planning. Nevertheless, there are relatively few reports on the integration of land use change model and planning support system, which fills the gap between the modelling target and the requirements of practical work by urban planners. On the basis of What If?, one of the leading PSSs in urban planning, the framework of CA – What If? model is proposed in Chapter 5. Taking the planning schemes, population growth and potential land use development scenarios into account, this model predicts the potential land use patterns and detailed evolutionary process in a hierarchical way. As a result, the simulation results, which are obtained under a combination of diversified spatial and non-variables, are more meaningful and relevant to the Infrastructure and Planning Department of Queensland, as well as the real estate industry.

1.4.3 The comprehensive evaluation of modelling outcomes Apart from the previous points, evaluation is also an important component of land use change modelling since it provides an objective assessment of its modelling accuracy and stability. In this thesis, various evaluation indices are applied and discussed in different chapters. These indices focus on not only single simulation output but also multiple replications, by both individual comparison and overall estimation, in macro and micro spatial scales, past and future temporal scales. They demonstrate the features of simulation outputs from different aspects. Consequently, it is more beneficial for the aforementioned department and industry sectors to identify the appropriateness of simulation models and tools, with the potentiality to be implemented and optimized in a further step.

Chapter 2 Land use change simulation and analysis using a vector-based cellular automata (CA) model

This chapter has been published in the journal of Environment and Planning B: Urban Analytics and City Science as: Lu Y, Laffan S, Pettit C, et al. Land use change simulation and analysis using a vector cellular automata (CA) model: A case study of Ipswich City, Queensland, Australia. Environment and Planning B: Urban Analytics and City Science. 2019. DOI: 10.1177/2399808319830971

Contribution: Yi Lu is responsible for data collection and processing, model implementation, part of the analysis work, and the written of initial manuscript.

Abstract

The loss of accuracy in vector-raster conversion has always been an issue for land use change models, particularly for raster-based Cellular Automata models. Here we describe a vector-based cellular automata (CA) model that uses land parcels as the basic unit of analysis, and compare its results with a raster CA model. Transition rules are calibrated using an artificial neural network (ANN) and historical land use data. Using Ipswich City in Queensland, Australia as the study area, the simulation results show that the vector and raster CA models achieve 96.64% and 93.88% producer’s spatial accuracy, respectively. In addition, the vector CA model achieves a higher kappa coefficient and more consistent misclassification frequency, while also having faster processing times. Consequently, the vector-based CA model can be applied to explore regulations of land use transformation in urban growth process, and provide a better understanding of likely urban growth to inform city planners.

2.1 Introduction

Land use change modelling, as an aid to understanding city changes, has been the subject of research for more than 30 years (Couclelis, 1989; Batty and Xie, 1994; Clarke et al., 1997). The theory and application of these models have been a key topic of research in GIS, and they are important tools for the analysis of land use dynamics (Verburg et al., 2004a). Cities have been identified as complex systems with dynamic and nonlinear characteristics, and the regulations of land use evolution can be determined from cities with the help of land use change models (Batty, 2009b; Yang et al., 2016). Additionally, the dynamics of potential future development of study regions can be achieved, leading to new insights of further analysis of the land-use change process under various scenarios (Shahumyan and Moeckel, 2017; Yang et al., 2018).

There are many types of land use change models, including Agent-based models (ABM)(Li and Liu, 2008; Long and Zhang, 2015), Cellular Automata (CA)(Li and Yeh, 2002a; Almeida et al., 2008a), Conversion of Land Use and its Effects (CLUE)(Verburg and Veldkamp, 2001), and What If? (Pettit, 2005; Klosterman and Pettit, 2005b). Among these, CA models have been applied in the field of land use change and urban development (Verburg et al., 2004b; Liu et al., 2008; Feng and Tong, 2018b), and derived new models such as SLEUTH (Jantz et al., 2004; Mahiny and Clarke, 2012) and FLUS (Liu et al., 2017b; Liang et al., 2018). These models are designed to analyse the relationship between driving forces and land use change, and predict its future development. In general, CA is considered a bottom-up urban model from which emergent patterns of land use change arise as derived from “simple” transition rules, which is in accordance with the essence of complexity science: a complex system is derived from the interactions of simple subsystems (Li et al., 2007).

Generally speaking, a CA model consists of four components: cells, cell space, transition rules and the neighbourhood configuration (Li et al., 2007). A cell is the basic unit of the

model, and represents some part of the real world. The cell space comprises the set of cells in the model, and spans the extent of the study area. Transition rules are of key importance in CA since they determine whether the state of a cell will be transformed at a time step. These rules can be described by a set of functions, and in most cases are the combination of driving factors, neighbourhood influences, constraints for sustainable development of the study area, as well as random disruption (White et al., 1997; Liu et al., 2010). The neighbourhood configuration defines the surrounding environment around a cell that influences it.

Transition rules are a significant component of CA models (Li et al., 2007), and many different methods have been used to calibrate them. There is not yet consensus about which is the “best” approach, but primarily, it is possible to obtain more accurate transition rules using artificial intelligence techniques than traditional techniques (Li et al., 2010).

While progress has been made on transition rules, the choice of data format is an active area of research, specifically raster versus vector. The raster data format is still most commonly used (Liu, 2008), partly because land use data are commonly derived from satellite images, and also because the simple data structure leads to simpler algorithms. However, there are several issues with the use of raster data for modelling geographic objects in land use change models. First, the size of a raster cell has a considerable impact on simulation dynamics in terms of both land-cover area and spatial structure (Ménard and Marceau, 2005). Second, the fixed neighbourhood configuration of raster CA inhibits their ability to accurately simulate geographical processes - urban space transformations rarely follow regular geometric patterns (Barreira-González et al., 2015). Third, land parcels and not grid cells are the elemental unit of planning instruments such as planning zones and urban growth boundaries. Planning support systems such as What If? and CommunityViz (Geertman et al., 2013; Pettit et al., 2015) typically are built upon vector

data models, so they can precisely incorporate such planning instruments into their models.

The use of vector data for CA modelling is a topic of research that is receiving increasing attention. Initial research used Voronoi diagrams with randomly distributed cells (Flache and Hegselmann, 2001), while O’Sullivan (2001) proposed a graph-CA to define the cell space by vertices and edges. However, as with raster-based models, these approaches are limited by the mismatch between the data format used and the phenomenon being modelled. As a result, cadastral land parcels have been widely used for the definition of cells in vector CA. In the past decade, the DLPS-VCA and MUGICA models have been developed at the cadastral land parcel scale, with globally-distributed study regions in Canada (Southwest Alberta), China (Shenzhen City) and Spain (Madrid region)(Moreno et al., 2009; Yao et al., 2017; Barreira-González et al., 2017). Generally speaking, for the purposes of land use modelling, land parcels are a better representation of the real world than raster cells (Long and Wu, 2017).

However, most of the aforementioned research focuses on the definition of vector cells and transition rules. The relationship between simulation results and sampling strategy for the calibration data has not been fully analysed. In addition, there is still a paucity of research on vector CA in areas experiencing rapid urbanization such as many parts of the world including Australia, China, India and the United States.

Consequently, the overall objective of this study is to assess the performance of CA models using different data formats. The specific question addressed is whether vector CA are more accurate than raster CA for modelling land use changes in cities experiencing rapid growth. Both vector and raster CA models are implemented to simulate land use changes over a 17-year period for a study area in Queensland, Australia.

This paper is divided into three sections following this introduction. In Section 2, the structure of CA models, methods of calibrating and validating transition rules, and the main analysis workflow are illustrated in more detail. Next, the driving factors and the simulation experiments of the study area are described in Section 3. Finally, the simulation results and key findings of this work are analysed and discussed in Section 4 and 5.

2.2 Structure of CA models

2.2.1 The differences between vector and raster CAs As noted above, the fundamental difference between vector and raster CAs is the smallest unit of analysis, with the cells being modelled as either square (raster) or irregular polygons (vector). This leads to differences in the neighbourhood configuration between the two types of model. For both models, there are two primary types of cells: centre and neighbour. The state of centre cell at an iteration is a function of the values of itself and its neighbours (Lu et al., 2015). In raster CA, Von Neumann and Moore neighbours are most commonly used, in which the neighbouring cells are typically assigned equal or similar weights. However, with respect to vector CA the diversity of polygon shapes means that the connections between neighbouring cells differ and thus there are differing impacts on the transformation probability of a centre cell. For example, a centre polygon might share a long boundary with one adjacent polygon, and thus there is a high potential interaction, and only a very short boundary with another and thus a lower interaction potential. In this work we define vector neighbours as those intersecting a buffer surrounding the central cell.

Generally speaking, the influence 퐶푒푙푙푖 obtains is a function of the areas of the neighbour cells (polygons) that are within the neighbourhood. As Figure 2.1 illustrates, neighbour cell N1 has a higher influence on the Centre cell than other neighbour cells due to the centroid distance between itself and the centre cell. In addition, use of a

buffer means that non-adjacent polygons (those that do not share a common boundary) that are within the buffer distance can still influence a centre cell, for example neighbour cells N2 and N3 in Figure 2.1. This is similar to the idea of the dynamic neighbourhood, which has been proposed by Moreno et al. (2009) in their VecGCA model: objects A and B are neighbours if they are separated by other objects which have states that are favourable to the change of state from A to B. This also means the neighbour sets are less sensitive to details in the polygon data set’s topology, for example where neighbouring polygons are non-adjacent because they are separated by roads.

Figure 2.1 Example neighbourhood configuration of vector CA using a buffer of fixed radius.

2.2.2 Workflow of CA modelling The general modelling process for raster and vector CAs are almost identical except that the vector CA includes information about the polygons (Figure 2.2). Firstly, the driving factors are identified. Secondly, both vector and raster cells are extracted from the database underpinning the study area. Thirdly, the values of driving factors on cells in different formats are identified. In addition, transition rules of CA models are discovered on the basis of randomly selected sample data across the study area. Then, historical 33

data in both vector and raster formats are imported to generate the simulated land use maps at the end of study period. Taking three spatial indicators to assess the accuracy of models is the final step.

Figure 2.2 Workflow of CAs for city growth modelling.

2.2.3 Transition rule discovery with artificial neural networks Transition rules are traditionally developed using multiple criteria evaluation (MCE)(Wu and Webster, 1998), principal components analysis (PCA)(Li and Yeh, 2002b), and logistic regression (Wu, 2002). Yet, they lack the ability to correct for correlations among driving factors (Li and Yeh, 2002a). With the development of artificial intelligence, more advanced methods are being introduced for the discovery and optimization of transition rules, for example ant colony optimization (ACO)(Liu et al., 2008; Yang et al., 2012; Thilak 34

and Amuthan, 2018), artificial immune system (AIS)(Liu et al., 2010), artificial neural network (ANN)(Almeida et al., 2008a), bee colony optimization (BCO)(Yang et al., 2013), cuckoo search (CS)(Cao et al., 2015), differential evolution (DE)(Feng and Tong, 2018a), genetic algorithm (GA)(Liu et al., 2014), particle swarm optimization (PSO)(Feng et al., 2011), and support vector machine (SVM)(Huang et al., 2009).

In this research, ANN is applied for the definition of transition rules. An ANN consists of layers (input, hidden and output) and neurons which are analogous to the structure of human brains(Civco, 1993). The number of neurons in the input layer of ANN is equal to the number of driving factors for the CA model, so the vector ANN comprises one more neuron than the raster version since the cell shape is also considered as a factor. Significant errors have been observed in previous experiments where the size of the ANN is too small (Guan et al., 2005). Therefore, the number of neurons in the hidden layer is set as 2푗 + 2 by taking Kolmogorov’s theorem as the reference (Wang, 1994; Almeida et al., 2008a), where j is the count of neurons in the input layer.

2.2.4 Accuracy assessment Three indicators are used to evaluate the simulation performances of the CA models. The first one is the producer’s spatial accuracy (Equation 1), which is the proportion of correctly predicted new residential cells, using the land use map for the later time steps as the reference data. 퐶 푃푟표푑푢푐푒푟′푠 푠푝푎푡푖푎푙 푎푐푐푢푟푎푐푦 = (1) 푇퐶 Where 퐶 is the number of correctly predicted transformed cells in the raster CA model, and 푇퐶 is the total number of transformed cells. For the vector CA, 퐶 and 푇퐶 are the correct and total area of transformed cells. Besides, for any research such as this, it is important to run multiple realisations of a model to assess stability.

Therefore, cumulative producer’s spatial accuracy (PA) is adopted, which refers to the average of sum accuracies in the output of vector and raster CAs (Equation 2).

∑푛 푃푟표푑푢푐푒푟′푠 푠푝푎푡푖푎푙 푎푐푐푢푟푎푐푦 퐶푢푚푢푙푎푡푖푣푒 푃퐴 = 푖=1 푖 (2) 푛 Where n is the total number of simulation experiments.

The kappa coefficient (Cohen, 1960), measures inter-rater agreement for categorical items.

푝0 − 푝푒 퐾푎푝푝푎 푐표푒푓푓푖푐푖푒푛푡 = (3) 1 − 푝푒

Where 푝0 is the relative observed agreement among raters (identical to accuracy), and 푝푒 is the hypothetical probability of chance agreement, using the observed data to calculate the probabilities of each observer randomly seeing each category (Viera and Garrett, 2005).

The quantity and distribution of misclassified cells has been calculated by the misclassification frequency. Taking the similar term from previous researchers as references (Huang et al., 2009; De Pinto and Nelson, 2009; Barthold et al., 2013) “Misclassified cells” are those that are simulated as changing from attribute A to B but remain A in reality, and the misclassification frequency of a specific cell can be depicted as (Equation 4):

푁푚 푀푖푠푐푙푎푠푠푖푓푖푒푑 푓푟푒푞푢푒푛푐푦 = (4) 푁푇

Where 푁푇 is the total number of conducted simulation experiments, and 푁푚 is the number misclassified cells. The misclassification frequency, categorized into five intervals, measures the degree of simulation error in these misclassified cells.

2.3 Application

2.3.1 Study area and data processing Ipswich city is the second oldest local government area in the Brisbane-South East Queensland (SEQ) region, Australia’s third largest metropolitan region. Ipswich city is located in the western growth corridor of the SEQ region, approximately 35 km west of Brisbane, the capital city of Queensland. It comprises an area of 1,090 km2, with a population of 200,000 (Ipswich, 2017). It is projected by the South East Queensland Regional Plan (Queensland Government, 2017) that the resident population will be as much as 455,000 in 2031. Previous work by Stimson (2012) predicted that there will be an increase of dwellings in Ipswich city from 2006 to 2016, along with urban renewal and urban consolidation. It is therefore important to understand the evolution of urban growth in Ipswich City for urban planners and policy-makers to better understand the city’s future.

Two suburbs in the eastern part of Ipswich City are used as the study area: Collingwood Park and Redbank Plains (Figure 2.3). The changes in these suburbs are typical of Ipswich City over the study period. In 2016, the area of these two suburbs was 2,571 ha. The main land uses are “Intensive uses”, “Production from natural environments”, and “Conservation and natural environments”, occupying 53.16%, 22.21% and 19.96% of the study area, respectively. The remainder of the study area comprises “production from Dryland agriculture and plantations” and “Water”, accounting for 4.39% and 0.28% of the study area, respectively.

Figure 2.3 The study area is Ipswich City in Queensland, Australia. Source of satellite image: ESRI, downloaded: 06 Jan 2018.

Polygons of Collingwood Park and Redbank Plains were extracted from the local government area dataset of Queensland. This was obtained from QSpatial, a state- owned geospatial portal of Queensland (Queensland Government, 2016) which is also the source of the land use maps (1999 and 2016) and most other spatial data used in this research (Table 2.1). Specific driving factors are extracted from ABS mesh block data, Baseline roads and tracks of Queensland, Land use mapping of Queensland (1999, 2016) and Protected areas of Queensland. By integrating Digital cadastral database (DCDB) and property address data, the main land use categories are identified.

Table 2.1 Data sources used in the case study.

Dataset Year Source

ABS mesh block 2006, 2016 Australian Bureau of Statistics

Local government areas 2016 (ABS)

Baseline roads and tracks of 2017 Queensland

Digital cadastral database (DCDB) 2017 Queensland Government Land use mapping of Queensland 1999, 2016

South East Queensland Regional 2009 Plan 2009–2031

The main land use transformation between 1999 and 2016 is from grazing native vegetation and residual native cover to residential area. The area of grazing native vegetation has been decreased by 233.95 ha between 1999 and 2016, representing 76.03% of all changed land use. Residual native cover is the category with the second largest reduction, as much as 66.3 ha, or 21.55% of the entire decreased category. There is a 200-ha increase of residential area during the same period, which is as much as 64.76% of all increased land use types.

The focus of this work is the transition of land uses to residential, thus the classification is simplified into residential and non-residential. Grazing native vegetation and residual native cover are classified as “non-residential”. Specifically, 188.47 ha non-residential land is being transferred into “residential” during 1999 to 2016, and 809.77 ha non- residential land in year 1999 remain unchanged in 2016. These “non-residential” and “residential” land parcels were defined as the vector cells of our CA model. According to the land use map metadata, the revised land use mapping has been improved through

the interpretation of the most suitable imagery, such as Landsat TM, ETM+ and OLI with a 30 m resolution. Therefore, the raster cells were derived by converting the land use map into 30 m resolution grids. Mesh block data, the smallest geographical area defined by Australian Bureau of Statistics (ABS, 2006; ABS, 2016b), was applied to calculate population density (Figure 2.4).

2.3.2 Driving factors Five groups of driving factors are identified as influencing land use change for our model (Table 2.2): accessibility, neighbourhood, protected area, population growth and cell shape. Slope gradient is not used in this work, as 95.1% of the study area is less than 10 degrees (using 25 m DEM of the SEQ region) and its influence on land use development can be considered negligible. Table 2.2 Driving factors of land use change in the study area.

Type Driving factor(s)

Raster CA Vector CA

Accessibility Commercial service (Distance to) Public service Highways Secondary Roads Neighbourhood 5×5 Moore Neighbour Intersect area of centre cell buffer (60 m) Protected area Restriction on city growth Population The changed density of population within a parcel in past decade growth Cell shape Not applicable The area of vector cell

Figure 2.4 Population density of the study area in ABS meshblocks (2006 and 2016).

Accessibility. It has been confirmed by previous researchers that land use transformation is usually dependent on a series of spatial variables in terms of accessibilities or proximities (Li and Yeh, 2002a; Wang et al., 2011a). Hence, distances to commercial service, public service, highways (parts of Ipswich Motorway M2 and Centenary Highway A5) and secondary roads (State Route 61), which are derived using the Near tool in ArcMap, are the four variables of the study area.

Neighbourhood configuration. A 5×5 Moore neighbourhood has been applied for the detection of neighbour cells in raster CA. Similarly, a buffer zone-based method has been adopted to define the neighbourhood configuration of vector CA on the basis of previous research (Moreno et al., 2008; Moreno et al., 2009; Barreira-González and Barros, 2016). Specifically, the vector CA’s neighbourhood configuration is identified as the intersection of the centre cell’s buffer zone and adjacent cells, which is proportional to the density of adjacent residential area. While the number of neighbours of a cell is constant in raster 41

format, a variable number of adjacent polygons would be observed with different vector centre cells. Considering the raster cell size (30 m) and extent of Moore Neighbourhood (5×5), 60 m, which is twice the raster cell size, is also set as the radius of the vector neighbourhood. Furthermore, when it comes to the radius of vector neighbourhood, we have also compared the neighbourhood configurations of raster and vector CAs. In raster CA, the size of each non-residential cell is 30 × 30 = 900 m2, and there are 4,377 residential cells within their neighbourhood at the beginning of simulation. For vector CA, the average size of all non-residential cells is 1,711 m2 except for 4 extremely large

2 cells (larger than 400,000 m which is much larger than transformed non-residential cells in historical data), and the total area of their neighbour cells is 7,697,694 m2. Considering this, the ratio of neighbour cells’ area with centre (non-residential) cells is quite similarly: 4,377 in raster and 4,498 in vector. Therefore, we can conclude that it is reasonable to adopt a 60 m buffer as the radius of vector neighbourhood, which is equal to the radius of raster neighbours.

Protected area. The spatial extent of protected areas, which represents areas reserved for the conservation of natural and cultural values, is a restriction for urban growth. Namely, no transfer is permitted within the range of protected areas.

Population growth. The difference of population densities between years 2006 and 2016, which could reflect the variation of migration, has been accepted as a key component of urbanization process by previous researchers (Han et al., 2009). They are calculated according to the resident numbers and areas of meshblocks within the study area. Considering the linear growth trend of population in Ipswich City since the beginning of this century (idcommunity, 2016), as well as data availability, the variation of population density during the past decade is the fourth driving factor.

Cell shape. Due to the diversity of vector cell shapes, it is necessary to use their area as inputs to the calibration of vector CA model. The transfer probability remains the determining factor as to whether a vector cell could be transformed or not.

2.3.3 Model implementation In year 1999, there were 4,537 non-residential vector cells in 1999, of which 3,152 (188.47 ha) were transformed into residential by 2016. The number of transformed and stable (those polygons that remain unchanged between time periods) raster cells are 2,121 and 9,091. For the purpose of ANN training, a random sample method has been adopted here. 50% of transformed cells and stable cells were randomly selected for training. The same number of raster cells for each model run was sampled to maintain the same ratio of area as the polygons in vector CA. The sampled data were imported to MATLAB R2014a for ANN calibration, where 70%, 15% and 15% of the sample data was used for training, validation and testing, respectively. Afterwards, the weights and biases of ANN are determined, and these are used as the transition rules of our CA models.

The number of iterations (time steps) required to ensure that spatial details can be simulated by CA models is often between 100 to 200 (Yeh and Li, 2006; Cao et al., 2015). Accordingly, the number of iteration is set to 100, with each iteration corresponding to 2 months given the simulated period (1999-2016). Each iteration comprises four steps. The first step is to calculate the values of driving factors on cells, normalizing them using the linear scaling function. The calculation of transfer probability is the next step. As Equations 1 - 4 indicate, the inputs of ANN are normalized values of driving factors on cells, and the outputs of ANN is the transfer probability of each non-residential cell, which is the foundation of selection. The third step is to determine which of these non- residential cells will be transformed. When the area (vector format) or number (raster format) of selected cells reaches the pre-set threshold, which is derived from the overall change of study area during 1999-2016, the selection process is completed. In this paper,

thresholds are set as 18,847 m2 and 21 polygons in raster and vector CAs. The final step is to update the land use data of the study area in order to calculate the fields of driving factors for the following iteration. Once all iterations are finished, the output of the experiment is the simulated land use map in year 2016. The prototype system of CA models has been programmed using ArcObjects 10.3 software development kit (SDK) on the Visual Studio 2013 platform.

2.3.4 Simulation results In this chapter, the real land use map of year 2016 in vector format is used for the assessment of vector-based ANN-CA model. Meanwhile, it has also been converted to raster format with 30 m resolution, and this is used as the reference data for the raster- based ANN-CA model. A series of model realisations were run to assess the variability of the model results, ending when the cumulative mean and standard deviation of the producer’s accuracy had stabilised. Figure 2.5 illustrates the process of land use simulation produced by the raster and vector CA models for one realisation of each model type. The newly transformed residential cells between years 1999 and 2016 reveal a rapid trend of urban expansion, particularly around the middle of the study region. At the beginning of simulation, the majority of north-western and southern parts of Redbank Plains are non-residential. Similarly, residential cells in Collingwood Park are mainly located around the middle of the regional centre, with a southwest - northeast distribution. When the CA models have completed 20 iterations of the simulation, dispersed non-residential cell which were in close proximity to residential areas in Redbank Plains, as well as those situated in the southwest of residential area in Collingwood Park, have been transformed to residential cells (Figure 2.5A and Figure 2.5B). The difference between vector and raster CAs was that more newly developed residential cells, generated by raster CA, occur along the middle of Henty Drive. In vector CA, extra residential cells appeared on the south of Eagle Street in the northern part of study area. It is also indicated by the output of CA model that this trend continued when

it came to the end of 60 iterations’ simulation progress with more residential cells developed in the south parts of two suburbs, and northeast part of Redbank Plains (Figure 2.5C and Figure 2.5D). Therefore, the new residential cells in vector CA were more clustered in the southern part of the study area. At the end of the simulation experiment, a greater number of cells between Rhondda Road Reserve and residential area in Collingwood Park were transformed from non-residential to residential, which was approximately equal to half of the residential cells in year 1999. In addition, new residential cells in the northeast part of Redbank Plains were detected, which connected the two-separated residential area into a whole. Besides, additional residential cells were observed in the south direction of 1999 residential area (Figure 2.5E and Figure 2.5F), mainly along Alawoona Street, Cedar Road, School Road and Halletts Road, the main roads of Redbank Plains in north-south and east-west directions. It was also demonstrated by the comparison of simulation outputs that the difference can still be detected at the edge of residential areas.

The cumulative producer’s spatial accuracies (PA) of the model realisations had stabilised by thirty realisations at 96.64% and 93.88% for the vector and raster CA models, respectively (Table 2.3 and Figure 2.6). Additionally, the average standard deviation of vector CA’s accuracy is 0.848, substantially less than for the raster CA which is 1.935. The cumulative kappa coefficient of vector and raster CA models are 0.960 and 0.925 after the 30 realisations.

Figure 2.5 Simulation process of land evolution using raster and vector CAs. Locations are labelled in Figure 2.3. Source of satellite image: ESRI, imagery downloaded: 18 April 2018.

Table 2.3 Overall error statistics.

Producer’s spatial accuracy (PA) Kappa coefficient

CA format Cumulative value Standard Cumulative value Standard (%) deviation (%) deviation

Raster 93.88 1.935 0.925 0.024

Vector 96.64 0.848 0.960 0.010

* The numeric precision used in this and subsequent tables is for consistency with other research publications in the field of land use change and CA modelling.

Cumulative producer’s spatial accuracy (%) Raster CA Vector CA

98.0 97.5 97.0 96.5 % 96.0 95.5 95.0 94.5 94.0 93.5 93.0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 Number of simulations

Figure 2.6 Cumulative producer’s spatial accuracy of CA models.

The misclassification frequency for each cell across the 30 realisations are divided into five categories in Table 2.4 and Figure 2.7: low, medium-low, medium, medium-high, and high. These correspond to the value ranges [1, 6], [7, 12], [13, 18], [19, 24] and [25, 30], respectively. For instance, if a cell has been identified as “misclassified” in 5 out of 30 simulation experiments, it will be classified as low frequency.

Table 2.4 Misclassification frequency (%).

CA format Low Medium-low Medium Medium-high High

Raster 66.79 13.53 7.38 6.15 6.15 Vector 61.37 10.57 3.46 8.84 15.76

Figure 2.7 Distribution of misclassified cells in raster (A) and vector (B) CA models. Source of base map: ESRI.

2.4 Discussion

The simulation results show that the vector CA approach has higher producer’s accuracy and kappa coefficient values than the raster CA. In comparison with traditional and widely used raster-based CA models, the vector CA model has reduced the loss of accuracy during data conversion. Specifically, the irregular geographical features can be

applied for CA modelling directly, rather than being converted to regularly spaced cells with fixed size. It is close to the nature of land use change in the real world. While the raster CA values are also high, there is no overlap between the two sets of accuracies. The producer’s spatial accuracy of the vector CA model has values ranging between 97.07 and 96.42, whereas the maximum and minimum values for the raster CA are 94.66 and 93.76. In addition, in all 30 realisations, the vector CA always has a higher producer’s spatial accuracy, and the differences between vector and raster CAs are in the range of [1.91, 3.32]. A similar situation occurs with the kappa coefficient, for which the maximum and minimum values for the vector CA are 0.98 and 0.94, and 0.96 and 0.88 for the raster CA. Therefore, the outputs of the vector CA are more accurate than the raster CA. This is in agreement with the findings of other researchers (Moreno et al., 2009; Wang and Marceau, 2013; Barreira-González et al., 2015).

There is also an obvious difference between the misclassification frequencies in the vector and raster CAs (Table 2.4). For the raster CA, 66.79% of misclassified cells are misclassified with a low frequency. Furthermore, 13.53% and 7.38% of misclassified cells belong to medium-low and medium misclassification frequency. The remainder of the misclassified cells can be categorized as medium-high and high. Regarding the misclassified cells from vector CA, low misclassification frequency is also most common, accounting for 61.37% of all incorrectly classified cell area. Nevertheless, high misclassification frequency ranks the second in the output of vector CA, which is 15.76%. The ratio of medium-low and medium-high misclassified cells is the third and fourth largest category (10.57% and 8.84%). Finally, the remaining 3.46% of misclassified cells have a medium misclassification frequency.

The spatial distribution of misclassified cells provides useful information. The misclassified raster CA cells are distributed across the study region, with three obvious groups: the west edge of Collingwood Park, the east side of Collingwood Drive, and areas

located to the south of Halletts Road. Furthermore, other misclassified cells are dispersed in the northwest, and east parts of Redbank Plains. Considering the misclassified cells that were generated by vector CA, their distribution is relatively concentrated, with more than 80% located in the west edge and middle of Collingwood Park, mainly around the residential area in year 1999. Generally speaking, the distribution of misclassified cells produced by the vector CA is more concentrated, which confirms that it has a better capability of error control and would be more beneficial for the work of model calibration.

It is also worth noting that the vector CA has faster processing times. For the analyses described here, the vector CA processing time was 24.3% that of the raster CA, while also resulting in a higher accuracy. This time difference is largely related to the number of spatial units in the simulations, and the average size of the polygons used in these analyses was approximately 2.5 times the raster cell size. It is worth considering how this applies to finer resolution data sets. If higher resolution satellite images are used to generate the land use maps then there would be a quadratic scaling in the number of grid cells and thus processing times, but one would not expect a substantial change in the number of land use polygons, thus the differences of processing times will diverge substantially as spatial resolutions increase.

What is more, metadata indicates that the 2017 land use map is derived from the 1999 baseline of the same land use dataset. There are only limited changes in vector cell shapes across the study area during the simulation period. The mechanism of land parcel subdivision has not been widely used in the literature, and a single set of cadastral data remains the dominant method for vector cell definition in CA modelling (Dahal and Chow, 2014; Barreira-González et al., 2015; Li et al., 2017c). While it is not included in this version of the ANN-CA model, the ability to simulate further land subdivision and

rezoning is an important topic for future research to more fully land use transformation processes.

2.5 Conclusions

In this research, the structure, parameters and workflow of vector and raster CAs have been compared, with a summarization of their similarities and differences. Collingwood Park and Redbank Plains, two suburbs of Ipswich City, Queensland, Australia, were used as a case study. Between 1999 and 2016, a 35.15% increase of the residential land occurred in the study area, much in a southerly direction, but also near existing residential area from 1999. We can conclude from the results of 30 realisations that the cumulative producer’s spatial accuracy of vector-based CA is 2.76% higher than raster- based CA. Additionally, the cumulative kappa coefficient of vector CA is 0.035 higher than that of the raster CA. Besides, the misclassified cells produced by vector CA has a higher spatial concentration. It is demonstrated that the vector CA model can not only produce a more accurate result in both macro and micro scales, but also narrow down the spatial distribution of misclassified outputs, which lay a solid foundation of the following work on model calibration and optimization.

The simulation and analysis results confirm that the vector-based CA model generates a more authentic output on urbanization process of the study area. Considering this, vector-based CA model could be an applicable tool to analyse the regulations and patterns of urban growth and aid the decision making of city planners. Therefore, the research of CA theories and applications on vector format is necessary and meaningful.

Chapter 3 The integration of vector-based cellular automata and partitioned rules for simulation of land use change

This chapter has been published in the 15th International Conference of GeoComputation as: Lu Y, Laffan S. The integration of vector-based cellular automata and partitioned rules for simulation of land use change. 15th International Conference on GeoComputation, Queenstown, New Zealand, Sep 18-21, 2019.

Contribution: Yi Lu is responsible for data collection and processing, model implementation, part of the analysis work, and the written of initial manuscript.

Abstract

Cellular Automata (CA) have formed an important part of the geocomputational and spatial analysis toolbox for three decades. Some of the most important components of CA are the transition rules to determine the changes in cell states across iterations. These are applied at the micro-sale, but lead to emergent patterns at the macro-scale. An important consideration for transition rules is the spatial extent over which they will be valid. One does not expect the same rules to apply equally across an entire region, yet most CA implementations only support one set of transition rules that are applied everywhere. In this paper, a vector CA model with spatially partitioned transition rules is proposed to identify the expansion of urban residential areas across heterogeneous study area. Initial experiments using two sub-regions of Ipswich, Queensland, Australia, indicate that the spatially partitioned approach can improve the accuracy of vector CA.

3.1 Introduction

CA models have been employed in the exploration of a wide variety of urban phenomena, from traffic simulation and regional-scale urbanization to land-use dynamics, polycentricity, historical urbanization, and urban development (Torrens and O'Sullivan, 2001). For spatial scientists and urban planners, there is also an urgent need to predict future developments and land use change in an understandable way. Numerous computer-based models have been developed to address these issues, including CA (Cellular automata)(Wu and Webster, 1998; Li and Yeh, 2002a; Liu et al., 2007), CLUE/CLUE-S (Conversion of Land use and its Effects at small region extent)(Verburg et al., 1999; Verburg and Overmars, 2007), MAS (Multi-agent system)(Heppenstall et al., 2011; O’Sullivan et al., 2016), SD (System dynamic)(He et al., 2006; Xu and Coors, 2012) and What If?(Pettit, 2005; Pettit et al., 2015).

Among all land use change models, CA and its extension models have been widely applied due to their capability of modelling complex spatial dynamics on the basis of a set of “simplified” transition rules (White et al., 1997). Abundant results have been achieved in the field of CA modelling, which can be classified into four groups: cell format (Flache and Hegselmann, 2001; Moreno et al., 2008), transition rules (Li et al., 2014; Almeida et al., 2008a), neighbourhood configuration (Moreno et al., 2009), sensitivity and uncertainties (Kocabas and Dragicevic, 2006; Şalap-Ayça et al., 2018). While transition rules have attracted more attention than other parameters, there is still a key problem to be discussed and solved: Does a single set of transition rules contain enough information for all sub-regions of the study area?

Here we propose a spatially partitioned CA model to address the above question. Instead of using one set of general transition rules across the entire study region, the transformations of cells are determined by spatially local rules, partitioned by sub- regions. Using a PACA (partitioned and asynchronous cellular automata) model, Ke et al.

(2016) simulated the process of urban growth during 2005 to 2013. Taking Yangtze River middle reaches megalopolis (YRMRM) as the study area, it is also indicated by Xia et al. (2019) that the development of partitioned transition rules for sub-regions can greatly improve both the overall and local accuracies of CA model. Concerning the aforementioned reports, it has been indicated that the spatial heterogeneity of urban growth can be better represented by using differential transition rules for partitioned zones. Additionally, these previous approaches used raster-based CA models, while vector CA models have the potential to improve land use change modelling, particularly in urban environments. The aim of this research is therefore to assess the effectiveness of a spatially partitioned vector CA model for land use change modelling.

3.2 Methodology

3.2.1 Using PSO for the discovery of transition rules In this research, particle swarm optimization (PSO) is utilized for the calibration of transition rules. PSO is a useful approach for the discovery of transition rules as it captures the complex non-linear processes of urban land use change (Feng et al., 2011) and deals well with the large number of calibration parameters (Pinto et al., 2017). In comparison with the previous ANN method, the PSO results are easier to understand with weights of all independent variables (driving factors).

Similar to the relationship between cell and CA, particle is the smallest unit of PSO, it equals to one potential solution of the target problem, and is comprised of two parts: velocities and positions

푃푎푟푡푖푐푙푒 = (푣푛, 푃푛) (1)

Where n is the dimension of target problem, 푣푛 and 푃푛 are the velocity and position of corresponding particle at a specific time point. They can be represented by n velocities and positions at time t:

푣 = (푣 , 푣 , … , 푣 , 푡) { 푛 1 2 푛 (2) 푃푛 = (푃1, 푃2, … , 푃푛, 푡)

The combination of velocity and position in each particle are updated according to individual and global best positions:

푣(푡 + 1) = 푤 ∗ 푣(푡) + 푐1 ∗ (푃 − 푃(푡)) + 푐2 ∗ (푃 − 푃(푡)) { 푖푏 푔푏 (3) 푃(푡 + 1) = 푃(푡) + 푣(푡 + 1)

Where 푤 is the weight of velocity, 푐1 and 푐2 are individual and global learning factors. 푃푖푏 is the best individual position of particle i, and 푃푔푏 is the best global position of all particles, namely the best one of all best individual positions. In addition, 푣(푡 + 1) is the velocity of a particle at time t+1, P(푡) and 푃(푡 + 1) are the positions of particle at time t and t+1, accordingly. Specifically, the nature of an individual particle can be taken as the combination of weights, and the space they explore is the potential weights of all driving factors. In Equation 2, 푣푛 determines how each weight value changes in the next step time t+1, and 푃푛 represents the values of weights for corresponding driving variables at time t. After comparing the existing and all previous positions during the optimization process, the one that could generate highest output value from predefined formula is regarded as the “best” position. Namely, when an individual particle has found its “best” (optimal) position, the simulated land use distribution from this group of weight combination will be produced with the highest accuracy.

The transfer probability P(Celli) of a single cell i can be calculated by Equation 4: 1 ( ) 푃 퐶푒푙푙푖 = 푖=푛 ∗ (푏0 + 푛표푟푁푒푖) (4) 1 + exp[−(푎0 + ∑푖=1 푓푖 ∗ 푤푖)]

Where 푃(퐶푒푙푙) is the probability of 퐶푒푙푙푖, 푎0 and 푏0 are two constants, 푓푖 푎푛푑 푤푖 are the normalized values of driving factors and corresponding weights, norNei is the normalized neighbourhood configuration. In this study, 푤푖 are derived from PSO method, and n is the number of driving factors. 55

3.2.2 Validation method

Two indices are used here to validate the performances of PSO-CA models: Cumulative producer’s spatial accuracy and misclassification frequency.

1. Cumulative producer’s spatial accuracy (CPSA) Producer’s spatial accuracy (PSA) has been widely used for the assessment of precision in the research field of land use modelling, which can be described as:

퐴푟푒푎푐표푟 푃푆퐴 = (5) 퐴푟푒푎푎푙푙

In Equation 5, the PSA of vector CA can be calculated by 푟푒푎푐표푟 and 퐴푟푒푎푎푙푙 , the correctly simulated and total area of cells.

Separate experiments are proposed in both the entire and sub-regions of the study area. Therefore, CPSA is defined as the mean value of PSA from all experiments:

푁푢푚 ∑ 푠푖푚 푃푆퐴 (6) 퐶푃푆퐴 = 푖=1 푁푢푚푠푖푚

Where 푁푢푚푠푖푚 refers to the number of simulation experiments that conducted under each combination of CA parameters.

2. Misclassification frequency (MF) Except for CPSA, the misclassification frequencies of all simulations, is used to assess the frequency a cell has been misclassified across the set of simulations (Huang et al., 2009; De Pinto and Nelson, 2009; Barthold et al., 2013):

푁푖푛푐표푟 푀퐹 = (7) 푁푎푙푙

Where 푁푖푛푐표푟 and 푁푎푙푙 records the number of incorrectly classified and total simulation counts.

3.2.3 Model implementation The implementation of both general and partitioned PSO-CAs can be summarized with three steps (Figure 3.1). At the beginning of simulation, the entire study area is divided into two sub-regions (administrative boundaries for this research). Sample data, which are required for PSO training, are then randomly selected from the spatial datasets. Afterwards, two separate PSOs, which represent the sets of general and partitioned transition rules, will be used for training, with the output of weights for different driving factors. After the completion of training, the simulated distribution of residential cells will be produced by the CA models with general and partitioned rules (general and partitioned PSO-CA). Finally, taking the land use map at the end of simulation period as reference, we evaluate the accuracy of both general and partitioned CAs.

Figure 3.1 Diagram of partitioned PSO-CA.

3.3 Case study

3.3.1 Study area and data processing Ipswich City is the second oldest local government area (LGA) of Brisbane-South East Queensland (SEQ) region. It is located approximately 35 km west of Brisbane, the capital city of Queensland, Australia. For the purpose of testing the effect of partitioned CA models in vector data format, two suburbs within Ipswich: Bellbird Park - Brookwater and Redbank Plains (Figure 3.2) are selected as the study area. These suburbs represent the typical trend of land use across Ipswich from 1999 to 2016. In 2016, the area of these two suburbs was 3,543.36 ha. The main land uses are “Intensive uses”, “Conservation and natural environments” and “Production from natural environments”, occupying 55.48%, 20.26% and 19.16% of the study area, respectively. The remainder 4.89% and 0.21% of the study area are classified as “Production from dryland agriculture and plantations” and “Water”. Polygons of the study area were extracted from the LGA dataset of Queensland. This was obtained from QSpatial, a state-owned geospatial portal of Queensland (Queensland Government, 2016) which is also the source of the land use maps (1999 and 2016).

Figure 3.2 The study area in Ipswich, Queensland, Australia.

Of the land converted to residential by 2016, 92.57% were from secondary land use classes “Grazing native vegetation” and “Other minimal use”, "Services (4.65%)" and "Land in transition (2.78%)" are the remaining sources. Specifically, the area of “Other minimal use” had been decreased by 450.73 ha between 1999 and 2016, representing 61.04% of all reduced land use. Besides, “Grazing native vegetation” suffers a 272.31 ha reduction, which is as much as 36.88% of the entire decreased category. Two land uses were excluded from the simulation: “Services” lands (mainly Schools and education

institutions), are managed under local state regional planning (Queensland Government, 2017), while “Land in transition” refers to the unknown land use. For analysis, “Other minimal use” and “Grazing native vegetation” are classified as “non-residential”, while “Intensive uses” is considered as “Residential”. Here we focus on the transformation of parcels from non-residential to residential, as this is the most common land use change in such environments (Seto and Shepherd, 2009; Fragkias et al., 2013). The transformed (from non-residential to residential) and stable (remain non-residential during study period) cells are summarized in Table 3.1. Table 3.1 The transformed and stable cells of study area during 1999 to 2016.

Transformed Stable

Count Area (ha) Count Area (ha)

Entire study 4,228 286.86 2,640 1,678.44 area

Bellbird Park - 2,091 157.25 1,117 798.64 Brookwater

Redbank Plains 2,137 129.61 1,523 879.80

3.3.2 Simulation process and result Each simulation experiment was run for 100 iterations in order to reproduce the subtle patterns of land use change (Cao et al., 2015). In general, one iteration can be summarized as “calculate, select, and update”. Calculate: at the beginning of an iteration, the transfer probabilities are calculated and assigned to non-residential layer. Select: newly transformed cells, where their attributes are converted to residential, are identified from the non-residential layer with relatively high transfer probabilities. Update: these selected cells are merged into the residential layer, and used to update the non-residential layer before next iteration. The main difference between general and partitioned CA is that the non-residential layer in partitioned CA will be updated by the

residential layers in partitioned sub-regions in order to avoid edge effects on its simulation outputs.

Two model configurations were applied. In the first configuration, a single set of transition rules was calibrated and applied using the full study area. In the second, two sets of transition rules were calibrated, one for each sub-region (Bellbird Park - Brookwater, Redbank Plains). In every experiment, 20% of the sample cells (including both transformed and stable), which achieved a balance between information richness and over-training, were randomly selected to calibrate the transition rules. On the basis of initial experiments, along with existing studies on calibration of transition rule (Feng et al., 2011; Liao et al., 2014), the weight of velocity v, individual and global learning factors c1 and c2 (in Equation 3) were set as 1, 1.5 and 1.5, respectively. Similarly, two constants a0 and b0 were both set as 1. The initial position and velocity of each particle was generated with random values in the intervals [-5, 5] and [-2, 2], and the maximum particle velocity was restricted to 1 to reduce unrealistic results due to an extreme velocity. Following Harrison et al. (2019), 1000 iterations were used for PSO training to ensure all particles were well-trained. Table 3.2 Comparison of average weights.

Study area Discom Discen Dispub Slope Disroad Dissta Dens Areacell Bellbird Park - -78.61 -80.21 46.22 -14.32 66.48 -4.70 125.64 -36.06 Brookwater Redbank Plains -8.88 -36.64 30.00 7.58 -27.39 -7.93 89.87 -22.75 Entire study -72.23 -22.94 -2.25 -17.51 35.07 -23.25 165.06 -57.50 area

According to Table 3.2, the importance of factors is illustrated by the absolute values of their weights. Positive and negative values indicate whether a relatively larger or smaller value makes a greater contribution to the conversion from non-residential to residential. Specifically, the population density in 2016 (Dens) is the most important factor for the

transformation from non-residential to residential in sub-regions and the entire study area. Nevertheless, the difference between general and partitioned CAs can be observed from the factor of 2nd largest contribution, which is identified as distance to district centre (Discen) in two sub-regions, as well as distance to commercial (Discom) in entire

rd study area by general transition rules. Similarly, Discom is also the factor with 3 largest contribution in Bellbird Park – Brookwater. Considering this, it is confirmed by the difference of weights that heterogeneity existed between sub-regions and entire study area, which leads to a further diversity of simulation results in the following. After obtaining the weights of general and partitioned CA, all simulation experiments are implemented with 100 iterations. To ensure stability of the solutions, 30 separate model runs were applied.

Figure 3.3 Simulation processes in entire/sub region(s).

Figure 3.3 illustrates the simulation processes of residential extension by general and partitioned PSO-CA. After the completion of 20 iterations, clustered new residential cells have been observed in southern part of previous residential area (Figure 3.3A) and along the Augusta Parkway of Bellbird Park - Brookwater (Figure 3.3B). As the simulation progresses, the new residential areas were further expanded from the previous time steps (Figure 3.3C and Figure 3.3D). At the end of the simulation experiment, new residential cells are evident on both sides of State Road 61 and Augusta Parkway, the main roads that bisect the sub-regions. The remaining scattered non-residential cells were distributed in western and north-eastern parts of Redbank Plains (Figure 3.3E and Figure 3.3F). Taking the real land use data as reference, the CPSA and MF of all simulation experiments are recorded in Table 3.3 and Figure 3.4.

Distribution of misclassified cells 100 90 80 70 60 50 40 30 20 10 0 Low Medium low Medium Medium high High

General PSO-CA (BB) General PSO-CA (RP) Partitioned PSO-CA (BB) Partitioned PSO-CA (RP)

Figure 3.4 Ratios of misclassified cells with different frequencies (%). * BB: Bellbird Park – Brookwater, RP Redbank Plains

Table 3.3 The PSA and CPSA values of sub and entire study areas.

Correct PSA CPSA Standard simulated area value value deviation (ha) (%) (%) (%)

1 123.36 78.45

2 131.32 83.51 Bellbird Park – ... 79.66 4.67 Brookwater 29 110.91 70.53

General 30 122.31 77.78

PSO-CA 1 119.46 92.17

2 116.53 89.91 Redbank … 90.45 1.41 Plains 29 116.58 89.94

30 114.00 87.95

1 134.25 85.37

2 135.66 86.27 Bellbird Park – ... 82.99 1.97 Brookwater 29 129.94 82.63

Partitioned 30 129.44 82.31 PSO-CA 1 106.45 82.13

2 107.84 83.20 Redbank 85.44 2.79 ... Plains 29 111.24 85.53

30 113.90 87.88

3.4 Discussion and conclusions

In this paper, vector PSO-CA models with both general and partitioned transition rules have been proposed for the simulation of residential expansion in Ipswich City, Queensland, Australia during 1999 to 2016. According to the comparison between outputs and reference map, the CPSA values of two sub-regions are 79.66% and 90.45% by general PSO-CA, as well as 82.99% and 85.44% in partitioned PSO-CA (Table 3.3). Specifically, while the difference between maximum and minimum PSA values is 25.03% for the general PSO-CA (92.48% and 67.45%), it is 13.07% (90.54% and 77.47%) for the partitioned PSO-CA. In addition, standard deviations of the simulation results are 4.67% (Bellbird Park – Brookwater, general PSO-CA), 1.41% (Redbank Plains, general PSO-CA), 1.97% (Bellbird Park – Brookwater, partitioned PSO-CA), and 2.79% (Redbank Plains, partitioned PSO-CA). Therefore, it can be concluded that general PSO-CA is more accurate in one sub-region while partitioned PSO-CA obtained a higher CPSA in another sub-region. Furthermore, the produced PSA values by partitioned PSO-CA are more stable.

Five levels of misclassification frequency have been selected, corresponding to ranges [1, 6] (Low), [7, 12] (Medium low), [13, 18] (Medium), [19, 24] (Medium high) and [25, 30] (High). As Figure 3.4 reveals, both general and partitioned PSO-CAs have demonstrated a “drop-rise” trend in misclassification frequency. For general PSO-CA, misclassified cells with low frequencies are the largest part, which occupied 43.61% and 40.64% in two sub-regions. Besides, 27.44% and 35.42% of the misclassified cells belong to high frequency, the second largest group in corresponding suburbs. The similar trend can also be detected in Redbank Plains for which partitioned PSO-CA produced 88.78% and 7.29% (Redbank Plains) misclassified cells with low and high frequencies. However, Bellbird Park – Brookwater is an exception where the misclassification frequency rates keep decreasing with the rise of frequency level. Overall, general transition rules make relatively spatially concentrated errors while the results produced by partitioned 65

transition rules are more diverse, which corresponds to a broader range of potential solutions.

According to this case study, it is confirmed that different simulation results will be generated while vector PSO-CA models are integrated with partitioned transition rules. It is difficult to distinguish which one is better at this stage: partitioned rules improve the spatial accuracy in one sub-region (Bellbird Park – Brookwater) while general rules obtain a more accurate output in another sub-region (Redbank Plains). Future work will assess the effects across a broader study region, including the integration of both general and partitioned transition rules on a per-subregion basis.

Chapter 4 A geographically partitioned cellular automata model for the expansion of residential areas

This chapter is prepared as a manuscript for submission to an international journal in related field: Lu Y, Laffan S, Pettit C. A geographically partitioned cellular automata model for the expansion of residential areas. Selected international journal, In Preparation.

Abstract

Many approaches and applications of cellular automata (CA) models have been reported, especially in the fields of land use change modelling and urban growth simulation. A key component of all these models is the transition rules, as they determine the transformation of cells at each iteration. However, the majority of previous studies use only one set of transition rules to reflect the regulations of land use change across the entire study region, and do not fully account for spatial heterogeneity. In this research, a vector-CA model has been implemented that calibrates transition rules for spatially partitioned subsets. The particle swarm optimisation (PSO) method is used for the calibration process. The changes in urban extent were modelled for a study region in Ipswich City, Australia, from the years 1999 to 2016. The results confirm that the spatially partitioned rules can generate more accurate and stable results compared with the rules calibrated using the whole study area, with a 2.13% increase of cumulative producer’s spatial accuracy (CPSA). The implementation of CA models with partitioned transition rules enables a better understanding of spatial heterogeneity in land use change.

4.1 Introduction

Since the integration of CA models with Geographic Information System (GIS) and Remote Sensing (RS) techniques in the 1990s (Couclelis, 1989; Batty and Xie, 1994; Clarke et al., 1997), this bottom-up modelling approach has been the subject of considerable research, especially in the fields of land use change simulation and urban growth analysis (Xie, 1996; Yang and Liu, 2005; Li et al., 2008; Chaudhuri and Clarke, 2013; Liu et al., 2017b). It has been projected that 70% of the world’s population will live in urban areas by 2050 under the current trend of global urbanization (United Nations, 2018), with a 200% increase of urban land cover from 2000 to 2030 (Fragkias et al., 2013). Considering this, research on CA models, which could provide better understanding of urbanization processes and related factors, is an important subject for spatial scientists and analysts.

The majority of current CA models are implemented using the raster data structure. It is acknowledged that fine resolution raster data can represent land use polygons (Moser et al., 2012), but only in aggregate form, and thus each raster cell will often have neighbouring cells that are within the same polygon. The inevitable loss of details caused by raster-based CA has led some researchers to develop vector-based CA (Semboloni, 2000; Benenson et al., 2002). For example, Moreno et al. (2009) introduced a VecGCA model with a dynamic neighbourhood, which judged the neighbour cells according to both their locations and attributes. Similarly, a patch-based CA has been developed by Wang and Marceau (2013) by taking the homogeneity of adjacent cells into consideration. Li (2017b) proposed a new edition of patch-based urban growth model with a heuristic watershed segmentation algorithm. Taking Shenzhen China as the study area, Yao (2017) built a dynamic land parcel subdivision (DLPS) CA to divide land parcels, which improves the simulation accuracy of vector-based CA at finer scales. These works have been implemented within the framework of vector-based CA models, which reduces the uncertainties of neighbourhood configuration in traditional raster-based CA.

Regardless of the spatial structure of CA models, transition rules are perhaps the most significant component. Transition rules determine the state of each cell at each iteration, leading to a model of the dynamic and macro-scale variation of land use across entire study region. Researchers have developed a range of methods to calibrate transition rules on the basis of artificial intelligence techniques such as artificial neural network (ANN)(Li and Yeh, 2002a; Almeida et al., 2008a; Omrani et al., 2017), ant colony optimization (ACO)(Liu et al., 2008; Yang et al., 2012; Thilak and Amuthan, 2018), bee colony optimization (BCO)(Yang et al., 2013; Naghibi and Delavar, 2016), cuckoo search (CS)(Cao et al., 2015), genetic algorithm (GA)(Hagenauer and Helbich, 2012; Mustafa et al., 2018a), particle swarm optimization (PSO)(Naghibi and Delavar, 2016; Feng et al., 2018). In comparison with traditional linear-based equations, these methods construct the framework of integrated solutions for advanced urban land dynamics simulation, comprising components (spatial and non-spatial dynamics entities), architecture (high level of integration), performance improvement (optimization solutions, interoperability, spatial-temporal synchronization), and evolutionary view (spatial DNA)(Wu and Silva, 2010). It leads to better representations of complex patterns of land use change at different spatial and temporal scales.

Among all methodologies for transition rules discovery, particle swarm optimisation (PSO) is a widely used evolutionary algorithm. It was initially proposed by Eberhart and Kennedy (1995), who abstracted the solution of a problem as the search of a target position by using the information sharing mechanism between a group of particles. The PSO method was later updated by introducing inertia weights, which improved the efficient search for a globally optimal solution (Shi and Eberhart, 1998). PSO is considered as a highly efficient global search algorithm as it is derivative-free and insensitive to the scaling of design variables (Schutte et al., 2003). Therefore, the combination of PSO with CA models have been adopted universally: particle swarm

cellular automata (PS-CA) in Condeixa-a-Nov, Portugal (Pinto and Antunes, 2010); multi- objective PSO (MOPSO) in Tehran, Iran (Masoomi et al., 2013); particle swarm optimization-based CA (PSO-CA) in UA-Shanghai region, China (Feng et al., 2018); neighbour decay cellular automata based on particle swarm optimization (PSO-NDCA) in Xiamen, China (Liao et al., 2014); and vector-based particle swarm optimization cellular automata (vector PSO-CA) in Ipswich, Australia (Lu and Laffan, 2018). In general, it is demonstrated that PSO is a feasible approach to capture the complex non-linear processes of urban land use change (Feng et al., 2011) and deals well with the large number of calibration parameters in a succinct structure (Ma et al., 2011; Pinto et al., 2017). Therefore, it is highly suitable for the calibration of transition rules, which reflects the complexity of urban systems. However, in spite of their advantages in the field of land use change simulation, the great majority of CA models have neglected an important aspect of land use change modelling: the spatial heterogeneity (Figure 4.1) of study regions in terms of land use categories, geographical entities, population density and their spatial distribution.

Figure 4.1 Spatial heterogeneity of a sample study region.

One does not expect the same set of transition rules to apply across a large region, as there might be substantial differences in land use policies or terrain that change the interactions of spatial variables. These kinds of substantial differences are defined as “Spatial heterogeneity”, which refers to the unevenly distributed land use categories, especially across large regions (Silva et al., 2006). The most obvious means of addressing this issue is to use spatially partitioned transition rules. For example, Ke et al. (2016) divided the study area into two sub-regions with k-means and knn-cluster, and discovered transition rules using the C5.0 decision tree algorithm in their partitioned and asynchronous cellular automata (PACA) model. Another partitioned particle swarm optimization - cellular automata (PSO-CA) model has been proposed at both regional, meso and city scales by taking terrain feature and administrative boundaries into consideration (Feng et al., 2018). Additionally, Xia (2019) integrated partitioned logistic- 71

CA with an improved gravitational field model for the simulation of historical urban evolution. The above-mentioned partitioned CA models have been implemented with case studies in Wuhan City, Shanghai City and surrounding areas (UA-Shanghai), as well as Yangtze River middle reaches megalopolis (YRMRM) of China, separately. It has been confirmed that CA models could better reflect the spatial variability of sub regions, with a further improvement of simulation capability, especially for small- and medium-sized cities. Nevertheless, these partitioned CA models were mainly developed using the raster data structure, which does not closely represent the spatial patterns of cadastral parcels. Furthermore, all these case studies were implemented in China, and very few reports have been observed in other study regions.

The main objective of this paper is to integrate and evaluate spatially partitioned transition rules with a vector-based CA model. It is anticipated that the heterogeneity characteristics of sub- regions will be better reflected by multiple sets of transition rules, where the simulated urban growth can be produced in a more accurate and stable pattern. The structure and workflow of CA model (general and partitioned), methodology for transition rules calibration and validation are illustrated in Section 4.2. In the following, Section 4.3 depicts a detailed case study by comparing the simulation results of general and partitioned CA models. Finally, the further discussions and summarization of key findings are presented in Sections 4.4 and 4.5.

4.2 Methodology

4.2.1 The structure of vector CA model As a dynamic spatial model, vector CA comprises the following components (Figure 4.2): 1) Vector cells with specific states, which represent the forms and attributes of corresponding geographical entities at specific time step t; 2) Vector cell space, where the simulation is implemented across the spatial extent of the study region; 3) Transition rules and 4) neighbourhood configuration. They determine the transfer probability of all

cells regarding different kinds of transformation, which is the combination of spatial variables’ forces, neighbourhood influences, constraints for development of study area, as well as random disruption (Liu et al., 2010). 5) Time step refers to the interval of discrete time in each iteration, which leads to spatial interactions between spatial variables and cell space within simulation period, and assists the generation of dynamic and more realistic simulation outcomes taking into account urban development processes.

Figure 4.2 Structure of vector CA model.

4.2.2 Calibration of transition rules In this paper, particle swarm optimization (PSO) has been utilized for the calibration of transition rules. Similar to the relationship between cell and CA, particle is the smallest unit of PSO, it represents one potential solution of the target problem, and is comprised of two parameters: velocity and position:

푃푎푟푡푖푐푙푒 = (푣푛, 푃푛) (1)

Where n is the dimension of target problem, 푣푛 and 푃푛 are the velocity and position of this particle at a specific time point. In specific, they can be represented by n velocities and positions at time t:

푣 = (푣 , 푣 , … , 푣 , 푡) { 푛 1 2 푛 (2) 푃푛 = (푃1, 푃2, … , 푃푛, 푡)

The velocity and position of each particle are updated according to individual and global best positions:

푣(푡 + 1) = 푤 ∗ 푣(푡) + 푐1 ∗ (푃 − 푃(푡)) + 푐2 ∗ (푃 − 푃(푡)) { 푖푏 푔푏 푃(푡 + 1) = 푃(푡) + 푣(푡 + 1) (3)

Where 푤 is the weight of velocity at time t, 푐1 and 푐2 are individual and global learning factors. 푃푖푏 is the best individual position of particle i, and 푃푔푏 is the best global position of all particles, namely the best one of all best individual positions. In addition, 푣(푡 + 1) is the velocity of a particle at time t+1, 푃(푡) and 푃(푡 + 1) are the positions of particle at time t and t+1, separately. Specifically, the nature of an individual particle can be taken as the combination of weights, and the space they explore is the potential weights of all driving factors. In Equation 2, 푣푛 determines how each weight value changes in the next step time t+1, and 푃푛 represents the values of weights for corresponding driving variables at time t. After comparing the existing and all previous positions during the optimization process, the one that could generate highest output value from predefined formula is regarded as the “best” position. Namely, when an individual particle has found its “best” (optimal) position, the simulated land use 74

distribution from this group of weight combination will be produced with the highest accuracy.

Afterwards, the transfer probability 푃(퐶푒푙푙푖) of cell i can be calculated by Equation 4: 1 ( ) 푃 퐶푒푙푙푖 = 푖=푛 ∗ (푏0 + 푛표푟푁푒푖) ∗ 퐶표푛푠() (4) 1 + exp[−(푎0 + ∑푖=1 푓푖 ∗ 푤푖)]

Where 푃(퐶푒푙푙) is the probability of 퐶푒푙푙푖, 푎0 and 푏0 are two constants, 푓푖 푎푛푑 푤푖 are the normalized values of driving variables and corresponding weights. In this study,

푤푖 are derived from PSO method, and n equals to the number of applied variables. 퐶표푛푠() is the constraint of development in the study area, norNei refers the normalized value of neighbourhood configuration (NC), where its original value is: 퐴푟푒푎 푁퐶 = 푟푒푠푖 (5) 퐴푟푒푎푐푒푙푙

In Equation 5, 퐴푟푒푎푐푒푙푙 is the area of centre cell, and 퐴푟푒푎푟푒푠푖 is the area of residential cells within the buffer zone of candidate cell. Specifically, the stochastic perturbation is not considered in order to exclude the uncertainties of randomization. Uncertainties during modelling might influence the validation of transition rules from the analysis result (Li et al., 2014).

To identify the best position of a particle, a fitness value is proposed which refers to the difference between calculated transfer probabilities and real land use change of all candidate cells during the simulation period. Smaller fitness values for a particle are closer to the actual land use map. It is the fitness value that determines whether the current position of particle is the “best” position during the training process. 푛

퐹푖푡(푃푎푟푡푖푐푙푒) = ∑ 퐴푏푠(푃(퐶푒푙푙푗) − 푇(퐶푒푙푙푗)) (6) 푗=1 As Equation 6 illustrates, the fitness value is calculated by the sum of the absolute value

(Abs) of the difference between the calculated transfer probability 푃(퐶푒푙푙푗 ) and real transfer value 푇(퐶푒푙푙푗), were 푗 is the count of sample cells, 푇(퐶푒푙푙푗) is a binary value

and indicates whether the sample 퐶푒푙푙푗 transferred (1) or not (0) at the end of simulation period.

4.2.3 Model validation Three indices are applied to evaluate the performances of PSO-CA models under general and partitioned transition rules: cumulative producer’s spatial accuracy, figure of merit and misclassification frequency.

Producer’s spatial accuracy (PSA) has been widely used for the assessment of precision in the research field of land use modelling, which can be described as: 퐴푟푒푎 푃푆퐴 = 푐표푟 (7) 퐴푟푒푎푎푙푙

Traditionally, the PSA of traditional raster CA is the ratio of 푁푢푚푐표푟 and Num푎푙푙, the correctly classified and total number of cells. Correspondingly, the PSA of vector CA can be calculated by 퐴푟푒푎푐표푟 and 퐴푟푒푎푎푙푙, the correctly simulated and total area of cells.

In this paper, separate experiments are proposed in each combination of transition rules in order to reduce the uncertainties during simulation processes. Therefore, CPSA is defined as the mean value of cumulative PSA from all experiments:

Numsim PSA  (8) CPSA = i=1 Numsim

Where Numsim records the number of simulation experiments that conducted under each combination of CA parameters.

Figure of merit, which was proposed by Klug et al. (1992), is the ratio of the intersection of the observed change and predicted change to the union of the observed change and predicted change (Perica and Foufoula‐Georgiou, 1996). The value of FoM ranges between 0 and 1, which indicates no overlap (0) to perfect overlap (1) between observed 76

and predicted change (Pontius et al., 2008). Compared with PSA, it takes multiple errors of observation into consideration, and can be depicted as: 퐶표푟 FoM = 푠푖푚 (9) 퐶표푟푠푖푚 + 퐸푟푟표푟1 + 퐸푟푟표푟2 + 퐸푟푟표푟3 where 퐶표푟푠푖푚 represents the cells that are observed and simulated as new category,

퐸푟푟표푟1 are the cells that simulated with new states but remain stationary by observation in reality, and 퐸푟푟표푟2 refers to cells which observed change but simulated as persistence. In addition, cells that being predicted with wrong category are classified as 퐸푟푟표푟3.

Misclassification frequency (MF), which evaluates the distribution and frequencies of misclassified cells, is the third proposed index of this study. On the basis of previous research (Huang et al., 2009; De Pinto and Nelson, 2009; Barthold et al., 2013), “Misclassified cells” are defined as cells that are transformed during simulation but which remain in the original state at the end of the simulation period. In Equation 10, the MF of a specific cell:

N MFi = mi (10) N ti

Where 푁mi and Nti are the count of misclassified and total simulation experiments of the cell.

4.2.4 Workflow of general and partitioned CA models The main objective of this study is to integrate the general and partitioned transition rules with CA models, and test their influences on land use change modelling of the study area. There are four main steps in the workflow (Figure 4.3): First, the entire study area is divided into N sub-regions on the basis of administrative boundaries and population distribution of different suburbs within the study area. Second, the discovery of transition rules using the PSO method, where sets of transition rules are generated from

randomly selected sample data of either the entire study area or each sub-region. Third, simulation experiments on land use change will be implemented by general and partitioned transition rules, separately. In the last step, the simulation results are assessed using the producer’s spatial accuracy, figure of merit and misclassification frequency statistics. In specific, the disturbance of uncertainty, which can be derived from modelling process, has been recognised as important issues of land use change simulation (Yeh and Li, 2006; Chen et al., 2017). Therefore, 30 duplicate simulation results have been generated for the purpose of reducing the uncertainties during simulation in this research, and to provide a more comprehensive evaluation of model stability under different types of transition rules.

Figure 4.3 Workflow of general and partitioned PSO-CAs.

4.3 Case study

4.3.1 Study area and data processing Ipswich City is one of the oldest provincial cities in the Southeast Queensland region (SEQ) of Australia, which was initially established in year 1860. It is located at 27°40’ S, 152°41’ E, approximately 45-km from Brisbane, the capital city of Queensland (Figure 4.4). According to the 2016 population census, there were 50,060 families in Ipswich City, with a total of 193,733 residents (ABS, 2016a).

Figure 4.4 Primary land uses of Ipswich City, Australia (year 2016).

(Source of vector map: QSpatial, downloaded: 22 Sep 2017)

The main data sources of the study area are listed in Table 4.1. Using the local government areas dataset, the spatial extent of Ipswich City can be identified. The land use mapping of Queensland, a product of the Queensland Land Use Mapping Program (QLUMP), is classified according to the 8th version of Australian Land Use and Management Classification (Commonwealth of Australia, 2016), and applied as the main data source of this research. Taking its primary land use as the reference, “Production from relatively natural environments” is the dominant land use in year 1999, which occupies 61.09% of the study area. Meanwhile, “Intensive uses” (The land which is subject to substantial modification, generally in association with closer residential settlement, commercial or industrial uses) and “Conservation and natural environments” are the second and third largest land use of the study area, are equal to 19.79% and 13.55% of the entire land use, respectively. The remaining 5.57% of the study area can be classified as “Production from irrigated/dryland agriculture and plantations” and “Water”. By overlapping the primary land use maps of the study area, it is revealed that there is an obvious transformation of Ipswich City during 1999 to 2016. In the year of 2016, the “Production from relatively natural environments” is still the largest primary land use, but its ratio has reduced by 3.89%. On the other hand, “Intensive uses” land has gained a 2.20% increase during the same period. In order to explore the details of land use transformation in a further step, the tertiary land use of the same study area is taken into consideration (Lu et al., 2019). Table 4.1 Data sources used in the case study.

Dataset Year Source

ABS mesh block 2006, 2016 Australian Bureau of

Local government areas 2016 Statistics (ABS)

Baseline roads and tracks of 2017 Queensland

Digital cadastral database (DCDB) 2017 Queensland Government

Digital elevation model (DEM) 2000

Land use mapping of Queensland 1999, 2016

South East Queensland Regional Plan 2009 2009–2031

Table 4.2 Land use statistics of study area(ha).

Primary land use 1999 2016

Conservation and natural environments* 16,255 17,691

Intensive uses* 23,736 26,381

Production from dryland agriculture and plantations 1,004 3,161

Production from irrigated agriculture and plantations 4,420 2,757

Production from relatively natural environments* 73,286 68,618

Water 1,252 1,345

* The primary land use “Conservation and natural environments”, “Intensive uses” and “Production from relatively natural environments” are the sources of second tertiary land uses “Residual native cover”, “Grazing native vegetation” and “Urban residential”, respectively.

Each primary land use in Table 4.2 corresponds to several types of tertiary land use. Among them, the trend of urbanization, which refers to the transformation from grazing native vegetation and residual native cover to urban residential, is the dominant conversion type between years 1999 and 2016. The overall transformed area is 1,225 ha, where 93.35% and 6.65% of these transformed cells are distributed in eastern and western parts of Ipswich City. It is also revealed by the latest national census data (ABS, 2016a) that the majority of residents are living in eastern Ipswich as well. Considering this, Ipswich City is identified as an ideal area for the research with obvious characteristics in both land use transformation and spatial heterogeneity (Figure 4.5). According to the administrative boundaries of suburbs and population distribution, the study region has been divided into two sub-regions: eastern Ipswich is defined as sub- region 1, while the remaining western part becomes sub-region 2.

On the basis of previous research and data availability, ten spatial variables have been identified, which can be categorized into six groups (Table 4.3): Table 4.3 Spatial variables of land use development.

Driving factor Spatial Mean Description group variable value

Discom 2,593.97 Distance to commercial facilities

Discen 11,183.00 Distance to city centre

Disedu 1,754.29 Distance to schools Proximity (m) Distance to main roads (highways Disroad 660.22 and secondary roads)

Dissta 2,531.45 Distance to train stations

The reclassified slope of the 25 m Terrain Slope 1.22 grids

Population Dens 22.03 Population density of year 2016 (persons/ha)

The the ratio of intersected area Neighbourhood Neigh 1.65 and centre cell in 90 m neighboor configuration configuration

The value of Natural Logarithm Shape of vector cell Areacell 6.96 function (ln) with area of vector (m2) cell as the input.

Not The spatial extent of protected Restriction Extentprot applicable areas, Ipswich

1. Proximity Proximity variables measure the accessibility of city amenities, community service centres, downtown area, nearest transportation network, and shopping malls, respectively (Huang et al., 2009). They are considered as the most commonly used spatial variables in land use change modelling. In this research, proximity variables are specified as the Euclidean distances from candidate cells to commercial facilities (Discom), city centre (Discen), schools (Disedu), main roads (Disroad, highways and secondary roads) and train stations (Dissta).The distribution of spatial variables in proximity group are listed in Figure 4.5, and the spatial heterogeneity of the study area is revealed by these variables. Specifically, the city centre of Ipswich City, which refers to its Central Business District (CBD), is located in the central eastern part of entire region (Figure 4.5D). Besides, the vast majority of commercial facilities and schools (Figure 4.5E and Figure 4.5F) are observed in sub-region 1, and only three of them have been detected in sub-region 2. In

comparison, main roads (Figure 4.5G) are evenly distributed across the entire study area. Additionally, railway stations (Figure 4.5H) are scattered along the central axis of Ipswich City in east-west direction.

2. Terrain According to Huang et al. (2009), topographic slope has a great impact on construction feasibility and cost. Slope was derived using a 25 m resolution digital elevation model (DEM). In general, the terrain of Ipswich City is flat, with the steepest part being the southwest. The slope values were reclassified into five classes: [1, 5], [6, 10], [11, 15], [16, 20], [20, 65] (unit: degrees).

3. Population density It has also been observed that the population density (Dens) of the study area is closely related to the expansion of urban land-use area, especially the development of residential areas (Bagan and Yamagata, 2015; Leao et al., 2018). In this study, 2016 population census data (ABS, 2016) is used as the third driving factor group.

4. Neighbourhood configuration Neighbourhood configuration (Neigh) quantifies how a single cell is influenced by its surroundings, which reflects the micro-level interactions, and leads to more complex and dynamic variations at the macro-level (Li et al., 2007). Thus, the neighbourhood configuration of vector CA is defined as the ratio of intersected area (of centre cell’s buffer and its neighbour cells) and the area of centre cell. A pre-tested neighbourhood sensitivity with 30, 60, 90, 120 and 150 m radii (Table 4.4) has been conducted with same ratio of sample data (20%). After 30 separate implementations in each radius, 90 m has achieved the highest spatial accuracy, and thus selected as the optimal radius for CA modelling in this chapter.

Table 4.4 Pre-test result of neighbourhood sensitivity

Radius (m) Pre-test count Average overall accuracy (%)

30 81.93 60 82.10 90 30 82.27 120 82.09 150 82.22

5. Shape of vector cell In this study, the form of vector cell is mainly derived from the digital cadastral database of Queensland, Australia, where each cadastral record corresponds to one vector cell.

Therefore, the area of vector cell (Areacell) is the final driving factor for our CA model.

6. Restriction The spatial extent of protected areas is taken as the ecological conservation redline of land use development. Namely, all transformation will be excluded within the range of protected areas.

Figure 4.5 The partition of sub-regions in study area. Noting the differing levels of influences from variables. 86

4.3.2 Model implementation During 1999 to 2016, 16,299 (1,225.81 ha) non-residential cells were converted into residential while the attributes of 13,019 (68,709.59 ha) non-residential cells remain the same. Taking the previous research on PSO-CAs (Feng et al., 2011; Rabbani et al., 2012; Liao et al., 2014) as reference, 50 particles were adopted for the training of PSO with 20% sample across Ipswich City, which reached a balance between information richness and spatial extent of study area. These samples were randomly selected from the converted and unconverted non-residential cells during simulation period (1999 to 2016).

The PSO-CA model is implemented in the prototype system, which is programmed using the ArcObjects 10.3 software development kit (SDK) on Visual Studio 2013 platform. Based on initial experiments and also existing studies on calibration of transition rules (Feng et al., 2011; Liao et al., 2014), the weight of velocity v, individual and global learning factors c1 and c2 (in Equation 3) were set as 1, 1.5 and 1.5, which represents a balance between the current and best positions of each individual particle, as well as global best positions among all particles. The initial position and velocity of each particle was generated with random values in the intervals [-5, 5] and [-2, 2], and the maximum velocity of particle was restricted to 1 in order to reduce uncertainty and avoid the influence from unrealistically extreme velocities. Following Harrison et al. (2019), 1000 training iterations were used for both general and partitioned PSO training to ensure all particles were well-trained under the same circumstances. In Figure 4.6, the curves of fitness values within one PSO training process, which represents the typical fitness curve of remaining processes by general and partitioned PSO-CA have been depicted. Specifically, the fitness value declines at a rapid speed in the initial training stage. With the increase of iteration count, the rate of decline decreases gradually. The peak of fitness values (1975.18 in general PSO-CA, 1935.95 in partitioned PSO-CA) were obtained at the beginning of PSO training. After 170 iterations, the PSO training in general PSO-CA tended to be stable, which reached 719.81, while the fitness value of the partitioned

PSO reached the balance point (where the previous decreasing trend become relatively stable) after 150 iterations. At the end of training, the final fitness values of general and partitioned PSO-CAs are 701.48 and 693.30, respectively. Considering this, it is concluded that the majority of training work has been completed within the first training stage, which is in accordance with the result from other papers (Liao et al., 2014).

Afterwards, the calibrated weights of all spatial variables were applied for the calculation of transfer probabilities during 100-iteration simulation processes. Each simulation iteration can be simplified into three steps: 1) Calculate the transfer probabilities of all non-residential cells; 2) Select converted cells with higher transfer potentials based on their probabilities in the current iteration; 3) Update the non-residential and residential layers according to the converted cells.

Figure 4.6 The fitness curves of PSO training in general and partitioned CAs.

4.3.3 Simulation results and assessment The average weights and fitness values obtained by both general and partitioned PSO- CAs are listed in Table 4.5. Specifically, the importance of variables is illustrated by the absolute values of their weights, where positive and negative values illustrate whether 89

a relatively larger or smaller value makes a greater contribution to cell conversion (from non-residential to residential). The population density in 2016 (Dens) is identified as the most important variable in general PSO-CA, with a weight value of 312.58. In addition, distances to nearest train stations (Dissta), schools (Disedu) and commercial facilities

nd th (Discom) are the variables ranking 2 to 4 in contribution, with average values of -217.73, -151.32 and -79.43, respectively. Different from Dens, it is revealed by the negative values that non-residential cells situated adjacent to these variables have a higher potential to be developed as residential. In addition, the weights of Disroad and Areacell appear to be equally important, and demonstrate that candidate cells with relatively longer distances to main roads and smaller areas are more likely to be converted.

Similarly, distance to city centre (Discen) can be classified as negative variables, but the influence is not as high as above-mentioned proximity variables. Finally, Slope has the minimum impact on the expansion of residential areas due to the relatively flat terrain of study area. Table 4.5 Comparison of average weights in general and partitioned PSO-CAs.

Partitioned PSO-CA Variables General PSO-CA Sub-region 1 Sub-region 2

Discom -79.43 -96.95 -252.74

Discen -24.37 -2.95 -303.28

Disedu -151.32 -76.69 52.72

Slope -13.38 -1.36 185.95

Disroad 45.02 60.49 -80.62

Dissta -217.73 -127.55 -115.38

Dens 312.58 253.20 292.97

Areacell -43.48 -32.78 -253.42

Fitness value 711.98 667.63 5.40

When it comes to the partitioned PSO-CA, the heterogeneity of sub-regions 1 and 2 are represented by the variation between variable weights. Regardless of value variation, the variable ranking of sub-region 1 is similar to the one in general PSO-CA. The only difference lies in the exchange of 3rd and 4th important variables. Nevertheless, an obvious value change of variable weights is detected in sub-region 2. Discen has been identified as the dominant variable for cell conversion, which is equal to -303.28, while Dens drops to the 2nd important variable (292.97) since relatively few people reside in

rd the western part of Ipswich City. Areacell is the 3 significant variable, which means the smaller cells have a higher priority to be transformed in sub-region 2. Besides, the proximity factor and terrain (Discom, Slope and Disroad) also play more vital roles in the selection of transformed cells with an increased absolute value of weights, except for

Dissta and Disedu.

With the assistance of Figure 4.7, we can identify the simulated expansion of residential areas by both general and partitioned PSO-CAs in one implementation. Overall, the main transformation of non-residential to residential cells occurred in the eastern part of Ipswich City, namely the regions between central Ipswich (West boundary), Karalee (North boundary), Camira (East boundary) and Swanbank (South boundary) (Figure 4.7G). With the commencement of simulation in General PSO-CA, scattered new- residential cells appeared on both sides of the M15 motorway (Figure 4.7A). Concerning iterations 20 to 60, the residential area between M2, M5 and M15 (three main highways of Ipswich) has been further expanded, as 15% of the non-residential area in aforementioned space were converted to residential (Figure 4.7C). Newly converted cells had an obvious increase at the end of the simulation, which can be reflected by the

gradually connected residential area across the east part of Ipswich City (Figure 4.7E). When it comes to the outputs by partitioned PSO-CA, the distribution of transformed cells was similar those from the general model. Nevertheless, a small amount of new residential cells appeared in the districts of Pine Mountain, Rosewood and Willowbank (Figure 4.7B). As the iterations progressed, these residential parts, situated around the original residential area of year 1999, were further expanded (Figure 4.7D and Figure 4.7F).

Taking the actual distribution of residential cells in year 2016 as the reference, the CPSA and FoM of general and partitioned CAs’ outputs have been summarized in Table 4.6. These indices are calculated from 30 separate simulations, which reduced the biases and uncertainties in a single simulation.

Figure 4.7 Simulated distribution of new residential cells (1999 to 2016) using general and spatially partitioned PSO-CA models. The majority of changes are observed between four suburbs (see panel G).

Table 4.6 Average values of CPSA and FoM for residential cells (year 2016).

Average of Standard FoM CPSA (%) deviation

General PSO-CA 81.97 0.35 0.70 Overall 84.10 0.10 0.72 Partitioned Sub-region 1 83.45 0.10 0.72 PSO-CA Sub-region 2 92.90 0.26 0.81

Apart from CPSA and FoM, MF records the frequency rates of all misclassified cells with 30 repeated simulation experiments. Five levels of MF: low, medium-low, medium, medium-high and high are classified on the basis of all MF values in Figure 4.8: [1, 6] (Low), [7, 12] (Medium-low), [13, 18] (Medium), [19, 24] (Medium-high) and [25, 30] (High).

Rates of misclassified cells 100.00% 88.53% 90.00% 80.00% 70.00% 60.00% 46.80%49.90% 50.00% 40.00% 35.19% 25.19% 30.00% 9.10% 20.00% 6.64% 10.91% 6.71% 9.17% 10.00% 5.07% 3.65% 1.01% 2.04% 0.10% 0.00% Low Medium-low Medium Medium-high High

General PSO-CA Partitioned PSO-CA (SR1) Partitioned PSO-CA (SR2)

Figure 4.8 Rates of misclassified cells with different frequencies (%).

4.4 Discussion

Of all variables in both the general and partitioned PSO-CA, Dens is the most significant. It is consistent with the findings from other researchers that population increase is the main variable for the development of residential area and urbanization processes

nd (Weber and Puissant, 2003; Chen et al., 2013). Dissta is the variable with 2 largest contribution to residential development in both general and partitioned (sub-region 1)

PSO-CAs. In terms of the remaining variables, Disedu, Discom, Disroad and Areacell are the 3rd, 4th, 5th and 6th important variables in general PSO-CA, while the ranking has been

th rd th th changed as 4 , 3 , 5 and 6 in partitioned PSO-CA (sub-region 1). Additionally, Discen and Slope are the 7th and 8th variables since the candidate cells are evenly distributed around the centre of Ipswich City, as well as the flat terrain of sub-region 1 (Eastern Ipswich).

Nevertheless, an obvious difference has been observed in regard to the partitioned PSO- CA in sub-region 2, it was operated in the western part of Ipswich City with only 81.56 ha of new residential cells, which corresponding to 6.65% of the entire cell transformation between years 1999 to 2016. According to the training results in sub- region 2, Discen has been identified as the most significant variable of residential development, with an absolute value of 303.28. Dens is still important in this area, which

nd rd is as much as 292.97 and ranks 2 . The 3 largest variable is Areacell, namely cells with relatively small area are likely to be transformed. This corresponds with the fact that the average area of transformed cells (3,054.56 m2) is much smaller than the average area

2 of all candidate cells (174,740.80 m ) in sub-region 2. Besides, Discom, Slope, Dissta, Disroad

th th and Disedu are the 4 to 8 important variable of sub-region 2, which affirms the existence of spatial heterogeneity in Ipswich City. Consequently, it is concluded that the degrees of contribution from spatial variables are switched in different sub-regions, especially for the rural areas, which are 10 km away from the city centre. Additionally, the numerical difference of weights for spatial variables have led to a substantial

improvement in the CPSA in sub-region 2, from only 4.33% (with general transition rules) to 92.90% (with partitioned transition rules) in average. It explains why more accurate results are produced by partitioned PSO-CA model compared with the general version.

As Table 4.6 demonstrates, the average CPSA value of general PSO-CA is 81.97%, with a 0.35% standard deviation. Meanwhile, the introduction of partitioned transition rules has increased the overall CPSA to 84.10%, where the standard deviation is only 0.10%. Additionally, the overall FoM of partitioned PSO-CA is 0.02 higher than the ones from general PSO-CA as well. Concerning the MF, it is also illustrated by Figure 4.8 that misclassified cells with high or low frequencies are the dominant types in both general and sub-region 1 partitioned PSO-CAs, which occupied 81.99% and 75.09% of all misclassified cells, separately. Moreover, misclassified cells with low frequency, which is as much as 88.53%, is the main category output by partitioned PSO-CA in sub-region 2. The results confirm that partitioned PSO-CA model in sub-region 1 has generated the most consistent misclassified cells in 30 repeated simulations with vector format data: 49.90% of the misclassified cells (high frequency) by partitioned transition rules (sub- region 1) have been produced at the same locations in 25 or more out of 30 simulations, where the ratios are 46.80% and only 3.65% in general and partitioned PSO-CAs (sub- region 2).

Apart from the proposed modelling framework in the current research, there is potential for the optimization of partitioned vector CA models. Firstly, in this research, two sub- regions are identified based on administrative boundaries and population distribution. They are used as the input of partitioned CA modelling. Nevertheless, urban heterogeneity consists of spatial differentiation in the physical, biological, and social structures of urban areas (Cadenasso et al., 2007; Zhou et al., 2017). Considering the complexity of spatial heterogeneity, more driving factors from different sectors need to be involved for the definition of sub-regions. A systematic analysis on the count of sub-

regions and modelling accuracy, which is likely to produce a set of optimal rules for this partition mechanism, deserves further exploration. Secondly, regardless of the widely observed dynamics in urban growth (Masek et al., 2000; Kumar et al., 2013; Siddiqui et al., 2018), the variation of spatial heterogeneity during simulation progress has been ignored. Namely, the dynamic characteristics existed in subregions as well. It could lead to geographic and temporal shifts of the influence of spatial variables, and need to be examined as well. Thirdly, it is still unknown whether the partitioned PSO-CA models could be further improved by dividing the sub-region into smaller sub-regions (multi- partitioned) according to the simulation result in the previous model. Therefore, the sensitivity of multi-partitioned CA models, which emphasizes the relationship between partitioned iteration-count and the spatial accuracy/stability of partitioned CA model, needs be tested with additional experiments. Fourthly, it seems necessary to set a baseline for sub-region partition. With the confirmation of long-lasting homogeneity (both physically and mentally) in specific residential communities (Blakely and Snyder, 1997; Hopkins, 2016), the operation of partitioned CA framework at a very fine scale would be an arguable proposal. It will split a residential community into smaller pieces, which generate a “better” but unrealistic output. Namely, a balance between modelling accuracy and reliability from a practical perspective needs to be achieved in order to promote its confidence.

4.5 Conclusions

This research represents a trial of integrating partitioned transition rules with a vector CA model. Based on simulation results and analysis, it has been concluded that partitioned transition rules in different sub-regions achieves a higher overall CPSA and FoM. Moreover, the misclassified cells with high frequency are more spatially concentrated than the ones from general PSO-CA. With the combination and comparison between two case studies in Chapters 3 and 4, the generality of this partitioned PSO-CA method has been validated. Overall, the partitioned PSO-CA

generates more accurate and stable land use patterns at both suburb and city scales. It is identified as a better choice of CA modelling for regions with spatially differentiated characteristics in urbanization, such as Ipswich. This is consistent with previous research using PSO-CA model in UA-Shanghai region (Feng et al., 2018).

During 1999 to 2016, a total of 1225.45 ha areas of non-residential cells have been converted to residential, which can be taken as the representation of the urbanization era of Australia during the past decades. The great majority of newly developed residential areas have been located in the eastern part of Ipswich City, which indicates its stronger urbanization trend. This urbanization trend is corresponding to the expectation of South East Queensland Regional Plan: Ipswich will accommodate the proportions of expansion growth in SEQ region (Queensland Government, 2017).

In summary, spatially partitioned transition rules could capture the relatively dynamic relationship between driving factors and residential development of the study area, which better represents its heterogeneity nature. Prospective work will be focused on partitioned vector CA model concerning its identification ability, dynamics and sensitivity in spatial heterogeneity, as well as the integration with different artificial intelligence algorithms to check the generality of aforementioned modelling framework. The spatial scale of variables, which might be over-estimated in further partitioned sub-regions with finer spatial scale, should be taken into consideration as well.

Chapter 5 The integration of cellular automata and What If? for scenario planning: Future residential expansion in Ipswich City

This chapter is prepared as a manuscript for submission to an international journal in related field: Lu Y, Laffan S, Pettit C. The integration of Cellular Automata and What If? for scenario planning: Future residential expansion in Ipswich City. Selected international journal, In Preparation.

Abstract

The ever-increasing volumes of available data for urban planning and management has led to the development of planning support systems (PSS) for the design of more flexible and people-oriented cities. In a time of rapid urbanization, there has also been a continued focus on land use change models to simulate its complex dynamics. However, the integration of land use change models with planning support systems has received comparatively little attention, despite its potential to provide a more comprehensive understanding of city futures over spatial and temporal scales. Considering this, a cellular automata (CA) land use change model has been coupled with the What If? PSS in this research. Using Ipswich City, South East Queensland (SEQ) region, Australia as the study case, its land use regulations and interaction with surroundings are analysed with multi- source data such as population variation and infrastructure distribution. Land suitability analysis and demand projections have been modelled using What If? with detailed process of residential expansion under specific scenario. This scenario is formulated to explore the possible future land development pattern. By using a scenario planning approach, the proposed CA – What If? model can be applied as a practical tool to analyse the future development of cities. Such data-driven models and tools enable urban 99

planners and policy-makers to explore future growth scenarios in the era of big data and rapid urbanization.

5.1 Introduction

Urban sprawl, which refers to the growth in urban areas with economic development and population migration, has been widely observed at a global-scale during the past couple of decades (Brueckner, 2000; Yu and Ng, 2007; Couch et al., 2008). According to the 2018 World Urbanization Prospects, 55% of the world’s population resides in urban areas, which is almost double the ratio around the middle of the 20th century. It is also projected that there will be a further 18% increase of new urban residents by the end of year 2050 (United Nations, 2018). Nevertheless, the conflict between growing demand for food and decreased agriculture land, as the main source of newly developed metropolitan areas, is a significant challenge to this global urbanization trend. In addition, a growing global population which continued to fuel the process of urbanization has an impact on climate change (Kalnay and Cai, 2003), deforestation (Grimmond, 2007), loss of biodiversity (González-Orozco et al., 2016) and species richness (McKinney, 2008). The exploration of possible urban development under different scenarios, as well as their potential influences on the earth system at multiple spatial and temporal scales. Through the exploration of these planning scenarios, governments could obtain a better understanding of the dynamics between urban spaces, human activities and the surrounding environment.

The idea of PSS can be dated back to the 1970s, when Lee Jr (1973) cast doubt on the construction of large-scale urban models. Such large-scale urban models (LSUMs) were essentially a series of mathematical models attempting to optimise the layout and sequencing of the city. The LSUM were not transparent nor did they lend themselves to collaborative or participatory planning and decision-making processes. Afterwards, PSSs emerged as more collaborative transparent computer tools aimed to assist planners and

100

policy- makers to explore a range of What If? scenarios. PSS have been defined as “Geo- information technology-based instruments that are dedicated to supporting those involved in planning the performance of specific tasks” (Geertman, 2006: 863-880). Through the application of PSS, urban planners and policy-makers can simulate the expansion of their cities in a more intuitive and reliable way. On the basis of modelling techniques, PSS can be categorized into four groups: large-scale, rule-based, state- change and cellular automata (CA) based (Klosterman and Pettit, 2005a). Modern examples that have seen practical use include UrbanSim (Waddell, 2002; Jin and Lee, 2018), What If? (Klosterman, 1999; Pettit et al., 2015; Pettit et al., 2020), CommunityViz (Kwartler and Bernard, 2001; Walker, 2017) and SLEUTH (Clarke and Gaydos, 1998; Rienow and Goetzke, 2015). Among all PSS, the What If? is one of the most widely used GIS based collaborative PSS which has been built to be transparent, flexible, and user- friendly system. What If? is based on the scenario planning approach and included dedicated modules for land supply, land demand and land allocation in order to generate holistic planning schemes (Pettit et al., 2015; Pettit et al., 2020).

The scenario planning approach was initially proposed by Royal Dutch/Shell for the generation and evaluation of its strategic options between late 1960s and early 1970s (Wack, 1985). With the awareness of urban growth and sustainable development (Naess, 2001), scenario planning has been used to forecast and analyse land use patterns, which is a complex issue that involves negotiations and compromises by various stakeholders (Li and Liu, 2008). It is also used as a tool to explore and evaluate the extensive uncertainties of possible future developments (Van Vuuren et al., 2012). The What If? PSS is predicated on a scenario planning approach, which enables planners and policymakers working with key stakeholder to explore an envelope of suitable future urban growth options for a particular region.

101

As a cellular automata mechanism based PSS model, SLEUTH has been used extensively used for scenario planning, with early reported applications in USA: Washington- Baltimore region (Jantz et al., 2004), Houston metropolitan area (Oguz et al., 2007), Tulare County, California (Onsted and Clarke, 2011). Afterwards, it was applied to other regions such as Isfahan Metropolitan Area, Iran (Bihamta et al., 2015) and Changzhou, China (Liu et al., 2017a). Similarly, CommunityViz is being used to assess the impacts of local policies and in Wyoming, USA (Lieske et al., 2003) and Wroclaw, Poland (Kazak et al., 2013), respectively. Furthermore, additional case studies with What If? have been reported in different Australian states: sustainable urban-growth scenarios for Hervey Bay in Queensland, 2021 (Pettit, 2005), land-use change scenarios for Mitchell Shire in Victoria, 2031 (Pettit et al., 2008) and land suitability scenarios of Perth-Peel region in Western Australia, 2031 (Pettit et al., 2015). Generally speaking, the practical value of exploring future city scenarios have been gradually accepted by both researchers and planners. However, the majority of PSSs operate on quantitative spatial planning models, which are static in nature and lack causal effects and random disturbances during urban development. Besides, even though PSS contains the module of land use change, it mainly focuses on the allocation and distribution of land parcels in specific planning applications (Geertman and Stillwell, 2004; Geertman and Stillwell, 2012), rather than the detailed processes of urban land use change, as well as its interactions with spatial variables under different circumstances.

In the field of land use change modelling, researchers have proposed hybrid models and systems by coupling different models to examine a single domain (O’Sullivan et al., 2016). Concerning the aforementioned problem, a hybrid interval-probabilistic land-use allocation model (IPLAM) has been developed and coupled with different development scenarios, which assisted land managers to obtain insight regarding trade-offs between environmental and economic objectives in land use system of Wuhan City, China (Tan et al., 2017). Meanwhile, Ghavami et al. (2017) designed a multi agent based land use

102

planning support system (MALUPS) to simulate the interactions between pre and automated negotiation phrases. It produced a final planning scheme with higher social utility and better spatial land use configurations. However, in spite of current success on hybrid models, there are still relatively few outcomes concerning the coupling of PSS and land use change models, which is expected to promote normative and goal-oriented strategic planning for the future (Couclelis, 2005). To fill this gap, this research proposes a hybrid framework for coupling a PSS with cellular automata, a commonly used land use change modelling approach. The hybrid framework combines top-down (macro-level) requirements of urban system from PSS in an objective and quantitative manner with bottom-up (micro-level) modelling approaches.

This paper is divided into four sections following this introduction: The general workflow of CA – What If? model and its implementation, including demand projection, scenario construction and evaluation, are described in section 5.2. Taking the Ipswich City of Queensland State, Australia as the study region, section 5.3 contains a case study of CA – What If?, which provides detailed allocations of urban development demand to specific locations under one specific scenario. The evaluation of this scenario is elaborated at the end of this section. Lastly, detailed analysis on urban development situation, as well as the key findings of this paper, are summarized in sections 5.4 and 5.5.

5.2 Methodology

5.2.1 General workflow of the CA – What If? model In this paper, the prediction of future scenario is executed in the Online What If? (OWI), a GIS-based planning support system, which is being made available through the Australian Urban Research Infrastructure Network (AURIN)(Pettit et al., 2013). Compared with the traditional desktop version, OWI provides open access for research and planning practitioners. An open access approach supports further

103

application of the tool in different context, such as the exploration of different urban growth scenarios in Pakistan (Hussnain et al., 2020).

The implementation of the CA – What If? model can be split into three stages: setup, scenario construction and allocation (Figure 5.1). Firstly, the uniform area zone file (UAZ) will be uploaded in Project Setup, with the specification of existing land use field. UAZ is the main data source for the OWI system, which assigns the current usage, future plan, and suitability factors of all parcels. In the next step, current and past information of population, housing and employment in different sectors and forecasted years will be input by the user. When it comes to the allocation setup, existing land use will be associated with planned land use, namely the definition of all potential land use conversions. In addition, Demographic Trends is the last step of setup stage, where future residential population and employment will be projected to specific years.

104

Figure 5.1 The implementation of integrated CA – What If? model.

Scenario construction is the second stage of the CA – What If? modelling. The scenario is a combination of development policies, represented by appropriate input data such as comprehensive plans, infrastructure plans, urban growth boundaries, and development restrictions on environmentally sensitive lands, which are linked to all parcels at different spatial scales (Waddell, 2002). In the current model, every scenario comprises three sub- scenarios: suitability, demand and allocation control. They determine the relative suitability, projected future demands, and the control of land use allocation procedures. The future urban development scenario in the projection year is formed by combining these sub-scenarios. Stages 1 (set-up) and 2 (scenario construction) can be classified as 105

“What If?” mode, namely the definition of suitability, demand and allocation strategy of land use scenarios using parameter settings. Finally, allocation of all land use demands to candidate parcels, which correspond to vector cells, is the third stage. It is implemented on the basis of their relative suitability for different land use demands, subject to the specified allocation controls. With this vector-based CA modelling approach, the What if? allocation algorithm is extended to incorporate neighbourhood proximity as a dynamic factor in the land use allocation process. Like other CA-based land use models such as Metronamica (Van Delden and Vanhout, 2018; Navarro Cerrillo et al., 2020) and SLEUTH (Kim and Batty, 2011; Chaudhuri and Clarke, 2013; Chaudhuri and Clarke, 2019), the third stage is composed of multiple iterations, and part of the land use demand will be met by the transformation of vector cells in a single iteration. It is in accordance with the nature of cities: complex systems mainly grow from the bottom-up, their size and shape following well-defined scaling laws that result from intense competition for space (Batty, 2008). In general, the integration of What If? scenario prediction and CA models, which are traditionally worked with historical data in raster format, is considered as one of the innovations in this research.

5.2.2 Population projection and scenario construction It has been demonstrated by Buhaug and Urdal (2013) that the global trend of urbanization is closely related to the continuous population growth, emphasizing the requirement of cross-disciplinary collaborations in urban planning and design. Therefore, the projection of future population is a necessary component of our study. In the CA – What If? model, the predicted population of specific year j can be estimated using the following equation:

푛 푃푗 = 푃푖 ∗ (1 + 푅푝) (1)

Where 푃푗 and 푃푖 are the population of projected and current (the base year for modelling) years, n is the difference between years i and j, 푅푝 is the rate of population growth, which can be derived from the past trend of residential demand:

106

푃푌푖 − 푃푌푗 푅 = (2) 푝 푌푖 − 푌푗

In Equation 2, 푃푦푖 and 푃푦푗 are the total population in years 푌푖 and 푌푗 (as historical data), Rp is the rate of population growth, namely the quotient of population increase and year gap. In OWI, the projection of future population is calculated under the assumption that a future increase of population will be static and equal to the previous value.

As Figure 5.1 indicates, the scenario is the combination of the three sub-scenarios:

Scenario = 푆푐푒푠 ∪ 푆푐푒퐷 ∪ 푆푐푒퐴퐶 (3)

Where 푆푐푒푠 , 푆푐푒퐷 and 푆푐푒퐴퐶 represent the land suitability, demand and allocation control sub-scenarios, respectively.

1. Suitability sub-scenario

The suitability sub-scenario (푆푐푒푠) determines the suitability of different parcels with various land uses. Suitability analysis factors in the opportunity and constraint layers using a standard multiple criteria evaluation (MCE) approach, which begun to be applied in spatial planning since the 1980s (Voogd, 1983). Such computer-based MCE approaches were integrated with GIS in the 1990s (Carver, 1991; Jankowski, 1995; Pettit and Pullar, 1999). The suitability scenario formulation in the CA – What If? model is based on a simple MCE weighted summation method (Marler and Arora, 2010) as illustrated in Equation 4.

S푗푘 = 푓푛 × 푤푛 (4) The suitability of a specific conversion from land use j to k is identical to the sum of driving factor 푓푛 multiplied by its corresponding weight 푤푛 . With the iteration of evaluation, the appropriateness of land use conversion of each parcel will be identified and classified into five categories: Low, Medium-low, Medium, Medium-high and High.

107

2. Demand sub-scenario

The demand sub-scenario ( 푆푐푒퐷 ) takes the population and employment growth projections defined in the Demand Setup, and computes the amount of residential land to accommodate the projected household growth.

In this scenario, the household number of the study area can be estimated by Equation 5:

푛 퐻푗 = 퐻푖 ∗ (1 + 푅ℎ) (5)

Similar to Equation 1, 퐻푗 and 퐻푖 represent the number of households in projected and current years, n is the time gap (unit: year) between them, and 푅ℎ refers to the rate of household growth, which is determined by the past residential data from Demand Setup:

퐻푦푖 − 퐻푦푗 푅 = (6) ℎ 푌푖 − 푌푗

Where 퐻푦1 and 퐻푦2 are the entire number of households of the study area in historical years 푌푖 and 푌푗, respectively.

Besides, the estimated demand of residential land 퐷푒푚푎푛푑푟푒푠푖 can be calculated as:

푃푓 퐵푖 × (1 − 퐼푅푖) ( ) ∗ 퐶표푢푛푡푖 (1−푉푅푖) 퐷푒푚푎푛푑푟푒푠푖 = ∑ (7) 퐷푒푛푓 푖

Where i is a particular type of residential housing, 퐵푖 , 퐼푅푖 and 푉푅푖 refer to the future breakdown percentage, future infill and future vacancy rates of this type. 푃푓 corresponds to the predicted future population, 퐶표푢푛푡푖 records the total number of residential housing i in current year. What’s more, 퐷푒푛푓 is the future density of residential housing 푖, which is the proportion of household number and residential area.

3. Allocation control sub-scenario

The allocation control sub-scenario (푆푐푒퐴퐶) administers the newly developed residential land by taking infrastructure, land use planning and growth patterns into consideration.

108

푆푐푒퐴퐶 judges whether 1) the parcel has public transport service infrastructure and 2) planned land use can be allocated with new residential demand. 1, 푆푎푡푖푠푓푖푒푑 표푟 푤푖푡ℎ표푢푡 푎푙푙표푐푎푡푖표푛 푐표푛푡푟표푙 푃푎푟푐푒푙 = { (8) 푖 0, 푁표푡 푠푎푡푖푠푖푓푖푒푑 푤푖푡ℎ 푎푙푙표푐푎푡푖표푛 푐표푛푡푟표푙 In short, any parcel i (i ≤ total number of parcels within study area) can be developed as new residential if it satisfies one of the following conditions: 1) Have corresponding public transport service infrastructure/Residential is considered as one of the planned land uses; 2) No allocation control is compulsory in this sub-scenario.

Furthermore, as a supplement of the suitability scenario, the selected growth pattern of allocation control scenario will influence the allocation of future land use demand. Different growth patterns identify different areas where parcels have the priority to be allocated.

Finally, after the combination of the sub-scenarios, the land use demand in the demand sub-scenario will be allocated to different locations in a parcel-level scale on the basis of (1) their relative suitability as user defined in the suitability scenario, and (2) the allocation controls in the allocation control scenario.

5.2.3 CA-based allocation process After the scenario construction, the CA model will allocate projected demand according to the result of suitability evaluation. CA is a discrete model with several applications in the simulation of land use change and urban expansion (Santé et al., 2010; Chaudhuri and Clarke, 2013; Li et al., 2017a). The fundamental unit of the CA model is whether a cell will be transformed or not. This is dependent on its transfer probability at a specific time step.

Traditionally, the transfer probability of 푐푒푙푙푖푗 is described as:

109

푡 푡 푡 푝푖푗 = 푃푐 × 훺푖푗 × 푐표푛푖푗 × 푅 (9)

푡 Where 푝푖푗 is the transfer probability of 푐푒푙푙푖푗 at time t, 푃푐 is the suitability of the

푡 current cell (which derived from constructed scenario), 훺푖푗 is the neighbourhood

푡 configuration of 푐푒푙푙푖푗, 푐표푛푖푗 is the construct of transformation in its current location, and R imitates the stochastic perturbation during real urban development (Wu, 2002; Li et al., 2007). The allocation process is composed of multiple iterations, and each iteration is in accordance with one month of time. Specifically, every iteration can be summarized with three key steps: Initially, the land suitability based on transfer probability as calculated using Equation 9 with all candidate cells being ranked in descending order. Next, the non-residential parcels with higher probabilities will be transformed to residential in the current iteration until the summed area of selected parcels reaches the pre-defined threshold. The threshold refers to the ratio of land use demand and iteration number. The third step is to update the residential and non- residential layers, recalculate the values of corresponding driving factors, and continue the allocation process in the next iteration.

5.2.4 Scenario evaluation There is inherent uncertainty in future sociology-economic land use drivers, which results in very different land-use dynamics and consequences (Popp et al., 2017). For providing a relatively comprehensive and objective assessment, the future development scenario will be evaluated with both historical data and a future planning scheme.

As previous research indicates, the evaluation methods of prediction results generally include a point-by-point comparison and an overall comparison (Liu et al., 2006), where the point-by-point validation method overlays the simulation results on the actual situation to compare the accuracy (Zhang et al., 2015). Considering this, three indices for scenario evaluation have been proposed in this section. 110

1. Mean patch size Mean patch size (MPS) equals to the total area of specific land use divided by the number of patches, which can be calculated as:

퐴푖 푀푃푆푖 = (10) 푁푢푚푖

Where 퐴푖 and 푁푢푚푖 are the sum area and number of patches in land use i, respectively. By comparing the MPS values in 2016 (real data) and 2031 (simulated scenario), the variability of patch (parcel) sizes can be identified.

2. Interquartile range As an indicator of descriptive statistics, interquartile range (IQR) is the difference of 3rd and 1st quartile (Q3 and Q1), namely the medians of second and first half of data samples. 퐼푄푅 = 푄3 − 푄1 (11) In comparison with MPS, IQR measures the spread of the value range around the median value, which could exclude the influence from outliers.

3. The compliance rate with planning scheme Apart from the MPS and IQR, the consistency between simulated land use development and planning scheme as the third index. It identifies whether the scenario development has been focused on the targets within specific suburbs and regions with priorities by the department of planning and development:

푁푢푚푖 퐴푖푠 CR푖 = ∑ (12) 퐴푠 푖

CR푖 is the compliance rate in land use i, 퐴푖푠 and 퐴푠 are corresponding to the area of simulated land use i within districts s and the total area of districts s, separately.

111

5.3 Case study

5.3.1 Study area description and data sources The City of Ipswich is a Local Government Area (LGA) in the South-East Queensland (SEQ) region, Australia (Figure 5.2). In the most recent census of population and housing, there were 50,060 households in Ipswich, with a total population of 193,773 (ABS, 2016a). As one of the fastest growing area in Australia, the population of Ipswich is projected to double by 2031, reaching 435,000 (Ipswich, 2015). As a result, Ipswich City is looking to create a prosperous future with new technology solutions and improve decision making within Council.

This research is focused on the future residential development, which represents the most common urbanization and suburbanization processes (Keil, 2018). The main objective is to assess the simulation and projection capability of proposed CA – What If? model, as well as put the modelling framework into practical use.

Figure 5.2 The location of Ipswich City. 112

5.3.2 Future scenario of Ipswich 1. Suitability sub-scenario In the suitability sub-scenario, the factors are derived from datasets sourced from the Australian Bureau of Statistics (ABS) and Queensland Spatial Catalogue (QSpatial, the open-data platform of Queensland Government). The specific factors and detailed categories are determined on the basis of previous What If? study case (Pettit et al., 2015) and data availability. Firstly, activity centre accessibility (F_Actcen) can be classified into two categories, parcels within 1000 m of commercial meshblocks (the smallest geographical area of ABS) near central Ipswich are regarded as “Near district centre”; while the remaining ones are tagged as “None”. Similarly, the thresholds of Education accessibility (F_Edu), Environment value (F_Envirvalue), Public transport accessibility (F_Pubacc) and Slope (F_Slope) are listed in Table 5.1. In addition, the priority areas of urban expansion (F_Urbanexp) are identified according to the South East Queensland Regional Plan (2009-2031), as a reflection of the designed planning scheme. In Table 5.1 The factors and categorized values of suitability sub-scenario.

Factors Data source Categories

(1) Near district centre: Activity centre Land use map within 1000 m of central commercial accessibility 2016 - QSpatial meshblocks (F_Actcen) (2) None: other parcels (1) High: within 800 m of education Education meshblocks Education accessibility (2) Medium: within 800-1,600 m of meshblock - ABS (F_Edu) education meshblocks (3) Low: other parcels

113

(1) Yes: within 1,000 m of protected Environment value Protected areas - areas (F_Envirvalue) QSpatial (2) No: other parcels (1) High suitability: 0-10 degrees Digital elevation Slope (F_Slope) (2) Medium suitability: 10-20 degrees model - QSpatial (3) Low suitability: more than 20 degrees Development (1) Priority expansion: within regional areas - South East development areas (Goodna, Redbank, Urban expansion Queensland Redbank Plains, Springfield, Booval, (F_Urbanexp) Regional Plan Yamanto and Brassall) 2009 - 2031 (2) Ordinary expansion: other parcels Public (1) High: within 800 m of train station Railway stations transportation (2) Medium: within 800-1,600 m of train and sidings - accessibility stations QSpatial (F_Pubacc) (3) Low: other parcels

2. Demand sub-scenario The demographic trends in Ipswich City can be obtained from historical census results (Table 5.2). In the What If? system, the future population in the demand sub-scenario is projected with a static population growth assumption, namely that population and vacancy rate will be stable according to the past and current demographic trends. Accordingly, OWI makes a projection of 1,113.93 ha of residential land use demand during 2016 to 2031. Table 5.2 Demographic trends in Ipswich (years 2006, 2011 and 2016).

Year 2006 Year 2011 Year 2016

Total Population (People) 140,181 166,904 193,733 Housing Units 53,320 60,935 68,674

114

(Dwellings - 2016 dwelling count, 2006/2011 dwelling type) Households (Dwellings - household 47,568 56,327 63,656 composition) Vacancy Rate 7.0% 7.6% 7.3% (Dwellings - dwelling count)

3. Allocation control sub-scenario According to the South East Queensland (SEQ) Regional Plan (Queensland Government, 2017), the state government is committed to develop a dispersed low-density urban settlement pattern in SEQ, including creating resiliency and connectivity within the transportation network. Therefore, a scenario “Growth along transport corridors” has been established in this case study. In this scenario, future residential growth is more likely to be allocated to the parcels near secondary roads.

4. Combined scenarios of future Ipswich development In order to validate the proposed scenario, suitability factors are applied to provide identical standards. The detailed parameters of suitability and allocation control sub- scenarios are summarized in Table 5.3.

Figure 5.3 shows the suitability maps of residential development produced by the suitability sub-scenario. According to the evaluation result, the majority of suitable land is situated in central and eastern Ipswich, and the specific categories of land parcels are listed in Table 5.4. Regarding the suitability of scenario “Growth along transport corridors”, nearly 60% of all candidate parcels are classified as medium-low category.

115

Another 31.39% of parcels with medium suitability have been observed. The remaining classes are medium-high (5.91%), low (2.53%) and high (0.86%).

Table 5.3 Detailed parameters of constructed scenario.

Combined Suitability sub-scenario Allocation sub-

scenario Factor Weights Types control scenario

Near activity centre (65) F_Actcen 80 None (10)

High (75) F_Edu 70 Medium (45) Low (25)

Yes (0) Growth F_Envirvalue 60 Priority allocation No (50) along along transport High suitability (80) transport corridors (less F_Slope 40 Medium suitability (40) corridors than 1000 m) Low suitability (10)

Priority expansion (90) F_Urbanexp 90 Ordinary expansion (40)

High (65) F_Pubacc 60 Medium (35) Low (15)

* The numbers in column “Types” indicate the weights of corresponding suitability factors Table 5.4 Suitability categories of land parcels.

Scenario: Growth along transport corridors Parcel count Area (ha) Ratio (%)

High 120 594.79 0.86 Medium-high 2,272 4,060.07 5.91 116

Medium 5,836 21,558.61 31.39 Medium-low 4,750 40,738.88 59.31 Low 40 1,734.71 2.53

Figure 5.3 Suitability maps of scenario residential development.

5.3.3 Land use allocation by CA models On the basis of suitability map, the CA-based allocation of future land has been implemented. The prediction process consists of 180 iterations to reflect the non-linear urban development at monthly intervals. The detailed allocation process of newly residential areas has been depicted in Figure 5.4. According to the historical land use transformation, the largest area of parcel is 87,061.64 m2. As a result, it is taken as threshold of selected candidate parcels. Namely, the area of newly developed residential parcel should be equal to or less than this threshold during the entire prediction process. 117

Figure 5.4 Predicted scenario developments of new residential areas (2016-2031).

Further extension of current residential areas (year 2016) has been produced by the CA – What If? model, as the main trend in Ipswich City regardless of the difference between parameters. Therefore, 93.65% of newly transformed residential cells are located in the central-eastern part of Ipswich City (Figure 5.4-B), while the remaining 6.35% scattered residential cells are situated in western Ipswich (Figure 5.4-A).

5.3.4 Evaluation of scenario development By combining the newly converted residential parcels (years 2016 - 2031) with the current residential area (year 2016), the mean value and standard deviation of patch size have been calculated below. Table 5.5 Mean value and standard deviation of patch size.

Mean patch Median patch Interquartile

size (m2) size (m2) range (IQR)

Reality (year 2016) 1,028.98 716.57 320.54 “Growth along transport corridors” scenario (year 1,449.93 721.16 365.03 2031)

118

Table 5.5 illustrates the general characteristics of parcel size of all residential land at the beginning (year 2016) and end (year 2031) of simulation period. In order to analyse the range of parcel sizes, they have been classified into five levels based on area extent: Small [0, 0.25), Medium-small [0.25, 0.5), Medium [0.5, 0.75), Medium-large [0.75, 1) and large (higher than 1) (ha). The detailed statistics of parcel classification is recorded in Figure 5.5.

Classification of parcel size (by area) 80.00% 70.41% 70.00% 61.02%

60.00%

50.00%

40.00%

30.00% 17.88% 11.31% 20.00% 12.72% 6.26% 3.53% 6.78% 6.58% 10.00% 3.51% 0.00% Small Medium-small Medium Medium-large Large

Reality, 2016 Scenario: Growth along transport corridors

Figure 5.5 Classification of parcel size (by area).

According to the South East Queensland Regional Plan (Queensland Government, 2017), the following 16 suburbs have been identified as existing urban areas or for priority new residential developments (Figure 5.6): Augustine Heights, Bellbird Park, Booval, Brassall, Brookwater, Camira, Goodna, Redbank, Redbank Plains, Spring Mountain, Springfield, Springfield Central, Springfield Lakes, Thagoona, Walloon, Yamanto. By spatial overlapping the scenario with aforementioned priority development areas, areas of newly developed residential parcels (years 2016 - 2031) within these suburbs are listed in Table 5.6.

119

Table 5.6 Areas of new residential parcels (years 2016 - 2031) within planned suburbs.

“Growth along transport corridors” scenario Suburb Residential area (ha) Ratio (%)

Augustine Heights 0.53 0.05 Bellbird Park 17.40 1.56 Brassall 47.12 4.23 Brookwater 77.16 6.93 Camira 3.36 0.30 Goodna 61.79 5.55 Redbank 104.41 9.37 Redbank Plains 18.68 1.68 Springfield 56.02 5.03 Springfield Central 25.12 2.26 Springfield Lakes 118.11 10.60 Walloon 14.59 1.31 Yamanto 48.10 4.32

All 592.39 53.18

* No newly residential land has been allocated in Booval, Spring Mountain, Thagoona suburbs

120

Figure 5.6 Urban areas for priority new residential developments.

5.4 Discussion

Concerning the MPS and IQR, there is an obvious difference of parcels between the years 2016 (reality) and 2031 (prediction. Specifically, the MPS was 1,028.98 m2 in year 2016, which has increased by 40.91% in the “Growth along transport corridors” scenario at the end of prediction period. Nevertheless, no obvious increment of median patch sizes and IQR have been detected during the same period. Therefore, it is revealed by MPS and IQR that larger parcels are more likely to be selected for residential development during years 2016 to 2031. As SEQ Regional Plan 2009 - 2031 indicates (Queensland Government, 2017), Ipswich’s population will be two times larger than the present, where higher density housing will be developed in more established urban areas.

121

Considering this, it is anticipated that more residential communities with higher floor space ratio (FSR) and more public facilities are likely to be built, which could better accommodate the upcoming residents since detached dwellings represent low density and high-rise apartment buildings are associated with high-density population (Sivam et al., 2012).

The transformation of dwelling types can be also explained by the variation of parcel classes (by area) (Figure 5.5). Although parcels with a smaller size are still the dominant category in both years 2016 and 2031, its ratio has decreased from 70.41% to 61.02% in “Growth along transport corridor” scenario. This downward trend is observed for the ratio of medium-small parcels as well, but relatively inconspicuous compared with variation in small parcels. Conversely, the ratio of large parcels (larger than 1 ha) has been predicted to increase by 11.30% accordingly. Besides, ratios of medium and medium-large parcels are relatively stable during the simulation period. Overall, the varying ratios of small, medium-small and large parcel categories have further confirmed the switch of transformed parcels in future Ipswich City. Additionally, the steady ratios of medium and medium-large parcels imply that the construction of medium-size residential blocks will be continued in Ipswich City, but not as much as high-density residential communities.

As Table 5.6 demonstrates, Springfield Lakes, Redbank and Brookwater are identified as the top three suburbs for the supply of future land use demand. They occupied 299.68 (26.90%) of the entire residential supply in the scenario “Growth along transport corridors”. Besides, Goodna and Springfield are predicted as the location of 61.79 and 56.02 ha additional residential parcels between the years 2016 and 2031, which is as much as 5.55% and 5.03% of all new residential areas and ranked the fourth and fifth among all state suburbs, separately. Generally speaking, 592.39 ha of non-residential areas in year 2016 will be transformed to residential in year 2031. Considering this, the

122

compliance rate of “Growth along transport corridors” scenario (53.18%) indicates that the proposed CA – What If? model could produce a similar pattern of future residential development with reasonable parameter settings, which is in line with the South East Queensland Regional Plan 2009–2031 (Queensland Government, 2017).

In the “Growth along transport corridors” scenario, priority allocation along secondary roads is the allocation control strategy. Consequently, as suburb represents, more newly transformed residential parcels are being observed within 1000 m buffer zones of motorways and motorways in the entire output of the “Growth along transport corridors” scenario (615.43 ha), especially on the southern part of M2 and northern part of M15, separately.

5.5 Conclusions

In this research, an integrated CA – What If? model has been proposed, and applied to a case study in Ipswich City, SEQ Region, Australia. On the basis of overall land use demand during the years 2016 and 2031, detailed allocation schemes under one potential urban development scenario is projected at monthly intervals and parcel-level spatial scale. The evaluation of land allocation schemes has been executed by taking both historical data and South East Queensland (SEQ) Regional Plan 2009 - 2031 into consideration.

It is anticipated that 1,113.93 ha of residential land is required according to the prediction of What If? system, where central and eastern Ipswich will be the continuous growth of residential areas. Springfield Lakes, Redbank and Brookwater are the planned suburbs (Table 5.6) with the largest area of residential transformation, which equal to 26.90% of the overall demand for residential land, respectively. In comparison, the predicted MPS value of “Growth along transport corridors” scenario have been obviously increased compared with the distribution of residential area in year 2016. Additionally, it is also confirmed that the integrated CA – What If? model can be applied for land use

123

demand prediction and allocation simulation. It is working as a refined tool for a more comprehensive understanding of the non-linear, unstable and uncertain world (de Roo, 2018).

In addition to existing functionality and features, further extensions to the proposed CA – What If? model is recommended. Firstly, the mechanism of land subdivision, which has been applied in previous residential urban growth studies (Wrenn and Irwin, 2015; Sun and Taplin, 2018), is suggested as future work. This would ensure the predicted future pattern of land use allocation is akin to the residential growth process in real world. Further research and development will also focus on the identification of the relationships between historical and future development trends of the study area. This would result in rigorous calibration of model parameters for scenario prediction and incorporated into the proposed current CA – What If? Model workflow. With the emergence of opening and shared data, it is anticipated that additional factors on human activities and their interactions with cities’ spatial structure (Schläpfer et al., 2014; Zhong et al., 2017) will be included, which presents a more detailed scenario analysis of future urban development for planners and policy-makers. Finally, the application of the current CA – What If? model would lead to the formulation and evaluation of multiple scenarios to understand the potential growth under different population projections, amongst other key assumptions.

124

Chapter 6 Conclusions and future research

The thesis has extended the theories and methodologies of land use change modelling and spatial analysis by focusing on vector-based CA models as the core component of the research. AI-based algorithms are applied to discover and evaluate the diversified influence from spatial factors and land use change patterns at multiple spatial/temporal scales. A prediction framework has been established to forecast the potential scenarios of urban growth under certain circumstance. It provides comprehensive support for decision making of future urban development.

Specifically, in Chapter 2, the artificial neural network (ANN) method has been applied for CA model calibration, two different formats (raster and vector) of modelling results are compared with each other. A case study is implemented in Collingwood Park and Redbank Plains, two state suburbs of Ipswich City, Queensland. It is demonstrated that the vector-based CA model can generate more realistic models of urbanization processes than raster-based models. Next, the particle swarm optimization (PSO) method is used for the calibration of updated vector-CA models in separate study regions: Bellbird Park - Brookwater and Redbank Plains state suburbs in Chapter 3; and the entire Ipswich City in Chapter 4. Unlike the “black-box” mechanism of the ANN method, the training result of PSO illustrates the specific values of all driving factors. Besides, it is also shown by the simulation outcome that partitioned transition rules could reduce the influence of spatial heterogeneity at a regional scale. Therefore, the vector CA models in Chapters 3 and 4 are regarded as updated versions of the ANN-CA model in Chapter 2. Additionally, Chapter 5 describes a model integrating vector CA and What If?. Based on the achievement in previous chapters, the application of the CA model has been extended from historical land use simulation to the prediction of future development scenarios. It could assist state government and planners to explore the potential patterns of urban development under different policies and strategies aligned to population projections. 125

In Chapter 6, three main conclusions are summarized in section 6.1. Additionally, the implications of potential research are elaborated in section 6.2.

6.1 Main conclusions

There are three main conclusions from this research: (1) The accuracy of CA models can be improved by applying the vector data format; (2) Partitioned transition rules reduce the influence of spatial heterogeneity at regional scale; and (3) Future scenario of urban development in both macro and micro scales can be predicted by integrating vector CA model with planning support system. These are now discussed in more detail.

Main conclusion 1: The accuracy of CA models can be improved by applying vector data format. At the beginning of this thesis, artificial neural networks (ANN) were used to discover the transition rules, and CA models in both vector and raster formats are compared with each other. Three indices, cumulative producer’s spatial accuracy (CPSA), cumulative Kappa coefficient and misclassification frequency have been adopted for the validation of models. The suburbs Collingwood Park and Redbank Plains, which represented of land use change characteristics of Ipswich City, are selected as the study area in Chapter 2.

According to the result of 30 duplicate simulations, both CPSA and cumulative Kappa coefficients of vector-based CA are higher than the index values from raster-based CA. These two indices demonstrate that vector-based CA has produced the distribution of residential and non-residential area with higher accuracy. Besides, it has also generated more misclassified cells (simulated as changing from non-residential to residential but remain non-residential in reality) at the same location out of 30 simulations. The fact that this error is consistent and stable means that it is more easily fixed, as part of future model developments.

126

As the initial stage of land use change modelling in this thesis, the vector-CA model is implemented under the setting that no variation of the form of vector cells occurs during the simulation period. The inter-connection between data formats and simulation accuracy has been validated in Chapter 2. It is consistent with the findings from related research in Canada (Stevens et al., 2007; Moreno et al., 2009; Wang and Marceau, 2013), China (Chen et al., 2014; Long and Shen, 2015; Yao et al., 2017), Portugal (Pinto and Antunes, 2010), Spain (Barreira-González et al., 2015; Barreira-González et al., 2019) and United States (Dahal and Chow, 2015).

Main conclusion 2: Partitioned transition rules reduce the influence of spatial heterogeneity at regional scale. After confirming the advantages of vector CA model, partitioned transition rules, which are proposed to eliminate the influence of spatial heterogeneity, is the second part of research in this thesis. For the discovery of transition rules, particle swarm optimization (PSO) has been applied for computing spatial variables’ importance. The effects of partitioned transition rules are tested by two case studies, separately.

Bellbird Park - Brookwater and Redbank Plains, two suburbs are used for the research of partitioned transition rules in Chapter 3. Each suburb is taken as a sub-region by the partitioned CA model. After training with sample data, the PSO algorithm has identified

Dens, Disroad, Dispub and Discen, Discom, Areacell as leading positive and negative variables in Bellbird Park - Brookwater. Namely, cells with higher/lower absolute values on these variables have a priority of transformation. Nevertheless, the leading variables in Redbank Plains have been changed to some extent. Consequently, Figure 6.1 indicates that there is a difference between all spatial variables in regard to their weights (absolute values) and type (positive or negative) in sub-regions. Concerning the simulation result, the CPSA values of two sub-regions are 79.66% and 90.45% in general PSO-CA, which

127

become 82.99% and 85.44% in partitioned PSO-CA. Namely, the partitioned PSO-CA has reached higher CPSA in one sub-region (Bellbird Park - Brookwater), and dropped in another sub-region (Redbank Plains). Therefore, it is difficult to determine the effects of partitioned transition rules in current case study, and another case study is executed in the following.

150.00 125.64

100.00 89.87 66.48 46.22 50.00 30.00 7.58 0.00 -4.70 -7.93 -14.32 -8.88 -22.75 -27.39 -50.00 -36.06 -36.64

-78.61 -100.00 -80.21 Areacell Dens Discen Discom Dispub Disroad Dissta Slope

2 * Areacell - Area of cell (m ), Dens - Population density (persons/ha), Discen - Distance to city centre (m), Discom - Distance to nearest commercial facilities (m), Dispub - Distance to nearest public services (m), Disroad - Distance to nearest roads (including motorways, highways, secondary roads and local connector roads, m), Dissta - Distance to nearest train station (m), Slope - The level of slope. Figure 6.1 The main spatial variables of sub-regions with positive/negative contribution, Chapter 3.

In the second case study of partitioned CA model, the entire Ipswich City is taken and divided into eastern (sub-region 1) and western (sub-region 2) parts according to the administrative boundaries and population distribution. PSO method is continuously used for the discovery of transition rules. According to the PSO training result (Table 2), there is an obvious difference between the two sub-regions, except for variables Dens and 128

Dissta. Specifically, in sub-region 1, Dismroad is the second important positive variable, while Discom, Disedu and Areacell are making negative contributions to residential development. When it comes to sub-region 2, Disedu and Slope have been transferred to positive variables, while the type Dismroad converts from positive to negative simultaneously. What’s more, even though Discom and Areacell are still identified as negative variables in sub-region 2, there is a 155.79 and 220.64 increase in their absolute values. Additionally, Discen exceeds Dens and working as the most important one among all variables in sub-region 2, which is close to 0 in sub-region 1. Taking the spatial variables of previous study region as reference, it is revealed by Chapter 4 that the spatial heterogeneity is more intense between sub-regions 1 and 2 of the entire Ipswich City.

400.00 292.97 300.00 253.20 185.95 200.00

100.00 60.49 52.72

0.00 -2.95 -1.36 -32.78 -100.00 -76.69 -96.95 -80.62 -127.55 -115.38 -200.00

-300.00 -253.42 -252.74 -303.28 -400.00 Areacell Dens Discen Discom Disedu Dismroad Dissta Slope

2 * Areacell - Area of cell (m ), Dens - Population density (persons/ha), Discen - Distance to city centre (m), Discom - Distance to nearest commercial facilities (m), Dispub - Distance to nearest public services (m), Dismroad - Distance to nearest main roads (including motorways, highways and secondary roads, m), Dissta - Distance to nearest train station

(m), Slope - The slope gradient. Figure 6.2 The main spatial variables of sub-regions with positive/negative contribution, Chapter 4.

129

It is also demonstrated by the simulation outputs that partitioned transition rules have obtained a better effect in Chapter 4: The CPSA value of predicted land use map is 81.97% by PSO-CA with general transition rules, and has been raised to 83.45% and 92.90% in sub-regions 1 and 2 of partitioned PSO-CA. Apart from CPSA, the figure of merit (FoM) of sub-regions 1 (0.72) and 2 (0.81) are also higher than the FoM of general CA (0.70). Consequently, CPSA and FoM values indicate that the partitioned transition rules are capable of identifying the distinctive features of sub-regions, which is essential for understanding the impacts of spatial variables at a finer scale, as well as the recognition of land use patterns with more details. This conclusion has been confirmed by the partitioned and asynchronous cellular automata (PACA)(Ke et al., 2016) and the particle swarm optimization-based CA (PSO-CA)(Feng and Tong, 2018a) models. In general, partitioned transition rules can produce more accurate simulation outputs in large-scale regions with obvious spatial heterogeneity, in comparison with adjacent state-suburbs in limited spatial extent.

Main conclusion 3: Future scenarios of urban development in both macro and micro scales can be predicted by integrating vector CA model with planning support system After the analysis on data format and transition rules of vector CA model, the framework of predicting urban development has been established by integrating vector CA with the What If? planning support system. Traditionally, planning support systems focus mainly on the scenario at the end of forecast period. Nevertheless, the process of land use demand allocation and transformation, which represent the specific city layout during the planning period, could not be reflected in a reasonable way. In comparison, this integrated model not only generates land use scenarios under different demands and strategies, but also provides the simulation of land use change processes at specific intervals. Namely, the complex future scenarios have been incorporated with the simulation process, which is a missing component of the commonly used rule-based planning support systems. The “Growth along transport corridors” scenario has been

130

developed using the CA – What If? model to forecast the future development of residential areas in Ipswich in Chapter 5.

In each scenario, six driving factors have been selected for suitability evaluation. According to the evaluation result of “Growth along transport corridors” scenario, the suitability of most parcel is classified as medium-low (59.31%) and medium (31.39%). Besides, the overall suitability of eastern Ipswich is obvious higher than western Ipswich, which are highly related to the spatial distribution of driving factors.

To assess its reasonability, mean patch size (MPS), interquartile range (IQR) and SEQ Regional Planning Scheme (Queensland Government, 2017) have been taken into consideration. It is concluded that more parcels with large size will be used for future urban development in Ipswich City. Regarding the scenario “Growth along transport corridors”, 11.30% of transformed parcels are predicted with large area (>= 1 ha) (years 2016 to 2013), which was only 6.58% in the past decades (years 1999 to 2016). In addition, there is a 40.91% increase of MPS values as well. Above-mentioned change of parcel sizes will lead to the growing number of residential communities with higher floor space ratios (FSR), which could accommodate the projected 435,000 population of Ipswich in 2031 (Ipswich, 2015). Sixteen suburbs with development priority are extracted according to the 2009-2031 SEQ Regional Planning Scheme (Queensland Government, 2017), where 53.18% predicted land use demand have been allocated.

Overall, the integrated CA – What If? model has been verified in terms of its capability in predicting and allocating land use demand, in macro and micro level, with a 17-year temporal period. It is also confirmed that the continuous expansion of residential areas along transport corridors, will be a reasonable strategy of future Ipswich. This is similar to the findings from Huston and Darchen (2014) that Ipswich is acting as the provider of accommodation for Brisbane City.

131

6.2 Implications for future research

Based on previous chapters and main conclusions, it is illustrated that the vector-based CA models can be coupled with different AI algorithms and a planning support system to identify the historical regulations of land use, and to predict its future development under certain circumstances. They are significant to the decision making in urban planning and management. Meanwhile, it is also believed that there is enough space for the exploration of corresponding research from this point forward.

6.2.1 The generality of vector CA modelling It has been confirmed that the vector-based framework could improve both overall accuracy and stability of CA models in Chapter 2. Nevertheless, as the previous literature review suggested, a diversity of algorithms has been proposed for the calibration and discovery of transition rules. Thus, it would be useful to inspect the generality of the vector-based CA framework by integrating it with different algorithms at a global scale. Besides, this thesis focuses on the transformation from non-residential to residential, which is the main trend of land use in Ipswich City in history (years 1999 to 2016). Notwithstanding, the complexity of urban systems (Batty, 2009a; White et al., 2015) has resulted in multiple land use transformations and mutual interactions in reality. Therefore, the generality of modelling framework concerning different types of land use change, should be further verified as well.

6.2.2 Sensitivity analysis of vector CA modelling in different aspects At present, researchers have made contributions to the sensitivity of cell size, neighbourhood configuration (Ménard and Marceau, 2005; Kocabas and Dragicevic, 2006) and transition rules (Li et al., 2014; Şalap-Ayça et al., 2018). In comparison, relatively few studies are reported on the sensitivity analysis of partitioned CA models. According to Saltelli et al. (2004), sensitivity analysis is defined as the study of how the uncertainty in the model output (numerical or otherwise) can be apportioned to

132

different sources of uncertainty from the model input. With the in-depth understanding of CA model, additional sensitivity analysis of other input parameters is required as well. In Chapter 3 and 4, the partitioned transition rules have been applied in study regions at different spatial scales to explore its effects on reducing the influence of spatial heterogeneity. The partitioned sub-regions are mainly determined by administrative boundaries and population distribution. However, there is no guarantee that spatial heterogeneity will strictly follow these factors alone. On the other hand, it is still uncertain if multi-partitioned sub-regions could be generated according to the misclassified cells in the initial outputs by partitioned rules. Therefore, a multi-partition CA could be formed with systematic analysis on the relationship between sub-region count, partition iteration and modelling results, which is another potential area to improve the performance of vector-based CA models in a further step.

Overall, with the research on the theories and methodologies and application of vector- based CA model, a better understanding on land use distribution and transformation patterns has been achieved in this thesis. AI algorithms and partitioned transition rules provide more objective way to explore the dynamic and non-linear characteristics of our cities. The integration with planning support system expands the functions of vector CA model from discovering historical regulations to predicting a more vivid future. It is anticipated that the research on CA model will continue to prosper with the ongoing “big-data” era, and contribute more to the research of spatial science and human-earth system.

133

References

Abbas KA and Bell MG. (1994) System dynamics applicability to transportation modeling. Transportation Research Part A: Policy and Practice 28: 373-390. Abolhasani S and Taleai M. (2020) Assessing the effect of temporal dynamics on urban growth simulation: Towards an asynchronous cellular automata. Transactions in GIS 24: 332-354. Abolhasani S, Taleai M, Karimi M, et al. (2016) Simulating urban growth under planning policies through parcel-based cellular automata (ParCA) model. International Journal of Geographical Information Science 30: 2276-2301. ABS. (2006) 2006 Census: Mesh Blocks. Available at: http://www.abs.gov.au/ausstats/[email protected]/mf/1209.0.55.002. ABS. (2016a) 2016 Census QuickStats: Ipswich (C). Available at: http://quickstats.censusdata.abs.gov.au/census_services/getproduct/census/20 16/quickstat/LGA33960. ABS. (2016b) 2016 Census: Mesh Blocks. Available at: http://www.abs.gov.au/ausstats/[email protected]/mf/1270.0.55.001. Al-sharif AA and Pradhan B. (2014) Monitoring and predicting land use change in Tripoli Metropolitan City using an integrated Markov chain and cellular automata models in GIS. Arabian journal of geosciences 7: 4291-4301. Almeida C, Gleriani J, Castejon EF, et al. (2008a) Using neural networks and cellular automata for modelling intra-urban land-use dynamics. International Journal of Geographical Information Science 22: 943-963. Almeida C, Gleriani J, Castejon EF, et al. (2008b) Using neural networks and cellular automata for modelling intra‐urban land‐use dynamics. International Journal of Geographical Information Science 22: 943-963. Araya YH and Cabral P. (2010) Analysis and modeling of urban land cover change in Setúbal and Sesimbra, Portugal. Remote Sensing 2: 1549-1563.

134

Arsanjani JJ, Helbich M, Kainz W, et al. (2013) Integration of logistic regression, Markov chain and cellular automata models to simulate urban expansion. International Journal of Applied Earth Observation and Geoinformation 21: 265-275. Asmussen S. (2008) Applied probability and queues: Springer Science & Business Media. Australian Bureau of Agricultural and Resource Economics and Sciences. (2019) Land use. Available at: http://www.agriculture.gov.au/abares/aclump/land-use. Bagan H and Yamagata Y. (2015) Analysis of urban growth and estimating population density using satellite images of nighttime lights and land-use and population data. GIScience & Remote Sensing 52: 765-780. Balmann A. (1997) Farm-based modelling of regional structural change: A cellular automata approach. European review of agricultural economics 24: 85-108. Barlovic R, Santen L, Schadschneider A, et al. (1998) Metastable states in cellular automata for traffic flow. The European Physical Journal B-Condensed Matter and Complex Systems 5: 793-800. Barreira-González P, Aguilera-Benavente F and Gómez-Delgado M. (2017) Implementation and calibration of a new irregular cellular automata-based model for local urban growth simulation: The MUGICA model. Environment and Planning B: Urban Analytics and City Science: 2399808317709280. Barreira-González P, Aguilera-Benavente F and Gómez-Delgado M. (2019) Implementation and calibration of a new irregular cellular automata-based model for local urban growth simulation: The MUGICA model. Environment and Planning B: Urban Analytics and City Science 46: 243-263. Barreira-González P and Barros J. (2016) Configuring the neighbourhood effect in irregular cellular automata based models. International Journal of Geographical Information Science: 1-20. Barreira-González P, Gómez-Delgado M and Aguilera-Benavente F. (2015) From raster to vector cellular automata models: A new approach to simulate urban growth with

135

the help of graph theory. Computers, Environment and Urban Systems 54: 119- 131. Barthold FK, Wiesmeier M, Breuer L, et al. (2013) Land use and climate control the spatial distribution of soil types in the grasslands of Inner Mongolia. Journal of arid environments 88: 194-205. Batisani N and Yarnal B. (2009) Urban expansion in Centre County, Pennsylvania: Spatial dynamics and landscape transformations. Applied Geography 29: 235-249. Batty M. (2008) The size, scale, and shape of cities. science 319: 769-771. Batty M. (2009a) Cities as Complex Systems: Scaling, Interaction, Networks, Dynamics and Urban Morphologies. Batty M. (2009b) Urban modeling. International Encyclopedia of Human Geography. Oxford, UK: Elsevier. Batty M, DeSyllas J and Duxbury E. (2003) The discrete dynamics of small-scale spatial events: agent-based models of mobility in carnivals and street parades. International Journal of Geographical Information Science 17: 673-697. Batty M and Xie Y. (1994) From cells to cities. Environment and Planning B: Planning and Design 21: S31-S48. Batty M, Xie Y and Sun Z. (1999) Modeling urban dynamics through GIS-based cellular automata. Computers, Environment and Urban Systems 23: 205-233. Benenson I, Omer I and Hatna E. (2002) Entity-based modeling of urban residential dynamics: the case of Yaffo, Tel Aviv. Environment and planning B: Planning and design 29: 491-512. Berjak SG and Hearne JW. (2002) An improved cellular automaton model for simulating fire in a spatially heterogeneous Savanna system. Ecological Modelling 148: 133- 151. Bihamta N, Soffianian A, Fakheran S, et al. (2015) Using the SLEUTH urban growth model to simulate future urban expansion of the Isfahan metropolitan area, Iran. Journal of the Indian Society of Remote Sensing 43: 407-414.

136

Blakely EJ and Snyder MG. (1997) Divided we fall: Gated and walled communities in the United States. Architecture of fear 320. Bousquet F and Le Page C. (2004) Multi-agent simulations and ecosystem management: a review. Ecological Modelling 176: 313-332. Brady M, Sahrbacher C, Kellermann K, et al. (2012) An agent-based approach to modeling impacts of agricultural policy on land use, biodiversity and ecosystem services. Landscape ecology 27: 1363-1381. Bright RM, Davin E, O’Halloran T, et al. (2017) Local temperature response to land cover and management change driven by non-radiative processes. Nature Climate Change 7: 296. Britz W, Verburg PH and Leip A. (2011) Modelling of land cover and agricultural change in Europe: Combining the CLUE and CAPRI-Spat approaches. Agriculture, Ecosystems & Environment 142: 40-50. Brown DG. (2014) Advancing land change modeling: opportunities and research requirements: National Academies Press. Brueckner JK. (2000) Urban sprawl: diagnosis and remedies. International regional science review 23: 160-171. Buhaug H and Urdal H. (2013) An urbanization bomb? Population growth and social disorder in cities. Global environmental change 23: 1-10. Cadenasso ML, Pickett ST and Schwarz K. (2007) Spatial heterogeneity in urban ecosystems: reconceptualizing land cover and a framework for classification. Frontiers in Ecology and the Environment 5: 80-88. Cao M, Tang Ga, Shen Q, et al. (2015) A new discovery of transition rules for cellular automata by using cuckoo search algorithm. International Journal of Geographical Information Science 29: 806-824. Cartwright SJ, Bowgen KM, Collop C, et al. (2016) Communicating complex ecological models to non-scientist end users. Ecological Modelling 338: 51-59.

137

Carver SJ. (1991) Integrating multi-criteria evaluation with geographical information systems. International Journal of Geographical Information System 5: 321-339. Chaudhuri G and Clarke KC. (2013) The SLEUTH land use change model: A review. International Journal of Environmental Resources Research 1: 88-104. Chaudhuri G and Clarke KC. (2019) Modeling an Indian megalopolis–A case study on adapting SLEUTH urban growth model. Computers, Environment and Urban Systems 77: 101358-101358. Chen J and Jiang J. (2000) An event-based approach to spatio-temporal data modeling in land subdivision systems. GeoInformatica 4: 387-402. Chen M, Liu W and Tao X. (2013) Evolution and assessment on China's urbanization 1960–2010: under-urbanization or over-urbanization? Habitat International 38: 25-33. Chen Y, Li X, Liu X, et al. (2014) Modeling urban land-use dynamics in a fast developing city using the modified logistic cellular automaton with a patch-based simulation strategy. International Journal of Geographical Information Science 28: 234-255. Chen Y, Liu X and Li X. (2017) Calibrating a Land Parcel Cellular Automaton (LP-CA) for urban growth simulation based on ensemble learning. International Journal of Geographical Information Science 31: 2480-2504. Civco DL. (1993) Artificial neural networks for land-cover classification and mapping. International Journal of Geographical Information Science 7: 173-186. Clarke KC and Gaydos LJ. (1998) Loose-coupling a cellular automaton model and GIS: long-term urban growth prediction for San Francisco and Washington/Baltimore. International Journal of Geographical Information Science 12: 699-714. Clarke KC, Hoppen S and Gaydos L. (1997) A self-modifying cellular automaton model of historical urbanization in the San Francisco Bay area. Environment and planning B: Planning and design 24: 247-261. Cohen J. (1960) A coefficient of agreement for nominal scales. Educational and psychological measurement 20: 37-46.

138

Commonwealth of Australia. (2016) The Australian Land Use and Management Classification (Version 8). Couch C, Petschel-Held G and Leontidou L. (2008) Urban sprawl in Europe: landscape, land-use change and policy: John Wiley & Sons. Couclelis H. (1985) Cellular worlds: a framework for modeling micro—macro dynamics. Environment and Planning A 17: 585-596. Couclelis H. (1989) Macrostructure and microbehavior in a metropolitan area. Environment and planning B: Planning and design 16: 141-154. Couclelis H. (2005) “Where has the future gone?” Rethinking the role of integrated land- use models in spatial planning. Environment and Planning A 37: 1353-1371. Coward M. (2012) Between us in the city: materiality, subjectivity, and community in the era of global urbanization. Environment and Planning D: Society and Space 30: 468-481. Cox D and Miller H. (1965) The theory of stochastic processes: Routledge. Curtis B, Willy C and Bischoff J. (2016) A Sensitivity Analysis for Deriving Dynamic and Evolutionary Rules in an Artificial Immune System-Cellular Automata Model. Proceedings of the International Conference on Modeling, Simulation and Visualization Methods (MSV). The Steering Committee of The World Congress in Computer Science, Computer Engineering and Applied Computing (WorldComp), 167. Dahal KR and Chow TE. (2014) An agent-integrated irregular automata model of urban land-use dynamics. International Journal of Geographical Information Science 28: 2281-2303. Dahal KR and Chow TE. (2015) Characterization of neighborhood sensitivity of an irregular cellular automata model of urban growth. International Journal of Geographical Information Science 29: 475-497. Dale VH. (1997) The relationship between land ‐ use change and climate change. Ecological applications 7: 753-769.

139

De Pinto A and Nelson GC. (2009) Land use change with spatially explicit data: a dynamic approach. Environmental and Resource Economics 43: 209-229. de Roo G. (2018) Ordering Principles in a Dynamic World of Change–On social complexity, transformation and the conditions for balancing purposeful interventions and spontaneous change. Progress in planning 125: 1-32. Deadman P, Brown RD and Gimblett HR. (1993) Modelling rural residential settlement patterns with cellular automata. Journal of environmental management 37: 147- 160. Deadman P, Robinson D, Moran E, et al. (2004) Colonist household decisionmaking and land-use change in the Amazon Rainforest: an agent-based simulation. Environment and planning B: Planning and design 31: 693-709. Eberhart R and Kennedy J. (1995) A new optimizer using particle swarm theory. Micro Machine and Human Science, 1995. MHS'95., Proceedings of the Sixth International Symposium on. IEEE, 39-43. Elsawah S, Guillaume JH, Filatova T, et al. (2015) A methodology for eliciting, representing, and analysing stakeholder knowledge for decision making on complex socio-ecological systems: From cognitive maps to agent-based models. Journal of environmental management 151: 500-516. Encinas AH, Encinas LH, White SH, et al. (2007a) Simulation of forest fire fronts using cellular automata. Advances in Engineering Software 38: 372-378. Encinas LH, White SH, del Rey AM, et al. (2007b) Modelling forest fire spread using hexagonal cellular automata. Applied mathematical modelling 31: 1213-1227. Evans TP and Kelley H. (2004) Multi-scale analysis of a household level agent-based model of landcover change. Journal of environmental management 72: 57-72. Fallah‐Fini S, Rahmandad H, Triantis K, et al. (2010) Optimizing highway maintenance operations: dynamic considerations. System Dynamics Review 26: 216-238.

140

Feng Y. (2017) Modeling dynamic urban land-use change with geographical cellular automata and generalized pattern search-optimized rules. International Journal of Geographical Information Science 31: 1198-1219. Feng Y and Liu Y. (2013) A heuristic cellular automata approach for modelling urban land- use change based on simulated annealing. International Journal of Geographical Information Science 27: 449-466. Feng Y, Liu Y and Batty M. (2015) Modeling urban growth with GIS based cellular automata and least squares SVM rules: a case study in Qingpu–Songjiang area of Shanghai, China. Stochastic Environmental Research and Risk Assessment: 1-14. Feng Y, Liu Y, Tong X, et al. (2011) Modeling dynamic urban growth using cellular automata and particle swarm optimization rules. Landscape and Urban Planning 102: 188-196. Feng Y and Tong X. (2018a) Calibration of cellular automata models using differential evolution to simulate present and future land use. Transactions in GIS 22: 582- 601. Feng Y and Tong X. (2018b) Dynamic land use change simulation using cellular automata with spatially nonstationary transition rules. GIScience & Remote Sensing: 1-21. Feng Y, Wang J, Tong X, et al. (2018) The Effect of Observation Scale on Urban Growth Simulation Using Particle Swarm Optimization-Based CA Models. Sustainability 10: 4002. Fera M, Fruggiero F, Lambiase A, et al. (2017) The role of uncertainty in supply chains under dynamic modeling. International Journal of Industrial Engineering Computations 8: 119-140. Filatova T. (2015) Empirical agent-based land market: Integrating adaptive economic behavior in urban land-use models. Computers, Environment and Urban Systems 54: 397-413.

141

Flache A and Hegselmann R. (2001) Do irregular grids make a difference? Relaxing the spatial regularity assumption in cellular models of social dynamics. Journal of artificial societies and social simulation 4. Forrester JW. (1970) Urban dynamics. IMR; Industrial Management Review (pre-1986) 11: 67. Fragkias M, Güneralp B, Seto KC, et al. (2013) A synthesis of global urbanization projections. Urbanization, biodiversity and ecosystem services: Challenges and opportunities. Springer, Dordrecht, 409-435. Gary Polhill J, Gimona A and Aspinall RJ. (2011) Agent-based modelling of land use effects on ecosystem processes and services. Journal of Land Use Science 6: 75-81. Geertman S. (2006) Potentials for planning support: a planning-conceptual approach. Environment and planning B: Planning and design 33: 863-880. Geertman S and Stillwell J. (2004) Planning support systems: an inventory of current practice. Computers, Environment and Urban Systems 28: 291-310. Geertman S and Stillwell J. (2012) Planning support systems in practice: Springer Science & Business Media. Geertman S, Toppen F and Stillwell J. (2013) Planning support systems for sustainable urban development: Springer. Ghavami SM, Taleai M and Arentze T. (2017) An intelligent spatial land use planning support system using socially rational agents. International Journal of Geographical Information Science 31: 1022-1041. Gibreel TM, Herrmann S, Berkhoff K, et al. (2014) Farm types as an interface between an agroeconomical model and CLUE-Naban land change model: Application for scenario modelling. Ecological indicators 36: 766-778. Gong W, Yuan L, Fan W, et al. (2015) Analysis and simulation of land use spatial pattern in Harbin prefecture based on trajectories and cellular automata—Markov modelling. International Journal of Applied Earth Observation and Geoinformation 34: 207-216.

142

González-Orozco CE, Pollock LJ, Thornhill AH, et al. (2016) Phylogenetic approaches reveal biodiversity threats under climate change. Nature Climate Change 6: 1110. Goodall B. (1987) The Penguin dictionary of human geography: Puffin Books. Grimm V, Augusiak J, Focks A, et al. (2014) Towards better modelling and decision support: documenting model development, testing, and analysis using TRACE. Ecological Modelling 280: 129-139. Grimm V, Berger U, DeAngelis DL, et al. (2010) The ODD protocol: a review and first update. Ecological Modelling 221: 2760-2768. Grimmond SUE. (2007) Urbanization and global environmental change: local effects of urban warming. Geographical Journal 173: 83-88. Guan D, Gao W, Watari K, et al. (2008) Land use change of Kitakyushu based on landscape ecology and Markov model. Journal of Geographical Sciences 18: 455-468. Guan Q, Wang L and Clarke KC. (2005) An artificial-neural-network-based, constrained CA model for simulating urban growth. Cartography and Geographic Information Science 32: 369-380. Guevara J, Garvin MJ and Ghaffarzadegan N. (2017) Capability Trap of the US Highway System: Policy and Management Implications. Journal of Management in Engineering 33: 04017004. Hagenauer J and Helbich M. (2012) Mining urban land-use patterns from volunteered geographic information by means of genetic algorithms and artificial neural networks. International Journal of Geographical Information Science 26: 963-982. Haines-Young R. (2009) Land use and biodiversity relationships. Land Use Policy 26: S178- S186. Han J, Hayashi Y, Cao X, et al. (2009) Application of an integrated system dynamics and cellular automata model for urban growth assessment: A case study of Shanghai, China. Landscape and Urban Planning 91: 133-141.

143

Hare M and Deadman P. (2004) Further towards a taxonomy of agent-based simulation models in environmental management. Mathematics and computers in simulation 64: 25-40. Harrison KR, Ombuki-Berman BM and Engelbrecht AP. (2019) A Parameter-Free Particle Swarm Optimization Algorithm using Performance Classifiers. Information Sciences 503: 381-400. He C, Okada N, Zhang Q, et al. (2008) Modelling dynamic urban expansion processes incorporating a potential model with cellular automata. Landscape and Urban Planning 86: 79-91. He C, Okada N, Zhang Q, et al. (2006) Modeling urban expansion scenarios by coupling cellular automata model and system dynamic model in Beijing, China. Applied Geography 26: 323-345. He C, Shi P, Chen J, et al. (2005) Developing land use scenario dynamics model by the integration of system dynamics model and cellular automata model. Science in China Series D: Earth Sciences 48: 1979-1989. Henderson JV and Wang HG. (2007) Urbanization and city growth: The role of institutions. Regional Science and Urban Economics 37: 283-313. Heppenstall AJ, Crooks AT, See LM, et al. (2011) Agent-based models of geographical systems: Springer Science & Business Media. Herold M, Goldstein NC and Clarke KC. (2003) The spatiotemporal form of urban growth: measurement, analysis and modeling. Remote sensing of Environment 86: 286- 302. Hewitt R and Diaz-Pacheco J. (2017) Stable models for metastable systems? Lessons from sensitivity analysis of a Cellular Automata urban land use model. Computers, Environment and Urban Systems 62: 113-124. Hopkins EA. (2016) The impact of community associations on residential property values. Housing and Society 43: 157-167.

144

Howard D, Howard PJA and Howard D. (1995) A Markov model projection of soil organic carbon stores following land use changes. Journal of environmental management 45: 287-302. Hu Y, Zheng Y and Zheng X. (2013) Simulation of land-use scenarios for Beijing using CLUE-S and Markov composite models. Chinese geographical science 23: 92-100. Huang B, Xie C, Tay R, et al. (2009) Land-use-change modeling using unbalanced support- vector machines. Environment and planning B: Planning and design 36: 398-416. Huang Q, He C, Liu Z, et al. (2014) Modeling the impacts of drying trend scenarios on land systems in northern China using an integrated SD and CA model. Science China Earth Sciences 57: 839-854. Hussnain MQ, Waheed A, Anjum GA, et al. (2020) A framework to bridge digital planning tools' utilization gap in peri-urban spatial planning; lessons from Pakistan. Computers, Environment and Urban Systems 80: 101451. Huston S and Darchen S. (2014) Urban regeneration: An Australian case study insights for cities under growth pressure. International Journal of Housing Markets and Analysis 7: 266-282. Iacono M, Levinson D, El-Geneidy A, et al. (2012) A Markov chain model of land use change in the Twin Cities, 1958-2005. Proceeding 10th Int. Symp. Spat. Accuracy Assess. Nat. Resour. Environ. Sci. 1-24. idcommunity. (2016) City of Ipswich, Population and dwellings. Available at: https://profile.id.com.au/ipswich/population?WebID=10&EndYear=2001&Data Type=UR. Ipswich Co. (2015) Advance Ipswich: For the future of our city and community. Ipswich Co. (2017) About Ipswich. Jankowski P. (1995) Integrating geographical information systems and multiple criteria decision-making methods. International Journal of Geographical Information Science 9: 251-273.

145

Janssen MA, Schoon ML, Ke W, et al. (2006) Scholarly networks on resilience, vulnerability and adaptation within the human dimensions of global environmental change. Global environmental change 16: 240-252. Jantz CA, Goetz SJ and Shelley MK. (2004) Using the SLEUTH urban growth model to simulate the impacts of future policy scenarios on urban land use in the Baltimore-Washington metropolitan area. Environment and planning B: Planning and design 31: 251-271. Jia B, Jiang R, Gao Z-Y, et al. (2005) The effect of mixed vehicles on traffic flow in two lane cellular automata model. International Journal of Modern Physics C 16: 1617- 1627. Jiang W, Chen Z, Lei X, et al. (2015) Simulating urban land use change by incorporating an autologistic regression model into a CLUE-S model. Journal of Geographical Sciences 25: 836-850. Jin J and Lee H-Y. (2018) Understanding residential location choices: an application of the UrbanSim residential location model on Suwon, Korea. International Journal of Urban Sciences 22: 216-235. Kalnay E and Cai M. (2003) Impact of urbanization and land-use change on climate. Nature 423: 528-531. Kamusoko C, Aniya M, Adi B, et al. (2009) Rural sustainability under threat in Zimbabwe– simulation of future land use/cover changes in the Bindura district based on the Markov-cellular automata model. Applied Geography 29: 435-447. Karafyllidis I. (2004) Design of a dedicated parallel processor for the prediction of forest fire spreading using cellular automata and genetic algorithms. Engineering Applications of Artificial Intelligence 17: 19-36. Karafyllidis I and Thanailakis A. (1997) A model for predicting forest fire spreading using cellular automata. Ecological Modelling 99: 87-97.

146

Kassie M, Zikhali P, Pender J, et al. (2010) The economics of sustainable land management practices in the Ethiopian highlands. Journal of agricultural economics 61: 605-627. Kazak J, Szewrański S and Decewicz P. (2013) Indicator-Based Assessment of Land Use Planning in Wrocław Region with CommunityViz. Proceedings REAL CORP: 1248- 1251. Ke X, Qi L and Zeng C. (2016) A partitioned and asynchronous cellular automata model for urban growth simulation. International Journal of Geographical Information Science 30: 637-659. Keil R. (2018) Extended urbanization,“disjunct fragments” and global suburbanisms. Environment and Planning D: Society and Space 36: 494-511. Kerridge J, Hine J and Wigan M. (2001) Agent-based modelling of pedestrian movements: the questions that need to be asked and answered. Environment and planning B: Planning and design 28: 327-341. Kim D and Batty M. (2011) Calibrating cellular automata models for simulating urban growth: Comparative analysis of SLEUTH and Metronamica. Centre for Advanced Spatial Analysis, Paper 176. Klosterman RE. (1999) The what if? Collaborative planning support system. Environment and planning B: Planning and design 26: 393-408. Klosterman RE and Pettit CJ. (2005a) An update on planning support systems. SAGE Publications Sage UK: London, England. Klosterman RE and Pettit CJ. (2005b) An update on planning support systems. Environment and planning B: Planning and design 32: 477-484. Klug W, Grippa G, Tassone C, et al. (1992) Evaluation of long range atmospheric transport models using environmental radioactivity data from the Chernobyl accident(the ATMES report). EUR(Luxembourg). Ko DW, He HS and Larsen DR. (2006) Simulating private land ownership fragmentation in the Missouri Ozarks, USA. Landscape ecology 21: 671-686.

147

Kocabas V and Dragicevic S. (2006) Assessing cellular automata model behaviour using a sensitivity analysis approach. Computers, Environment and Urban Systems 30: 921-953. Kotir JH, Smith C, Brown G, et al. (2016) A system dynamics simulation model for sustainable water resources management and agricultural development in the Volta River Basin, Ghana. Science of the total environment 573: 444-457. Koziatek O and Dragićević S. (2017) iCity 3D: A geosimualtion method and tool for three- dimensional modeling of vertical urban development. Landscape and Urban Planning 167: 356-367. Kumar DS, Arya D and Vojinovic Z. (2013) Modeling of urban growth dynamics and its impact on surface runoff characteristics. Computers, Environment and Urban Systems 41: 124-135. Kuper A. (2004) The social science encyclopedia: Routledge. Kwartler M and Bernard RN. (2001) CommunityViz: an integrated planning support system. Planning support systems: Integrating geographic information systems, models and visualization tools: 285-308. Laffan SW, Skidmore AK and Franklin J. (2012) Geospatial analysis of species, biodiversity and landscapes: introduction to the second special issue on spatial ecology. International Journal of Geographical Information Science 26: 2003-2007. Langroodi RRP and Amiri M. (2016) A system dynamics modeling approach for a multi- level, multi-product, multi-region supply chain under demand uncertainty. Expert Systems with Applications 51: 231-244. Le QB, Park SJ, Vlek PL, et al. (2008) Land-Use Dynamic Simulator (LUDAS): A multi-agent system model for simulating spatio-temporal dynamics of coupled human– landscape system. I. Structure and theoretical specification. Ecological Informatics 3: 135-153. Leao SZ, Troy L, Lieske SN, et al. (2018) A GIS based planning support system for assessing financial feasibility of urban redevelopment. GeoJournal 83: 1373-1392.

148

Lee Jr DB. (1973) Requiem for large-scale models. Journal of the American Institute of Planners 39: 163-178. Li FJ, Dong SC and Li F. (2012) A system dynamics model for analyzing the eco-agriculture system with policy recommendations. Ecological Modelling 227: 34-45. Li X, Chen Y, Liu X, et al. (2017a) Experiences and issues of using cellular automata for assisting urban and regional planning in China. International Journal of Geographical Information Science 31: 1606-1629. Li X, Gong P, Yu L, et al. (2017b) A segment derived patch-based logistic cellular automata for urban growth modeling with heuristic rules. Computers, Environment and Urban Systems 65: 140-149. Li X and Liu X. (2008) Embedding sustainable development strategies in agent‐based models for use as a planning tool. International Journal of Geographical Information Science 22: 21-45. Li X, Liu X and Li S. (2010) Intelligent GIS and spatial optimization. Beijing: Science Press. Li X, Liu X and Yu L. (2014) A systematic sensitivity analysis of constrained cellular automata model for urban growth simulation based on different transition rules. International Journal of Geographical Information Science 28: 1317-1335. Li X, Yang Q and Liu X. (2008) Discovering and evaluating urban signatures for simulating compact development using cellular automata. Landscape and Urban Planning 86: 177-186. Li X, Yeh A, Liu X, et al. (2007) Geographical Simulation Systems: Cellular Automata and Multi-Agents. Beijing: Science Press. Li X and Yeh AG-O. (2002a) Neural-network-based cellular automata for simulating multiple land use changes using GIS. International Journal of Geographical Information Science 16: 323-343. Li X and Yeh AG-O. (2002b) Urban simulation using principal components analysis and cellular automata for land-use planning. Photogrammetric engineering and remote sensing 68: 341-352.

149

Li Z, Guan X, Wu H, et al. (2017c) A novel k-means clustering based task decomposition method for distributed vector-based CA models. ISPRS International Journal of Geo-Information 6: 93. Liang X, Liu X, Li X, et al. (2018) Delineating multi-scenario urban growth boundaries with a CA-based FLUS model and morphological method. Landscape and Urban Planning 177: 47-63. Liao J, Tang L, Shao G, et al. (2014) A neighbor decay cellular automata approach for simulating urban expansion based on particle swarm intelligence. International Journal of Geographical Information Science 28: 720-738. Lieske SN, Hamerlinck JD, Feeney DM, et al. (2003) Aquifer Protection and Community Viz™ in Albany County, Wyoming. Ligtenberg A, Bregt AK and Van Lammeren R. (2001) Multi-actor-based land use modelling: spatial planning using agents. Landscape and Urban Planning 56: 21- 33. Ligtenberg A, Wachowicz M, Bregt AK, et al. (2004) A design and application of a multi- agent system for simulation of multi-actor spatial planning. Journal of environmental management 72: 43-55. Liu D, Zheng X and Wang H. (2020) Land-use Simulation and Decision-Support system (LandSDS): Seamlessly integrating system dynamics, agent-based model, and cellular automata. Ecological Modelling 417: 108924. Liu J, Zhang G, Zhuang Z, et al. (2017a) A new perspective for urban development boundary delineation based on SLEUTH-InVEST model. Habitat International 70: 13-23. Liu X, LI X, AI B, et al. (2006) Multi-agent Systems for Simulating and Planning Land Use Development. Acta Geographica Sinica 10: 011. Liu X, Li X, Liu L, et al. (2008) A bottom‐up approach to discover transition rules of cellular automata using ant intelligence. International Journal of Geographical Information Science 22: 1247-1269.

150

Liu X, Li X, Shi X, et al. (2010) Simulating land-use dynamics under planning policies by integrating artificial immune systems with cellular automata. International Journal of Geographical Information Science 24: 783-802. Liu X, Li X, Yeh AG-O, et al. (2007) Discovery of transition rules for geographical cellular automata by using ant colony optimization. Science in China Series D: Earth Sciences 50: 1578-1588. Liu X, Liang X, Li X, et al. (2017b) A future land use simulation model (FLUS) for simulating multiple land use scenarios by coupling human and natural effects. Landscape and Urban Planning 168: 94-116. Liu Y. (2008) Modelling urban development with geographical information systems and cellular automata: CRC Press. Liu Y, Feng Y and Pontius RG. (2014) Spatially-explicit simulation of urban growth through self-adaptive genetic algorithm and cellular automata modelling. Land 3: 719-738. Long Y and Shen Z. (2015) V-BUDEM: a vector-based Beijing urban development model for simulating urban growth. Geospatial analysis to support urban planning in Beijing. Springer, 91-112. Long Y and Wu K. (2017) Simulating Block-Level Urban Expansion for National Wide Cities. Sustainability 9: 879. Long Y and Zhang Y. (2015) Land-use pattern scenario analysis using planner agents. Environment and Planning B: Planning and Design 42: 615-637. López E, Bocco G, Mendoza M, et al. (2001) Predicting land-cover and land-use change in the urban fringe: a case in Morelia city, Mexico. Landscape and Urban Planning 55: 271-285. Lu Y, Cao M and Zhang L. (2015) A vector-based Cellular Automata model for simulating urban land use change. Chinese geographical science 25: 74-84. Lu Y and Laffan S. (2018) The Use of Particle Swarm Optimization for a Vector Cellular Automata Model of Land Use Change (Short Paper). 10th International

151

Conference on Geographic Information Science (GIScience 2018). Schloss Dagstuhl-Leibniz-Zentrum fuer Informatik. Lu Y, Laffan S, Pettit C, et al. (2019) Land use change simulation and analysis using a vector cellular automata (CA) model: A case study of Ipswich City, Queensland, Australia. Environment and Planning B: Urban Analytics and City Science: 2399808319830971. Ma S, He J, Liu F, et al. (2011) Land-use spatial optimization based on PSO algorithm. Geo-spatial Information Science 14: 54-61. Magliocca N, Safirova E, McConnell V, et al. (2011) An economic agent-based model of coupled housing and land markets (CHALMS). Computers, Environment and Urban Systems 35: 183-191. Mahiny AS and Clarke KC. (2012) Guiding SLEUTH land-use/land-cover change modeling using multicriteria evaluation: towards dynamic sustainable land-use planning. Environment and planning B: Planning and design 39: 925-944. Marler RT and Arora JS. (2010) The weighted sum method for multi-objective optimization: new insights. Structural and multidisciplinary optimization 41: 853- 862. Masek J, Lindsay F and Goward S. (2000) Dynamics of urban growth in the Washington DC metropolitan area, 1973-1996, from Landsat observations. International Journal of Remote Sensing 21: 3473-3486. Masoomi Z, Mesgari MS and Hamrah M. (2013) Allocation of urban land uses by Multi- Objective Particle Swarm Optimization algorithm. International Journal of Geographical Information Science 27: 542-566. Matthews RB, Gilbert NG, Roach A, et al. (2007) Agent-based land-use models: a review of applications. Landscape ecology 22: 1447-1459. McKinney ML. (2008) Effects of urbanization on species richness: a review of plants and animals. Urban ecosystems 11: 161-176.

152

McNamara DE and Keeler A. (2013) A coupled physical and economic model of the response of coastal real estate to climate risk. Nature Climate Change 3: 559-562. Ménard A and Marceau DJ. (2005) Exploration of spatial scale sensitivity in geographic cellular automata. Environment and planning B: Planning and design 32: 693-714. Mitsova D, Shuster W and Wang X. (2011) A cellular automata model of land cover change to integrate urban growth with open space conservation. Landscape and Urban Planning 99: 141-153. Mizraji E. (2006) The parts and the whole: inquiring how the interaction of simple subsystems generates complexity. International journal of general systems 35: 395-415. Moghadam HS and Helbich M. (2013) Spatiotemporal urbanization processes in the megacity of Mumbai, India: A Markov chains-cellular automata urban growth model. Applied Geography 40: 140-149. Mohammad M, Sahebgharani A and Malekipour E. (2013) Urban growth simulation through cellular automata (CA), analytic hierarchy process (AHP) and GIS; case study of 8th and 12th municipal districts of Isfahan. Geogr. Tech 8: 57-70. Moreno N, Ménard A and Marceau DJ. (2008) VecGCA: a vector-based geographic cellular automata model allowing geometric transformations of objects. Environment and planning B: Planning and design 35: 647-665. Moreno N, Wang F and Marceau DJ. (2009) Implementation of a dynamic neighborhood in a land-use vector-based cellular automata model. Computers, Environment and Urban Systems 33: 44-54. Morgan F and O’Sullivan D. (2009) Using binary space partitioning to generate urban spatial patterns. 4th International Conference on Computers in Urban Planning and Urban Management. 1-16. Moser G, Serpico SB and Benediktsson JA. (2012) Land-cover mapping by Markov modeling of spatial–contextual information in very-high-resolution remote sensing images. Proceedings of the IEEE 101: 631-651.

153

Muller MR and Middleton J. (1994) A Markov model of land-use change dynamics in the Niagara Region, Ontario, Canada. Landscape ecology 9: 151-157. Munshi T, Zuidgeest M, Brussel M, et al. (2014) Logistic regression and cellular automata- based modelling of retail, commercial and residential development in the city of Ahmedabad, India. Cities 39: 68-86. Mustafa A, Cools M, Saadi I, et al. (2017) Coupling agent-based, cellular automata and logistic regression into a hybrid urban expansion model (HUEM). Land Use Policy 69: 529-540. Mustafa A, Heppenstall A, Omrani H, et al. (2018a) Modelling built-up expansion and densification with multinomial logistic regression, cellular automata and genetic algorithm. Computers, Environment and Urban Systems 67: 147-156. Mustafa A, Rienow A, Saadi I, et al. (2018b) Comparing support vector machines with logistic regression for calibrating cellular automata land use change models. 51: 391-401. Myint SW and Wang L. (2006) Multicriteria decision approach for land use land cover change using Markov chain analysis and a cellular automata approach. Canadian Journal of Remote Sensing 32: 390-404. Naess P. (2001) Urban planning and sustainable development. European Planning Studies 9: 503-524. Nagel K and Schreckenberg M. (1992) A cellular automaton model for freeway traffic. Journal de physique I 2: 2221-2229. Naghibi F and Delavar MR. (2016) Discovery of Transition Rules for Cellular Automata Using Artificial Bee Colony and Particle Swarm Optimization Algorithms in Urban Growth Modeling. ISPRS International Journal of Geo-Information 5: 241. Nasiri V, Darvishsefat AA, Rafiee R, et al. (2019) Land use change modeling through an integrated Multi-Layer Perceptron Neural Network and Markov Chain analysis (case study: Arasbaran region, Iran). Journal of Forestry Research 30: 943-957.

154

Navarro Cerrillo RM, Palacios Rodríguez G, Clavero Rumbao I, et al. (2020) Modeling Major Rural Land-Use Changes Using the GIS-Based Cellular Automata Metronamica Model: The Case of Andalusia (Southern Spain). ISPRS International Journal of Geo-Information 9: 458. New South Wales Government. (2019) Standard Instrument—Principal Local Environmental Plan. In: legislation N (ed). Newburn D, Reed S, Berck P, et al. (2005) Economics and land‐use change in prioritizing private land conservation. Conservation biology 19: 1411-1420. Newland CP, Maier HR, Zecchin AC, et al. (2018) Multi-objective optimisation framework for calibration of Cellular Automata land-use models. 100: 175-200. O'Sullivan D. (2001) Graph-cellular automata: a generalised discrete urban and regional model. Environment and planning B: Planning and design 28: 687-705. O'Sullivan D. (2008) Geographical information science: agent-based models. Progress in Human Geography 32: 541-550. O’Sullivan D, Evans T, Manson S, et al. (2016) Strategic directions for agent-based modeling: avoiding the YAAWN syndrome. Journal of Land Use Science 11: 177- 187. Oguz H, Klein A and Srinivasan R. (2007) Using the SLEUTH urban growth model to simulate the impacts of future policy scenarios on urban land use in the Houston- Galveston-Brazoria CMSA. Research Journal of Social Sciences 2: 72-82. Omrani H, Tayyebi A and Pijanowski B. (2017) Integrating the multi-label land-use concept and cellular automata with the artificial neural network-based Land Transformation Model: an integrated ML-CA-LTM modeling framework. GIScience & Remote Sensing 54: 283-304. Onsted JA and Clarke KC. (2011) Forecasting enrollment in differential assessment programs using cellular automata. Environment and planning B: Planning and design 38: 829-849.

155

Overmars KP and Verburg PH. (2007) Comparison of a deductive and an inductive approach to specify land suitability in a spatially explicit land use model. Land Use Policy 24: 584-599. Pasaoglu G, Harrison G, Jones L, et al. (2016) A system dynamics based market agent model simulating future powertrain technology transition: Scenarios in the EU light duty vehicle road transport sector. Technological Forecasting and Social Change 104: 133-146. Perica S and Foufoula‐Georgiou E. (1996) Model for multiscale disaggregation of spatial rainfall based on coupling meteorological and scaling descriptions. Journal of Geophysical Research: Atmospheres 101: 26347-26361. Pettit C, Biermann S, Pelizaro C, et al. (2020) A Data-Driven Approach to Exploring Future Land Use and Transport Scenarios: The Online What If? Tool. Journal of Urban Technology 27: 21-44. Pettit C, Keysers J, Bishop I, et al. (2008) Applying the what if? planning support system for better understanding urban fringe growth. Landscape Analysis and Visualisation. Springer, 435-454. Pettit C and Pullar D. (1999) An integrated planning tool based upon multiple criteria evaluation of spatial information. Computers, Environment and Urban Systems 23: 339-357. Pettit CJ. (2005) Use of a collaborative GIS-based planning-support system to assist in formulating a sustainable-development scenario for Hervey Bay, Australia. Environment and planning B: Planning and design 32: 523-545. Pettit CJ, Klosterman RE, Delaney P, et al. (2015) The online what if? Planning support system: A land suitability application in Western Australia. Applied Spatial Analysis and Policy 8: 93-112. Pettit CJ, Klosterman RE, Nino-Ruiz M, et al. (2013) The online what if? Planning support system. Planning support systems for sustainable urban development. Springer, 349-362.

156

Pfaffenbichler P, Emberger G and Shepherd S. (2010) A system dynamics approach to land use transport interaction modelling: the strategic model MARS and its application. System Dynamics Review 26: 262-282. Phillips SJ, Anderson RP and Schapire RE. (2006) Maximum entropy modeling of species geographic distributions. Ecological Modelling 190: 231-259. Pinto N, Antunes AP and Roca J. (2017) Applicability and calibration of an irregular cellular automata model for land use change. Computers, Environment and Urban Systems 65: 93-102. Pinto NN and Antunes AP. (2010) A cellular automata model based on irregular cells: application to small urban areas. Environment and planning B: Planning and design 37: 1095-1114. Pontius GR and Malanson J. (2005) Comparison of the structure and accuracy of two land change models. International Journal of Geographical Information Science 19: 243-265. Pontius RG, Boersma W, Castella J-C, et al. (2008) Comparing the input, output, and validation maps for several models of land change. The Annals of Regional Science 42: 11-37. Popp A, Calvin K, Fujimori S, et al. (2017) Land-use futures in the shared socio-economic pathways. Global environmental change 42: 331-345. Puertas OL, Henríquez C and Meza FJ. (2014) Assessing spatial dynamics of urban growth using an integrated land use model. Application in Santiago Metropolitan Area, 2010–2045. Land Use Policy 38: 415-425. Queensland Government A. (2016) Queensland Spatial Catalogue (QSpatial). Queensland Government A. (2017) South East Queensland Regional Plan 2017. Available at: https://dsdmipprd.blob.core.windows.net/general/shapingseq.pdf. Rabbani A, Aghababaee H and Rajabi MAJJoARS. (2012) Modeling dynamic urban growth using hybrid cellular automata and particle swarm optimization. 6: 063582.

157

Rahim FHA, Hawari NN and Abidin NZ. (2017) Supply & Demand of Rice in Malaysia: A System Dynamics Approach. Int. J Sup. Chain. Mgt Vol 6: 1. Rahman M. (2016) Detection of land use/land cover changes and urban sprawl in Al- Khobar, Saudi Arabia: An analysis of multi-temporal remote sensing data. ISPRS International Journal of Geo-Information 5: 15. Rienow A and Goetzke R. (2015) Supporting SLEUTH–Enhancing a cellular automaton with support vector machines for urban growth modeling. Computers, Environment and Urban Systems 49: 66-81. Rimal B, Zhang L, Keshtkar H, et al. (2017) Monitoring and modeling of spatiotemporal urban expansion and land-use/land-cover change using integrated markov chain cellular automata model. ISPRS International Journal of Geo-Information 6: 288. Robinson DT, Brown DG, Parker DC, et al. (2007) Comparison of empirical methods for building agent-based models in land use science. Journal of Land Use Science 2: 31-55. Roodposhti MS, Aryal J and Bryan BA. (2019) A novel algorithm for calculating transition potential in cellular automata models of land-use/cover change. Environmental Modelling & Software 112: 70-81. Sahrbacher C, Schnicke H, Happe K, et al. (2005) Adaptation of the agent-based model AgriPoliS to 11 study regions in the enlarged European Union. Institute of Agricultural Development in Central and Eastern Europe (IAMO). Şalap-Ayça S, Jankowski P, Clarke KC, et al. (2018) A meta-modeling approach for spatiotemporal uncertainty and sensitivity analysis: an application for a cellular automata-based Urban growth and land-use change model. International Journal of Geographical Information Science 32: 637-662. Saltelli A, Tarantola S, Campolongo F, et al. (2004) Sensitivity analysis in practice: a guide to assessing scientific models: Wiley Online Library. Sanders L, Pumain D, Mathian H, et al. (1997) SIMPOP: a multiagent system for the study of urbanism. Environment and planning B: Planning and design 24: 287-305.

158

Sang L, Zhang C, Yang J, et al. (2011) Simulation of land use spatial pattern of towns and villages based on CA–Markov model. Mathematical and Computer Modelling 54: 938-943. Santé I, García AM, Miranda D, et al. (2010) Cellular automata models for the simulation of real-world urban processes: A review and analysis. Landscape and Urban Planning 96: 108-122. Saysel AK, Barlas Y and Yenigün O. (2002) Environmental sustainability in an agricultural development project: a system dynamics approach. Journal of environmental management 64: 247-260. Schelhorn T, O'Sullivan D, Haklay M, et al. (1999) STREETS: An agent-based pedestrian model. Schläpfer M, Bettencourt LM, Grauwin S, et al. (2014) The scaling of human interactions with city size. Journal of the Royal Society Interface 11: 20130789. Schmolke A, Thorbek P, DeAngelis DL, et al. (2010) Ecological models supporting environmental decision making: a strategy for the future. Trends in ecology & evolution 25: 479-486. Schreinemachers P and Berger T. (2011) An agent-based simulation model of human– environment interactions in agricultural systems. Environmental Modelling & Software 26: 845-859. Schulze J, Müller B, Groeneveld J, et al. (2017) Agent-based modelling of social-ecological systems: achievements, challenges, and a way forward. Journal of artificial societies and social simulation 20. Schutte JF, Koh B, Reinbolt JA, et al. (2003) Scale-independent biomechanical optimization. FLORIDA UNIV GAINESVILLE DEPT OF ELECTRICAL AND COMPUTER ENGINEERING. Schwarz N, Kahlenberg D, Haase D, et al. (2012) ABMland-a tool for agent-based model development on urban land use change. Journal of artificial societies and social simulation 15: 8.

159

Seitzinger SP, Gaffney O, Brasseur G, et al. (2015) International Geosphere–Biosphere Programme and Earth system science: Three decades of co-evolution. Anthropocene 12: 3-16. Semboloni F. (2000) The growth of an urban cluster into a dynamic self-modifying spatial pattern. Environment and planning B: Planning and design 27: 549-564. Sertel E, Ormeci C and Robock A. (2011) Modelling land cover change impact on the summer climate of the Marmara Region, Turkey. International Journal of Global Warming 3: 194-202. Seto KC and Shepherd JM. (2009) Global urban land-use trends and climate impacts. Current Opinion in Environmental Sustainability 1: 89-95. Shahumyan H and Moeckel R. (2017) Integration of land use, land cover, transportation, and environmental impact models: Expanding scenario analysis with multiple modules. Environment and Planning B: Urban Analytics and City Science 44: 531- 552. Shamshad A, Bawadi M, Hussin WW, et al. (2005) First and second order Markov chain models for synthetic generation of wind speed time series. Energy 30: 693-708. Shannon CE. (2001) A mathematical theory of communication. ACM SIGMOBILE mobile computing and communications review 5: 3-55. Shen Q, Chen Q, Tang B-s, et al. (2009) A system dynamics model for the sustainable land use planning and development. Habitat International 33: 15-25. Shepherd S. (2014) A review of system dynamics models applied in transportation. Transportmetrica B: Transport Dynamics 2: 83-105. Shi T and Gill R. (2005) Developing effective policies for the sustainable development of ecological agriculture in China: the case study of Jinshan County with a systems dynamics model. Ecological economics 53: 223-246. Shi Y and Eberhart R. (1998) A modified particle swarm optimizer. Evolutionary Computation Proceedings, 1998. IEEE World Congress on Computational Intelligence., The 1998 IEEE International Conference on. IEEE, 69-73.

160

Siddiqui A, Siddiqui A, Maithani S, et al. (2018) Urban growth dynamics of an Indian metropolitan using CA Markov and Logistic Regression. The Egyptian Journal of Remote Sensing and Space Science 21: 229-236. Silva J, Farinas M, Felfili J, et al. (2006) Spatial heterogeneity, land use and conservation in the cerrado region of Brazil. Journal of Biogeography 33: 536-548. Sivam A, Karuppannan S and Davis MC. (2012) Stakeholders’ perception of residential density: a case study of Adelaide, Australia. Journal of Housing and the Built Environment 27: 473-494. Stevens D, Dragicevic S and Rothley K. (2007) iCity: A GIS–CA modelling tool for urban planning and decision making. Environmental Modelling & Software 22: 761-773. Stimson R, Bell M, Corcoran J, et al. (2012) Using a large scale urban model to test planning scenarios in the Brisbane‐South East Queensland region. Regional Science Policy & Practice 4: 373-392. Subedi P, Subedi K and Thapa B. (2013) Application of a hybrid cellular automaton- markov (CA-Markov) model in land-use change prediction: A case study of Saddle Creek Drainage Basin, Florida. Applied Ecology and Environmental Sciences 1: 126-132. Sun Y and Taplin J. (2018) Adapting principles of developmental biology and agent-based modelling for automated urban residential layout design. Environment and Planning B: Urban Analytics and City Science 45: 973-993. Sun Z and Müller D. (2013) A framework for modeling payments for ecosystem services with agent-based models, Bayesian belief networks and opinion dynamics models. Environmental Modelling & Software 45: 15-28. Suryani E, Chou S-Y and Chen C-H. (2010) Air passenger demand forecasting and passenger terminal capacity expansion: A system dynamics framework. Expert Systems with Applications 37: 2324-2339.

161

Tako AA and Robinson S. (2012) The application of discrete event simulation and system dynamics in the logistics and supply chain context. Decision support systems 52: 802-815. Tan S, Zhang L, Zhou M, et al. (2017) A hybrid mathematical model for urban land-use planning in association with environmental–ecological consideration under uncertainty. Environment and Planning B: Urban Analytics and City Science 44: 54-79. Tang J, Wang L and Yao Z. (2007) Spatio‐temporal urban landscape change analysis using the Markov chain model and a modified genetic algorithm. International Journal of Remote Sensing 28: 3255-3271. Thilak KD and Amuthan A. (2018) Cellular automata-based improved ant colony-based optimization algorithm for mitigating ddos attacks in vanets. Future Generation Computer Systems 82: 304-314. Tian G, Ma B, Xu X, et al. (2016) Simulation of urban expansion and encroachment using cellular automata and multi-agent system model—a case study of Tianjin metropolitan region, China. Ecological indicators 70: 439-450. Torrens PM and O'Sullivan D. (2001) Cellular automata and urban simulation: where do we go from here? : SAGE Publications Sage UK: London, England. Tsolakis N and Anthopoulos L. (2015) Eco-cities: An integrated system dynamics framework and a concise research taxonomy. Sustainable Cities and Society 17: 1-14. Turner BL, Skole D, Sanderson S, et al. (1995) Land-use and land-cover change. Science/Research Plan. Global Change Report (Sweden). Turner BL, Wuellner M, Nichols T, et al. (2016) Development and evaluation of a system dynamics model for investigating agriculturally driven land transformation in the north central United States. Natural Resource Modeling 29: 179-228. Turner WR, Nakamura T and Dinetti M. (2004) Global urbanization and the separation of humans from nature. Bioscience 54: 585-590.

162

Ulli‐Beer S, Gassmann F, Bosshardt M, et al. (2010) Generic structure to simulate acceptance dynamics. System Dynamics Review 26: 89-116. United Nations. (2018) World Urbanization Prospects: The 2018 Revision. Valbuena D, Verburg PH, Bregt AK, et al. (2010) An agent-based approach to model land- use change at a regional scale. Landscape ecology 25: 185-199. Van Delden H and Vanhout R. (2018) A Short Presentation of Metronamica. Geomatic Approaches for Modeling Land Change Scenarios. Springer, 511-519. Van Vuuren DP, Kok MT, Girod B, et al. (2012) Scenarios in global environmental assessments: key characteristics and lessons for future use. Global environmental change 22: 884-895. Vanegas CA, Aliaga DG, Benes B, et al. (2009) Visualization of simulated urban spaces: Inferring parameterized generation of streets, parcels, and aerial imagery. IEEE Transactions on Visualization and Computer Graphics 15: 424-435. Vanegas CA, Kelly T, Weber B, et al. (2012) Procedural generation of parcels in urban modeling. Computer graphics forum. Wiley Online Library, 681-690. Veldkamp A and Fresco L. (1996) CLUE: a conceptual model to study the conversion of land use and its effects. Ecological Modelling 85: 253-270. Veldkamp A and Lambin EF. (2001) Predicting land-use change. Elsevier. Verburg P and Overmars K. (2007) Dynamic simulation of land-use change trajectories with the CLUE-s model. Modelling land-use change. Springer, 321-337. Verburg PH. (2006) Simulating feedbacks in land use and land cover change models. Landscape ecology 21: 1171-1183. Verburg PH, De Koning G, Kok K, et al. (1999) A spatial explicit allocation procedure for modelling the pattern of land use change based upon actual land use. Ecological Modelling 116: 45-61. Verburg PH, de Nijs TC, van Eck JR, et al. (2004a) A method to analyse neighbourhood characteristics of land use patterns. Computers, Environment and Urban Systems 28: 667-690.

163

Verburg PH, Eickhout B and van Meijl H. (2008) A multi-scale, multi-model approach for analyzing the future dynamics of European land use. The Annals of Regional Science 42: 57-77. Verburg PH and Overmars KP. (2009) Combining top-down and bottom-up dynamics in land use modeling: exploring the future of abandoned farmlands in Europe with the Dyna-CLUE model. Landscape ecology 24: 1167. Verburg PH, Schot PP, Dijst MJ, et al. (2004b) Land use change modelling: current practice and research priorities. GeoJournal 61: 309-324. Verburg PH, Soepboer W, Veldkamp A, et al. (2002) Modeling the spatial dynamics of regional land use: the CLUE-S model. Environmental management 30: 391-405. Verburg PH and Veldkamp A. (2001) The role of spatially explicit models in land-use change research: a case study for cropping patterns in China. Agriculture, Ecosystems & Environment 85: 177-190. Viera AJ and Garrett JM. (2005) Understanding interobserver agreement: the kappa statistic. Fam Med 37: 360-363. Voogd H. (1983) Multicriteria evaluation for urban and regional planning: Pion London. Wack P. (1985) Scenarios: Uncharted Waters Ahead. Harvard Business Review September–October. Waddell P. (2002) UrbanSim: Modeling urban development for land use, transportation, and environmental planning. Journal of the American planning association 68: 297-314. Waiyasusri K, Yumuang S and Chotpantarat S. (2016) Monitoring and predicting land use changes in the Huai Thap Salao Watershed area, Uthaithani Province, Thailand, using the CLUE-s model. Environmental Earth Sciences 75: 533. Walker D. (2017) The planners guide to CommunityViz: The essential tool for a new generation of planning: Routledge.

164

Wang F. (1994) The use of artificial neural networks in a geographical information system for agricultural land-suitability assessment. Environment and Planning A 26: 265- 284. Wang F, Hasbani J-G, Wang X, et al. (2011a) Identifying dominant factors for the calibration of a land-use cellular automata model using Rough Set Theory. Computers, Environment and Urban Systems 35: 116-125. Wang F and Marceau DJ. (2013) A Patch‐based Cellular Automaton for Simulating Land‐use Changes at Fine Spatial Resolution. Transactions in GIS 17: 828-846. Wang H, Li X, Long H, et al. (2011b) Development and application of a simulation model for changes in land-use patterns under drought scenarios. Computers & geosciences 37: 831-843. Wang H, Zhang B, Xia C, et al. (2019) Using a maximum entropy model to optimize the stochastic component of urban cellular automata models. 1-23. Wang L, Zhang X and Zhang H. (2010) Principle and structure of CLUE-S model and its progress. Geography and Geo-information Science 26: 73-77. Wang Q, Liu R, Men C, et al. (2018) Application of genetic algorithm to land use optimization for non-point source pollution control based on CLUE-S and SWAT. Journal of Hydrology 560: 86-96. Wang Y, Monzon A and Di Ciommo F. (2015) Assessing the accessibility impact of transport policy by a land-use and transport interaction model–The case of Madrid. Computers, Environment and Urban Systems 49: 126-135. Weber C and Puissant A. (2003) Urbanization pressure and modeling of urban growth: example of the Tunis Metropolitan Area. Remote sensing of Environment 86: 341- 352. Weiss G. (1999) Multiagent systems: a modern approach to distributed artificial intelligence: MIT press.

165

Weng Q. (2002) Land use change analysis in the Zhujiang Delta of China using satellite remote sensing, GIS and stochastic modelling. Journal of environmental management 64: 273-284. White R and Engelen G. (1993) Cellular automata and fractal urban form: a cellular modelling approach to the evolution of urban land-use patterns. Environment and Planning A 25: 1175-1199. White R and Engelen G. (2000) High-resolution integrated modelling of the spatial dynamics of urban and regional systems. Computers, Environment and Urban Systems 24: 383-400. White R, Engelen G and Uljee I. (1997) The use of constrained cellular automata for high- resolution modelling of urban land-use dynamics. Environment and planning B: Planning and design 24: 323-343. White R, Engelen G and Uljee I. (2015) Modeling Cities and Regions as Complex Systems: From Theory to Planning Applications: MIT Press. Wickramasuriya R, Chisholm LA, Puotinen M, et al. (2011) An automated land subdivision tool for urban and regional planning: Concepts, implementation and testing. Environmental Modelling & Software 26: 1675-1684. Wickramasuriya R, Chisholm LA, Puotinen M, et al. (2013) A method to dynamically subdivide parcels in land use change models. International Journal of Geographical Information Science 27: 1497-1513. Wiseman N and Patterson Z. (2016) Testing block subdivision algorithms on block designs. Journal of Geographical Systems 18: 17-43. Wolfram S. (1983) Statistical mechanics of cellular automata. Reviews of modern physics 55: 601. Wrenn DH and Irwin EG. (2015) Time is money: An empirical examination of the effects of regulatory delay on residential subdivision development. Regional Science and Urban Economics 51: 25-36.

166

Wu F. (1998) SimLand: a prototype to simulate land conversion through the integrated GIS and CA with AHP-derived transition rules. International Journal of Geographical Information Science 12: 63-82. Wu F. (2002) Calibration of stochastic cellular automata: the application to rural-urban land conversions. International Journal of Geographical Information Science 16: 795-818. Wu F and Webster CJ. (1998) Simulation of land development through the integration of cellular automata and multicriteria evaluation. Environment and planning B: Planning and design 25: 103-126. Wu N and Silva EA. (2010) Artificial intelligence solutions for urban land dynamics: a review. Journal of Planning Literature 24: 246-265. Wu Q, Li H, Wang R, et al. (2006) Monitoring and predicting land use change in Beijing using remote sensing and GIS. Landscape and Urban Planning 78: 322-333. Xia C, Zhang A, Wang H, et al. (2019) Modeling urban growth in a metropolitan area based on bidirectional flows, an improved gravitational field model, and partitioned cellular automata. International Journal of Geographical Information Science: 1-23. Xie Y. (1996) A generalized model for cellular urban dynamics. Geographical Analysis 28: 350-373. Xu Z and Coors V. (2012) Combining system dynamics model, GIS and 3D visualization in sustainability assessment of urban residential development. Building and Environment 47: 272-287. Yang J, Liu W, Li Y, et al. (2018) Simulating Intraurban Land Use Dynamics under Multiple Scenarios Based on Fuzzy Cellular Automata: A Case Study of Jinzhou District, Dalian. Complexity 2018. Yang J, Su J, Chen F, et al. (2016) A local land use competition cellular automata model and its application. ISPRS International Journal of Geo-Information 5: 106.

167

Yang J, Tang Ga, Cao M, et al. (2013) An intelligent method to discover transition rules for cellular automata using bee colony optimisation. International Journal of Geographical Information Science 27: 1849-1864. Yang Q, Li X and Shi X. (2008) Cellular automata for simulating land use changes based on support vector machines. Computers & geosciences 34: 592-602. Yang X and Liu Z. (2005) Using satellite imagery and GIS for land‐use and land‐cover change mapping in an estuarine watershed. International Journal of Remote Sensing 26: 5275-5296. Yang X, Zheng X and Lv L. (2012) A spatiotemporal model of land use change based on ant colony optimization, Markov chain and cellular automata. Ecological Modelling 233: 11-19. Yao Y, Liu X, Li X, et al. (2017) Simulating urban land-use changes at a large scale by integrating dynamic land parcel subdivision and vector-based cellular automata. International Journal of Geographical Information Science: 1-28. Yeh AG-O and Li X. (2006) Errors and uncertainties in urban cellular automata. Computers, Environment and Urban Systems 30: 10-28. Yu W, Zang S, Wu C, et al. (2011) Analyzing and modeling land use land cover change (LUCC) in the Daqing City, China. Applied Geography 31: 600-608. Yu XJ and Ng CN. (2007) Spatial and temporal dynamics of urban sprawl along two urban–rural transects: A case study of Guangzhou, China. Landscape and Urban Planning 79: 96-109. Zeng Y, Yu M and Li S. (2018) Urban Expansion Modeling Approach Based on Multi-Agent System and Cellular Automata. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences 42: 3. Zhang H, Jin X, Wang L, et al. (2015) Multi-agent based modeling of spatiotemporal dynamical urban growth in developing countries: simulating future scenarios of Lianyungang city, China. Stochastic Environmental Research and Risk Assessment 29: 63-78.

168

Zhang H, Zeng Y and Bian L. (2010) Simulating multi-objective spatial optimization allocation of land use based on the integration of multi-agent system and genetic algorithm. International Journal of Environmental Research 4: 765-776. Zhong C, Schläpfer M, Müller Arisona S, et al. (2017) Revealing centrality in the spatial structure of cities from human activity patterns. Urban Studies 54: 437-455. Zhou F, Xu Y, Chen Y, et al. (2013) Hydrological response to urbanization at different spatio-temporal scales simulated by coupling of CLUE-S and the SWAT model in the Yangtze River Delta region. Journal of Hydrology 485: 113-125. Zhou W, Pickett ST and Cadenasso ML. (2017) Shifting concepts of urban spatial heterogeneity and their implications for sustainability. Landscape ecology 32: 15- 30.

169