Urban growth modeling for the assessment of future climate and disaster risks: approaches, gaps and needs

Urban climate-related disaster risks are set to rise, driven by the interaction of two global megatrends: urbanization and climate change. A detailed understanding of whether, where and how cities are growing within or into hazard-prone areas is an urgent prerequisite for assessing future risk trajectories, risk-informed planning, and adaptation decisions. However, this analysis has been mostly neglected to date, as most climate change and disaster risk research has focused on the assessment of future hazard trends but less on the assessment of how socio-economic changes affect future hazard exposure. Urban growth and expansion modeling provide a powerful tool, given that urban growth is a major driver of future disaster risk in cities. The paper reviews the achievements lately made in urban growth and exposure modeling and assesses how they can be applied in the context of future-oriented urban risk assessment and the planning of adaptation measures. It also analyses which methodological challenges persist in urban growth and exposure modeling and how they might be overcome. These points pertain particularly to the need to consider and integrate (1) urban morphology patterns and potential linkages to exposure as well as vulnerability, (2) long-term time horizons to consider long-term developments, (3) feedbacks between urbanization trajectories and hazard trends, (4) the integration of future urban growth drivers and adaptation responses, (5) feedbacks between adaptation and urbanization, and (6) scenarios, which are developed within a commonly defined scenario framework.


Introduction
The risk of disasters in relation to environmental hazards has been increasing globally and there is strong reason to assume that it will continue to rise, driven by climate change, local environmental degradation and socio-economic changes, notably population growth and urbanization in hazard-exposed areas such as coasts [1,2].However, when exploring future trends in disaster risk, most scientific attention over the last two decades has been directed towards assessing changes in the climate system and the implications for natural hazard patterns (e.g.flooding, or landslide) [3].The widely-accepted standard definition of risk by the Intergovernmental Panel on Climate Change (IPCC) sees risk as not only driven by hazards, but as the product of the dynamic interaction between hazards, exposure and vulnerability [1,4].Therefore, in order to properly understand potential future risk trajectories and potential options for adaptation, the dynamic assessment of risks cannot be limited to the assessment of future trends in hazards alone, but also needs to take into consideration potential shifts in exposure and vulnerabilities [5][6][7][8][9][10][11].
Urbanization is one of the key trends shaping humanity's future exposure, vulnerability and thus risk to natural hazards.A rapidly growing body of literature provides evidence that cities need to be considered as primary hotspots of climate-related disaster risks [12].They are home to over 50% of the world's population and are expected to host two thirds of humankind by the middle of this century [13].Cities also host a majority of the world's physical and economic assets and are often places of high social vulnerability and precariat, especially in developing countries and many emerging economies [14].In transition countries and emerging economies, urbanization leads to great shifts not only with respect to the question where human life and economic activity takes place, e.g.whether new exposure is generated by cities growing into hazardous zones such as floodplains [15].It also matters in terms of how social and economic systems are configured, having strong implications on societal exposure [16].
An understanding of future urban growth pathways is hence of critical importance for examining humanity's risk trajectories at large.Yet, whilst the modeling of future hazard trends has received a lot of scientific attention and funding, the assessment of urbanization-led trajectories in exposure has for long been neglected.This rift triggers four key questions which serve to guide the subsequent sections of this paper: 1. Which requirements would urban growth models ideally need to fulfill in order to allow for their application towards urban exposure and eventually urban risk modeling?2. Which advancements have been made in modeling future urban growth, densification and expansion, and which potentials do distinct types of models offer for urban exposure and risk modeling?3. Which approaches have been used explicitly in the context of modeling future exposure trajectories of cities in the context of climate and disaster risks?4. Which knowledge gaps and methodological challenges persist in urban growth and exposure modeling-and what would be central elements of a research agenda aiming to overcome them?
To answer these questions, we assessed the stateof-the-art in urban growth modeling (section 3) and carried out a systematic and structured literature review on the use of urban growth models in future exposure studies (section 4).Based on these, we identify and highlight major research gaps in section 5 and suggest a future research agenda for taking the field forward.

Requirements of urban exposure modeling
Urban growth models should ideally possess certain characteristics and capabilities for the simulation of future trajectories of urban exposure towards natural and climatic hazards.These requirements presented in the following were identified and compiled based on the analysis of studies in the field of risk management involving academic literature [5,7,8,[16][17][18][19][20] and practical experiences [21,22] that address current challenges in the assessment of future risks.
When employed for the projection of future exposure within the context of climate and disaster risk assessments, urban growth models ideally should have the following thematic capabilities: 1. Ideally, they should be capable of representing different types of urban morphology and the respective socio-economic profiles, including population densities and hence numbers.For example, for urban exposure considerations in the environmental hazard context, it makes a significant difference, whether a residential area, an industrial zone, or an area with office buildings is exposed to a hazard.Sheer information on the urbanized area per se is therefore only to a limited extent meaningful for risk management purposes.2. They should also be capable of expressing urban densification, next to sprawl.While many other applications are concerned primarily with sprawl, densification considerations are of great relevance for exposure assessments due to their influence on the many hazards, e.g. for flooding, where densification reduces the capacity for inner-city flood retention, or for heat waves, where densification itself intensifies heat island effects.3. Further, they may also need to refer to longer time frames to the end of the century required, for instance, for the context of sea level rise exposure and the fit of long-lasting urban infrastructure such as major port facilities and different time-steps within them.4. At the same time, they should provide timestep information within this longer window, in order to inform planning purposes and policies with a shorter duration, e.g.land use zoning and development plans, as well as adaptive adjustments to adaptation pathways.5.In doing so, such models also should be able to generate a range of scenarios with different assumptions on socio-economic and climatic boundary conditions, rather than producing business-as-usual projections only.A major objective of modeling work for exposure and risk assessments is to explore the width of possible future trajectories in order to prove possible risk reduction and adaptation measures against them.6. Models also should be capable of considering the potentially changing composition and relative importance of different drivers of urban change over time.Especially with increasing climate change impacts, factors of climatic or socio-economic nature shaping urban growth and densification might come into play which might not even exist today, e.g.restrictive land use zoning in flood plains, economic transitions in a city, or gentrification processes.7.In this sense, models need to be able to represent social and political decisions of transformative character and the societal responses to them.This includes new planning approaches for sustainable urban development, such as the implementation of swamp city approaches, or the relocation of settlements at risk.8. Ideally such models should even be capable of integrating adaptation measures, e.g.ring dykes, flood-proofing of buildings or renaturing of rivers, and their effects on future growth, densification and exposure.9.In doing so, advanced models would benefit from not being limited to urbanization effects in the core urban area, but also accounting for surrounding land use changes from urbanization and industrialization effects, which need to be considered for e.g. the flood models on catchment level which produce the hazard layers.10.In order to be relevant for concrete risk and contingency management, advanced exposure models should ideally be able to represent a spacetime-trajectory, meaning the temporal shifts of exposed elements over a day or year.This is important as business districts, for example, typically increase their population manifold during the day, when compared to their nighttime population.11.Ideally, models will be capable to deal with and move in between different geographical scales, in order to generate nested scenarios that fit to the needs of actors at various scales whilst being coherent in their modeling approach, assumptions and results.12.In the sense of model application, it would be valuable if such models are transferrable and compatible for applications in various physical and socio-economic conditions, i.e. in different cities, countries, and geographical locations.
Considering the above set of demands, the next section reviews existing modeling types against these criteria, in order to assess the implications and suitability of existing urban growth models for exposure simulation.

Urban growth modeling: types and implications for urban exposure simulation
Urban growth modeling is a long-established and continuously growing research field providing a bet-ter understanding of past evolutions of urban systems and future growth trends [23].The first model approaches already popped up, when social and political developments in the second half of the 20th century raised questions specific to the urban contexts, such as intense traffic due to the rapid increase in the number of cars [24] or conflicts from to urbanization of agricultural land [25].Thus, the field of urban modeling emerged in order to contribute to the understanding of how cities might evolve over time [26][27][28].However, early models had only limited validity due to limitations in accuracy, computing power and data availability [29].Since the late 1990ies, improving data storage capacities and computing power and the increasing availability of highresolution satellite imagery have pushed the development of novel approaches exploiting the new opportunities forward.Within the last twenty years, a wide range of urban growth models has been developed with different mechanisms and foci depending on their scope of applications.
The following chapter introduces the four classes of urban growth models and presents an overview, how far they already meet the requirements for urban exposure modeling defined in chapter 2.

Classes of urban growth models
Urban growth is defined as the change of land use/land cover (LULC) towards urban land use and the evolution of urban land [23].Urbanization is driven by many socio-economic and environmental processes and underlies different geographical boundary conditions, such as housing, migration, land use, infrastructure, economic activities and networks.Urban growth models represent the different mechanisms in a numerical way to support the understanding of past urbanization processes and project future scenarios.Therefore, an urban growth model can be seen as a representation of the complex interaction among a set of driving factors triggering urbanization that change across temporal (short-or long term) and spatial (local, regional, global) scales and organizational (municipal, provincial, national, supranational) levels [23].Two basic methodological mechanisms exist for modeling urban growth: Process-based approaches usually start from a set of rules defining urban growth derived from theoretical considerations and apply them to study cases [30].Data-focused approaches relate observed land cover changes to different factors driving urban growth, such as transportation arteries or terrain properties [31].Today, four major classes of urban growth models can be distinguished, which feature different combinations of the two mentioned mechanisms: cellular automata (CA), Agent-based (ABM), land use transport (LUT), and machine learning based approaches (MLA) [23,[32][33][34][35], which are explained in the following: 3.1.1.CA models CA models are often designed for regional case studies and are characterized by a high flexibility and simplicity [23].They use spatially adhering arrays of regular or irregular shaped cells and rely on a set of conversion rules derived from the analysis of historic urbanization patterns.In this regards, different urbanization rules and their spatiotemporal distributions are considered such as socio-economic developments (e.g. transportation networks, land use or economic activity) and geographical conditions (e.g.terrain) [23].The workflow in a CA model usually consists of the collection of data (usually remote sensing products), the analysis of the historic drivers of urban growth and the formation of transition rules using calibration methods [36].In CA models, the units are.CA models.The transition of a cell from other land use to urban area is determined in a probability-based approach, which implicitly assumes that decisions are made on the unit level.Each grid cell interacts with its neighbors via selforganization [37].A popular model using a cellular automat approach is the SLEUTH model (acronym for slope, land use, exclusion, urban, transportation, hillshade) which has been applied widely on various cities over the last years to simulate both urban and land cover changes on regional scale [38].

ABMs
ABMs form a family of process-driven computational models for small-scale applications that first appeared in the 1980s [32].This class of models simulates the actions and interactions of autonomous agents over space and time that trigger urban development and changes in the urban structures [39].Agents can thereby represent individuals or collected units such as residents, urban land developers or governmental institutions.Depending on their specific purposes, ABMs use a range of indicators for projecting urbanization patterns, such as (urban) land use types and densities, access to areas for recreation, infrastructure and public services [40].ABMs apply a series of synchronized simple operations representing the individual behaviors of the different agents forming holistic dynamics of the (usually complex) total system.In contrast to CA models with usually fixed grid structures, the neighboring units in ABM vary with time as the agents can transfer and interact among themselves and with their environment [32].Therefore, AMBs are often used to assess complex responses of agents to change, e.g.adaptation of households to climate change [41].The widely used NetLogo platform, for example, is an AB modeling environment which enables the modeling of urban growth due to the agents' responses to internal and external changes [42].

LUT models
LUT models typically use data-driven mechanisms applied on local to regional scales and use zones as spatial inputs.Similar to ABMs, LUT models consider different socio-economic developments such as population growth or economic growth as drivers of urbanization depending on their specific purpose [23].The models assume that land use and transportation are inseparably linked [43-45]: While land use defines the necessity for spatial interaction, transport expresses the land accessibility [46].The models make use of functions like: 'journey-to-work' and 'journeyto-shop' to allocate population among the residential zones in a workplace region.UrbanSim is a popular microeconomic model for LUT models, which relies on location choice for houses and organizations and the transportation network for simulating urban development [47-50].

MLA
A relatively new but increasingly popular approach in urban growth modeling is the group of machine learning based methods to simulate urban growth.Model applications often focus on a regional scale and simulate land use change and/or transitions of urban structures and densities.Usually, ML techniques are applied to identify relevant drivers of urbanization from a large set of different input parameters (e.g.socio-economic, legal, and environmental boundary conditions) and evaluate the transition potentials for urbanization within the study area based on the analysis of historic time-series data [51].Machine learning methods are usually combined with CA approaches to simulate the urbanization patterns based on the obtained suitability matrices, as methods solely using ML methods tend to show a lower performance [52].Different approaches exist, which couple CA with Artificial neural networks [51, 53, 54], or random forest [55], or support vector machines [56], for instance.Compared to the classical CA models using statistical methods, ML based models have the advantage of a higher flexibility in tackling the complexity of urban growth and interrelations between relevant drivers [57] but with less transparency on the process of urbanization.They also have a higher demand for input data of high quality and completeness to identify the relevant drivers.Therefore, a careful selection of the ML approach is recommended [58].

Application potentials for urban exposure modeling
Due to their distinctive characteristics, the different classes of urban growth models vary in their suitability and application potentials for assessing future exposure to environmental risks.The presented evaluation against the requirements for urban exposure simulation defined in chapter 2 is based on the analysis of recent review articles that assessed the stateof-the-art in urban growth modeling [23,36,40,52,59,60].
The classical CA model approach fulfills a set of characteristics, which are required for the applicability in exposure modeling: Due to the mixed dataand process driven approach, the models are usually capable of projecting urban growth for different time steps in longer timeframes considering potential changes in the drivers of urbanization [36].Recent developments include the simulation of densification processes [61] and the expansion from city scale up to national scale [62].Often, the assumption of units making decisions holds for modeling urban growth and exposure, particularly when the focus is on the dynamics of the physical aspects.Several techniques are available for the generation of scenarios such as the manipulation of the driving factors for urbanization or the suitability of land.CA Model approaches usually allow for the simulation of land use changes and can therefore consider developments in the hinterland as additional surface-sealing through future urbanization and their effects on the exposure.
Also, the ABM class meets many of the defined criteria for urban exposure simulation.The flexibility of the ABMs in incorporating a larger set of driving forces enables the projection of alternative future scenarios.This can be highly beneficial for assessing the effect of (un)desired policy strategies and urbanization dynamics on future exposure.A growing number of applications of ABMs exist, which addresses questions of urban morphology and densification ranging e.g. from the simulation of land use changes in general [63], societal changes in the future development of residential areas [64,65] or the formation of slums [66], for instance.However, ABMs very often do not have a generic structure, as they are designed for a specific study area, so adjustments are needed, when transferring or upscaling the model to other sites [67].
LUT models, in contrast, tend to be most beneficial for modeling urban development challenges related to transportation (e.g.travel time), rather than responding to the criteria for urban exposure modeling presented in section 2.
In most aspects, ML-based approaches meet similar applicability requirements for urban exposure simulations as the classical CA model approaches.However, a growing number of publications have already used ML techniques to simulate not only urban growth but additionally future changes in urban morphologies and densities [68][69][70].A second major difference to CA models concerns the transferability, as ML-based models have higher requirements in data quality, which poses a challenge for data-poor areas [51].
To sum up, CA models, agents based models and MLA meet many identified requirements for urban exposure simulation, although applied in different contexts and other research fields.(See also the overview presented in table 1).

Application scales and input data for urban exposure simulations
Today, mapping current and simulating future urban growth is applied across quite different spatial scales with different implications for urban exposure modeling: Most urban growth models are built to analyze trends and inform solutions on city level, so that approaches and applications are tailored towards regional or city scale assessment [71].CA and ABMs are often applied for such applications on local to regional scales.These models suit local exposure simulations especially with the need for high resolutions to consider e.g.hazards with high small-scale variations as floods, or landslides.However, applying these models on larger scales than the regional level quite quickly escalates the computational challenges and needs.
With the increasing availability of highresolution, global remote sensing data, large-scale urban growth modeling has come to focus leading a variety of global modeling assessments and enabling the applications of urban growth models even in data-scarce environments [72,73].Large-scale models are regarded as vital inputs for global questions on exposure and overview studies.However, these models entail limited information on local exposed elements prone to hazards with confined boundaries and high spatial uncertainty (e.g.river floods).
On all spatial scales, urban growth models typically must use and integrate different data set types, with different contents, data collection methods and temporal as well as spatial resolutions.Most models rely heavily on remotely sensed time-series information on land cover changes [55].Other frequently used data sets include, for example, population distribution or infrastructure networks.Datasets therefore often have diverging temporal and spatial resolution, e.g. between remote sensing and census data, which have to be adjusted by interpolation techniques.
When incorporating modeling results into the policy-making process, it is essential to assess the suitability of modeling approaches for the considered scale of application, transparently communicate uncertainties and ensure a correct interpretation of results [74].

Explicit studies on urban growth modeling in the assessment of future disaster exposure
The increasing awareness to link climate and disaster risks to urban planning and management has led to a rising number of studies exploring the potentials and challenges in the application of urban growth models for scenarios of future exposure.This chapter presents an analysis how far the current model approaches meet the twelve proposed criteria for modeling future trajectories of urban exposure towards natural and climatic hazards.For this purpose, a structured review of current literature was carried out that explicitly used on urban growth modeling for future disaster exposure.A total of 54 publications were identified and assessed in terms of applied modeling approach for future urban disaster exposure (see supplementary materials).The search generated 1095 peer-reviewed journal publications in English covering the time period of 1995-2021.In the next step, each publication was manually reviewed and crosschecked by a reviewing team in terms of the three criteria:

Methodological approach of the structured literature analysis
(1) whether it modeled the process of urban growth, development or expansion over time, (2) whether it linked urban growth to natural hazard exposure, and (3) whether it investigated the future urbanization trends and scenarios.
This selection process resulted in a total of 48 publications, which applied urban growth modeling to assess future exposure to natural hazards.A second review with a similar setup was carried out on 3 June 2022 to consider the latest advances since the publication of the IPPC report [1].This resulted in 6 further studies, which correspond to the presented criteria.

Geographical locations, investigated hazard types, and spatial scales
Figure 1 shows the distribution of studied areas and their scales of application of the 54 identified publications investigating future urban growth in the context of exposure scenarios.High-income and uppermiddle income countries are researched to the highest degree and within this group especially the United States of America: 9 of the 54 studies (17%) focus on cities, or regions that are located in the USA.In contrast, low-income countries, which often have highest vulnerabilities to environmental hazards, are the least studied category of countries.Only three studies located in Africa could be identified: Two large-scale studies on continental [75] and multinational scale [76] and one regional study on the city of Addis Ababa, Ethiopia [77].
Thereby, the main focus of the 54 studies is on future flood exposure: a total of 89% (48 n) of the identified publications are related to flood hazards (storm surge, pluvial and fluvial flooding, and sea level rise).37% of the studies (20 n) analyses the combined effects of changes in future hazards and urban expansion.Usually independently generated hazard maps are combined with urban growth projections [78][79][80].Several publications apply a chain of different models with the output of the urban expansion model serving as the input for the hazard model.To be more specific, the impact of urbanization on flood generation processes is usually assessed through landuse change projections that serve as input for hydrological or hydraulic models.The study of Huong and Pathirana [81] is a typical example, where the Dinamica EGO land-use change model [82] is applied to generate the input for a land surface processes model and a flood generation model.The different models are applied incrementally, which means that feedback loops of flood hazard changes on urban growth were neglected.For this reason, reciprocal influences between hazard change and urbanization are not fully explored.This would only be possible by a dynamically coupled modeling approach, which has not yet been investigated intensively so far within the literature on future urban exposure simulation.
One of the challenges in the context of flood risk studies is the determination of future surface permeability factors needed for the simulation of runoff-patterns.For urban areas, this problem is addressed by simulating densification processes, which were then translated into different categories of imperviousness [83].This suggests the potential of urban growth modeling for a more accurate projection of future flood hazard patterns.
Other types of natural hazards have been researched so far to a minor degree in combination with future urban growth.Only 22% of the studies (12 n) investigate future exposure to earthquakes (4 n), landslides, (3 n), forest fire (1 n), volcanic eruptions (1 n), and erosion (1 n) usually by utilizing official or independently generated hazard maps.
Four types of application scales can be categorized, when analyzing the sizes of the study areas: one third of the studies (38.9%,21 n) focused on urbanization and hazard change scenarios at the city or local scale.They limited their study areas to the cities or urban regions of interest.Typically, the extent of the study area is defined along the local administrative boundaries, which often allows for the direct use of administrative planning documents and the evaluation of local planning strategies [78,[84][85][86].
Several studies go beyond local administrative boundaries to additionally consider the influences of peri-urban developments and form the second group.For example, Khan et al [87] examine the whole metropolitan region of Dhaka, Bangladesh or Calderon and Silva [88] focus on the Metropolitan Area of Costa Rica to consider the strong urban-rural linkages in these regions.Usually, the focus of this kind of studies lies on the development of future urban growth scenarios.Future exposure is evaluated by existing hazard maps provided by governmental or other institutional authorities.
The third group that can be identified focuses on flood risk assessments on local to regional scale and consequently selected the study areas by the extent of investigated rivers basins due to requirements of applied hydrological models (26 n, 48.1%).In this group of studies, urbanization projections are needed as a driver of hazard intensities additionally to climate change.However, for larger river basins different, locally varying processes of urbanization need to be considered in the simulation of future urban growth patterns.Beckers et al [89], for instance, addressed this challenge using a statistical approach, which is driven by population projections on municipal scale to investigate the flood risk for residential areas in the Meuse river catchment in Belgium.In this way they projected a set of future scenarios considering the small-scale differences in the urbanization patterns of the 19 municipalities located along the investigated river.
Studies on large-scale extents forming the fourth group are the least investigated category: seven studies (13.0%) carried out assessments on multiple countries or continental scale and only one approach carried out a global-scale projection of exposure of urban areas [90].

Assessment of approaches against the requirements for urban exposure modeling
The analysis of identified literature against the criteria for future urban exposure modeling shows large gaps between the proposed requirements for urban exposure simulation and their implementation within existing modeling approaches of the identified literature: the twelve criteria for future urban exposure modeling are only partially fulfilled to different degrees by existing modeling approaches (see figure 2).Whereas the utilization of alternative scenarios and the consideration of land use change feedbacks, for instance, are already of widespread application, applied model approaches of urban growth consider densification processes or space-time trajectories into future projections only exceptionally.
Classical CA models and MLA are the dominant model types currently applied in future exposure studies (see figure 3).This is due to the comparably long tradition of CA models and the recent boom of machine learning approaches in urbanization modeling.In the next paragraphs the twelve criteria are evaluated also in terms of the adjustments of the applied models to project scenarios of future urban exposure.

Urban morphologies and densification processes
The simulation of urban morphology change (Criterion 1) and densification processes (Criterion 2) beside urban growth is not yet established in studies of exposure simulations, but 18 publications already integrate the simulation of the different characteristics of urban land into modeling approaches (33.3%).If urban morphologies are considered, most authors follow the concept of urban land use classes and distinguish between different urban forms as residential, commercial, industrial, or transportation structure types [86,[91][92][93].Urban morphology changes are projected by simulating the shifts between the different urban land-use classes in a probability-based approach.Also the modeling of change in building densities follows a similar approach.Building densities are thereby usually represented by discrete classes of low to high development, fractions or densities [20,81,83,94].
Several advanced approaches have combined the simulation of densification processes and urban land use change by the combination of the two urban characteristics [20, 75, 87, 92, 95-97] (7 n, 12.9%).Tierolf et al [20], for instance, subdivided the urban class into villages, suburban sparse, suburban dense, urban dense and industry.To project future transitions, they applied the model CLUMondo [98] which is used to simulate land use changes.CLUMondo is a typical representative of CA based models that simulates demand-driven land use changes and their feedbacks.
Despite these improvements, this level of detail is not yet sufficient for exposure modeling on local to regional scale.The residential class, for instance, covers a range of different building types ranging from multistory buildings to detached houses.Those building types, however, might go along with huge differences in their vulnerability profiles to natural hazards not only due to the constructions and materials but also the differences in socio-economic conditions of their residents.For this reason, an even deeper  differentiation is needed to increase the level of detail in risk assessment scenarios.

Time-frame and time-step information
The literature analysis shows that long term exposure pathways (Criterion 3) have not been in the center of attention in the research on future urban exposure simulation.Most publications focus on short and medium term projections ranging until 2030 or the middle of the century; only ten publications (18.5%) analyze periods extending the year 2050.The assessment of Criterion 4 'Time-step information' shows that 70.4% (38 n) of the studies evaluate only the end points of the projection time spans, although urban growth models are theoretically able to cover several time steps (see table 1).However, the short projection periods make an analysis of the development steps redundant.
The strong focus on shorter periods may also be reasoned in the applied modeling approach: the derivation of future urbanization drivers by the assessment of historic developments may not represent future urban growth trends to a full degree and miss important (future) factors of urban growth [99].Due to this, the simulation of the complex, multidependent, and non-linear processes of urbanization patterns with their high uncertainties poses a challenge for the projection of long-term trajectories that needs to be addressed.

Alternative scenarios and drivers of urbanization
The simulation of alternative future urbanization scenarios (Criterion 5) is the requirement, which is met to the to the highest degree by the analyzed studies: 30 out of the total of 54 publications (55.6%) have investigated different future urbanization scenarios that go beyond the basic scenario, which projects the continuation of historic growth.
The analysis further shows that the methodological approaches used to create scenarios for future urbanization are not coherent within the identified publications.Boundary conditions for scenarios deviating from the trends-to-be-continued pathway are usually chosen based on pragmatic approaches and data availability.As simulated future scenarios are not result from a standardized scenario creation approach, the explicit comparability of obtained results beyond the individual studies is not possible.The Shared Socio-Economic Pathways framework (SSPs), which would offer such a common scenario framework for future urbanization scenarios, was used only in two studies [19,100].Also participatory approaches have not been integrated into the generation of scenarios.Participatory scenario development approaches offer the integration of local knowledge and divers perspectives, capacity building and increase the usability of scenarios and modeling results by stakeholders.Only two studies mention the engagement of stakeholders in the development of scenarios and for validation purposes [92,101].
Three techniques of operationalizing scenarios and manipulating the different drivers of urbanization can be found in the analyzed studies (Criterion 6): the first approach is most straightforward and usually referred to as 'business as usual projection' .The parameterizations obtained from the analysis of historic growth patterns are kept and neither the influence of drivers nor the probabilities for urbanization are changed.This means that the past trends are consistently continued into the future.42.5% (23 n) of the studies have limited the future projections to this scenario.
The other frequently used approach is the manipulation of the transition maps that describe the spatial probabilities of land use change and the urbanization potentials.25.9% of the studies (14 n) use this approach to simulate scenarios with regulative or restrictive character as policy interventions on landuse.Examples include the implementation of master plans [102], the introduction of protection zones [79,103], or risk-sensitive planning scenarios and zoning regulations [91,96,102].Apart from this class of scenarios, suitability maps have also been used to depict the level of attractiveness for urbanization.Votsis [104] for instance, project future financial effects of sea level rise on the housing market and consequently on urbanization patterns by considering property prices via different classes of suitability for urbanization.
Studies manipulating transition maps usually assume the continuation of the past urbanization patterns, as the compositions and influences of urban drivers remain unchanged and the simulated interventions are from static nature.One outstanding approach in this matter is presented by Logan et al [105]: They directly extend the suitability map for urbanization by a tsunami awareness factor that is derived from the duration since the last tsunami event.The urban growth model is dynamically coupled to a tsunami model, which projects the appearance of tsunamis in the study area.As a tsunami triggers the awareness factor, the probability map expressing the suitability for urbanization is adjusted at each time step.In this way, different behavioral adaptation effects are considered in the urbanization patterns.
A considerable number of 17 studies (31.5%) exist using a third approach of operationalizing scenarios: they investigate changes in the speed and drivers of urbanization apart from land suitability.This means from the operational perspective that the parameters describing urban growth patterns are actively varied.One explored way is the variation the probability threshold of a pixel turning into urban area or the amount of areas being urbanized [76,94].Several studies go one step further: they dynamically simulate urbanization trajectories by using future projections of the drivers as population numbers or urban land demand [19, 88,89].However, the relationships between risks to natural hazards and urbanization patterns are not studied in a fully dynamical way.
One outstanding approach in this matter integrates socio-economic vulnerability profiles to flooding in their machine learning based modeling approach to simulate a scenario of urban growth and morphology changes adapted to the rising risk of sea level rise in 2030 and 2080 [93].(several approaches as the SLEUTH model have an additional model component that intrinsically stipulates or dampens the speed of urban growth [106].However, the authors do not count them into this category, if nature and composition of urban growth drivers remains unchanged).

Social and political decisions and adaptation measures
Criterion 7 'Social and political decisions' addresses the need for model approaches that reflect the effects of transformational processes in a society on the nature of urban growth through the introduction of new planning paradigms: three studies (5.6%) use scenarios of compact and sprawled growth patterns or high and low growth speed [87,94,107].However, the future political and societal boundary conditions, that facilitate those growth strategies, were not presented, so that they cannot be counted for this criterion.
A more structured approach is presented by the authors of Wolff et al [19], who simulated future urbanization scenarios for Mediterranean countries in Europe applying a ML based urban growth model.They derive future land demands based on the projected developments presented in the SSPs Framework.Future land use demand is then used to define the future degrees of newly urbanized areas.In this way, the simulations of urban growth are accompanied by the SSP narratives, which describe future developments of growth patterns with the corresponding socio-economic developments and population numbers.
One further notable study, which also used a comprehensive scenario approach, includes projections for biophysical and socio-economic boundary conditions as climatic conditions, population density or farm sizes.Demand-driven scenarios of urbanization are simulated in business as usual, liberalization and self-sufficiency futures using the Dyna-CLUE model [108].
While societal processes and political decisions are not in the focus of analyzed studies, the investigation of adaptation strategies (Criterion 8) is much more researched with 26.0% of the studies (14 n).The most commonly evaluated measures are zoning practices and the exclusion of areas at risk.These scenarios are usually applied in studies of regional scale and operationalized by the adjustment of the suitability map or the exclusion layer according to current or future hazards maps [84,96,102,103].
Hemmati et al [91] applies a different planning scenario based on the implementation of risksensitive socio-economic incentives.The authors assume that the construction of infrastructural facilities in risk safe areas as schools and entertainment centers leads to a shift of future urbanization patterns into regions exposed to lower risks.For this purpose, a CA model is used, which integrates access to infrastructure as a driving factor of urbanization patterns.Votsis [104] also simulates the effect of the adaptation of the settling patterns due to market adjustments by the coupling of the flood risk level to property price.This scenario is operationalized by an adjustment of the probability map.
Logan et al [105] present the only approach, which considers behavioral effects on urbanization patterns driven by the experience of hazard occurrences.The authors follow the assumption that with increasing experience of a hazard event, vulnerability decreases due to learning and adaptation effects of a community.This means in practical that urbanization is less likely in those areas, where the inhabitants have a higher awareness due to the recent experience of a tsunami.To analyze the effects of the two different strategies of adaptation measures, a risk awareness component was integrated into a CA based urban growth model: on top of the classical determination of urbanization probabilities, a hazard awareness factor was included, which diminishes over time with the increasing duration since the last tsunami event.Two different adaptation strategies to tsunami risk are then analyzed: the installation of seawalls with different heights and the raise of community awareness.Modeling results showed in the long-term perspective a higher effectivity of soft-adaptive measures, which keep the awareness of community at a higher level.
Only one study presents an analysis of the mutual or exclusive effects of nature protection, mitigation and adaptation practices on future exposure on a national scale [92].A set of ten land userelated adaptation measures was simulated such as the stricter protection of open space, the preservation of urban green spaces or densification of urban areas.For this purpose, the land Use scanner model is applied, which is a optimization based land-use change model.An indicator-based impact analysis shows the partly contradicting effects of climate change mitigation and adaptation measures.This study stands out as flood risk adaptation measures and their impacts on urbanization patters are evaluated from a more holistic perspective, which also includes the mutual influences on climate mitigation and nature protection goals.

Land-use change
The modeling of land-use change and feedbacks between rural and urban areas belongs to the already established requirements for exposure assessments (Criterion 9): 25 out of 54 studies (46.3%) consider land-use change in the simulation of future exposure.This may be reasoned in the fact that a high share of publications focusses on the investigation of future exposure to fluvial and pluvial flooding.Typical representatives of this kind of studies investigate the influence of future urbanization on future flood areas using hydrological or rainfall-runoff models, which require comprehensive land-use change scenarios of the entire river basins [79,89,93,94,103,107].Due to a high number of established LULC models, very often CA or ML based approaches are applied that were not explicitly developed for urban growth modeling, such as Dyna-CLUE [109] or land transformation model [110].
Land-use change is a complex phenomenon which is driven by more influence factors that are different in their nature or effects compared to those relevant for urbanization.The analysis of the 25 publications shows that applied scenarios often lack of a description or alignment of the additional drivers that are relevant for other land-use changes apart from urbanization.However, one outstanding study in this aspect is presented by Szwagrzyk et al [108], who apply three scenarios, which consider a variety of biophysical and socio-economic drivers of the future potential developments of each of the land use classes and their interdependencies, such as slope, soil conditions, distance to road infrastructure, population density, land ownership, or average farm sizes [111].

Space-time trajectories, geographical scales and transferability
So far the modeling of space-time trajectories of exposed objects (Criterion 10) has not been researched in the context of future exposure simulation.Only Logan et al [105] consider this criterion by investigating the effect of different adaptation strategies to tsunami risk as installing seawalls and raising community awareness on urbanization patterns.Applying a CA based modeling approach with an integrated awareness component, the authors can show that a high awareness level from hazard experience impedes the population from settling in riskprone areas.With the temporally decreasing awareness of the last hazard event, areas, which are also at higher risk, become increasingly settled.If those areas are then affected by a tsunami, they are 'demolished' and the land falls back into the undeveloped class and areas at lower risk become more attractive again.This example can be interpreted as a first step towards the consideration of the spatiotemporal shifts of exposed objects.In this way, it shows the principal suitability of CA-based approaches for the modeling of spacetime trajectories.
Also Criterion 11-the geographical scalabilityhas not been researched intensively (3.7%, 2 n): studies on urbanization scenarios in the context of future exposure usually focus on one geographical scale.The interdependencies between supra-national, national, regional and local drivers are partly covered in the narratives describing the boundary conditions of applied scenarios.However, nested approaches which enable the consistent simulation of global to local scale urban growth scenarios under changing climate conditions, have not yet been investigated.One identified approach might point into this direction: Wolff et al [19] applies an urban growth model individually for each country of the whole Mediterranean region.Assuming coherent scenarios for each model run enables to merge the individual projections to one overarching urban growth scenario.
Beckers et al [89] go one step further: they project flood risk under different scenarios for a river basin covering 19 municipalities.To reflect the different developments patterns and velocities within the region, they use a regional model and apply a statistical approach downscaled to the municipal level.In this way they are able to consider also small-scale developments on a higher spatial level.
Also the transferability of urban growth models to different physical and socioeconomic environments (Criterion 12) is addressed and discussed only in 9 studies (16.7%), mostly in the form of limitations on certain fields of application or general recommendations.However, the robustness of an urban growth model to varying boundary conditions can also be indirectly demonstrated over time by the successful application of the model to multiple study areas with different boundary conditions.Nevertheless, the analysis of this study clearly shows the low level of research activity in low-income countries, which makes these regions a blind spot when assessing the general transferability of the model in this way.

Future research agenda
In order to support the planning of adaptation strategies in cities, scenarios on future exposure trends are urgently required.Such scenarios should consider the prospective trends of hazards and urbanization driving urban exposure, in an integrated manner.The review has shown that urban growth models offer the potential to address these questions.They can help to overcome the current mismatch where assessments of future urban risk trends are superimposing future hazards trends (e.g.flooding or sea-level rise) with current patterns of urban extent, morphology socio-economic characteristics.Urban growth models are therefore essential for exploring and strategically assessing urban exposure trajectories in the future.Especially process-based approaches such as CA models and machine learning based methods have already shown the general suitability in this regard.However, the review showed that the majority of existing future urban exposure studies was conducted at different scales in high income countries, while low and lower-middle income countries remain under-researched.This is in stark contrast to the most recent population and urbanization prospects which project most rapid urban growth in low and lowermiddle income countries, particularly in Africa and Asia [13]; countries and regions which at the same time often are also amongst the most vulnerable to environmental hazards.
The review also shows that due to the high level of complexity and the large number of processes that must be taken into account for the projection of future urban exposure, it will probably not be possible to integrate all of the 12 criteria from chapter 2 into one specific modeling approach, even in the long term.Instead, the different strengths of the four urban growth model approaches should be taken into account in the integration of the requirements.However, in order to tap the full potential of urban risk and adaptation science, the next generation of urban growth models needs to be advanced in several ways: First, models need to be able to project not only urbanization trends, but also to describe the nature of (future) urbanized areas.Whilst urban growth models typically provide information on whether and when a particular spatial entity (e.g. a grid cell) turns from non-urban into urban land use, models to date are poorly equipped to provide information on the types of urban land use and their accompanied social and economic vulnerability characteristics.For disaster risk and climate change adaptation science, it makes a significant difference whether an'urban cell' on a map represents, for instance, a high-income neighborhood, an informal slum, an industrial estate, or an office complex.In other words, urban growth models need to be expanded to provide information and assess feedback with processes of socio-economic change, such as gentrification or industrialization.
Second, an increasing focus on projecting timeframes that go beyond short-term planning scenarios is recommended.A scenario calculation running up to the end of the century holds great value for analyzing long-term trends and raising awareness.Hence, models need to be designed and presented in a way to provide various time-steps that adhere to practical decision-making processes but also include long-term trends that affect urbanization patterns.
Third, urban expansion models need to incorporate new drivers of urbanization as feedback with climatic or other environmental changes and their impacts.Climate change will progress over the next decades and its impacts will increasingly become visible.A key research question for models will therefore be to explore whether, when and how highly exposed cities will see changes-potentially reductions-in their growth and migration patterns, due to the very impacts of climate change.
Fourth, models must be made capable of integrating the effects of societal trends and adaptation responses on urbanization patterns.The drivers of urbanization that were crucial in the past may no longer have the same importance in the future, and new forces may even emerge that influence future land use.For this reason, also adaptation responses might be triggered in different forms.This can include, for example, digitalization of society, which might reduce the pressure on urban areas, or autonomous adaptation decisions, such as the outmigration of leading economic players, which can trigger migrations of entire economic clusters.
Fifth, they also need to integrate the large variety of adaptation practices more strongly from local to regional and global actors apart from zoning practices and restrictions.Indirect measures such as information campaigns or learning from neighbors, for instance, might play a huge role in shaping individual and eventually collective adaptation decisions in the future.Urban growth models need to be (further) adjusted in a way that these adaptation practices and their effects on future exposure can be evaluated under different boundary conditions.ABMs are suitable tools to simulate these processes in a direct way, but also CA models can integrate them via the manipulation of the probability layers for urbanization.
Sixth, models should increasingly be designed and used to explore the full range of potential future trajectories in urban risk and adaptation, using for example the SSP architecture.The establishment of a coherent scenario framework would support the comparability of results across scales and regions and allow for a better understanding of developments specific to the study areas.Models therefore need to be tweaked to also speak to hitherto not considered directions of change, e.g. the retreat of existing cities from coastal zones through de-urbanization.

Conclusions
Along with climate change, urbanization is one key driver of risk to natural hazards.While past research activities have focused on assessing future hazard trends under climate change conditions, there has been less focus on examining future urbanization pathways and their contribution to environmental risks.Although the field of urbanization modeling is already established in many application areas, the use of urban growth models in future exposure and risk assessments is a relatively new development.To assess the current state of the art and future research needs in the use of urban growth models for urban exposure simulation, a structured literature review was conducted.The results indicate that there is a need to consider and integrate (1) urban morphology and density patterns and linkages to vulnerability (2), time horizons, which consider long-term effects on urbanization (3), feedbacks between hazard impacts and urbanization trajectories (4), the integration of future urban drivers and adaptation responses (5), urbanization effects of adaptation practices as collective and individual behavioral change, and (6) scenarios, which are developed within a commonly defined scenario framework to harness the potential of urban growth modeling with regard to the assessment of future urban exposure and risk to environmental hazards.
With the suggested improvements, urban growth models have the much-needed potential to become an essential tool not only in the assessment of future risk trajectories, but also for the evaluation of adaptation strategies and their consequences.

ORCID iDs
A structured literature analysis was carried out in the database Web of Science Core Collection on 20 September 2021.The following query function was used to search in the TOPIC (TS) of publications, which includes title, abstract and keywords: (TS = ('urban growth') OR TS = (urbanization) OR TS = ('urban development')) AND (TS = (exposure) OR TS = (risk) OR TS = (vulnerability) OR TS = (resilience)) AND (TS = (model * ) OR TS = (simulat * )) AND (TS = (future) OR TS = (scenario))

Figure 1 .
Figure 1.Locations of subnational and local studies and scales of application of the identified 54 publications on future urban exposure modeling.

Figure 2 .
Figure2.Evaluation of the criteria for urban exposure simulations ( * number of the publications that have not necessarily carried out the integration of space-time trajectories, geographical scales and the transferability of approaches but at least discuss the topics).

Figure 3 .
Figure 3. Distribution of the applied model approaches applied by the 54 studies presenting future exposure scenarios with CA: cellular automata, AB: agent-based, LUT: land use transport, MLA: machine learning based approach.

Table 1 .
Evaluation of the model type characteristics and general model suitability with regard to the twelve identified criteria for future urban exposure modeling.
++ standard use; + first use cases; o no applications so far.
Chaudhuri G and Clarke K 2013 The SLEUTH land use change model: a review Environ.Resour.Res. 1 88-105 [39] Abar S, Theodoropoulos G K, Lemarinier P and O'Hare G M P 2017 Agent based modelling and simulation tools: a review of the state-of-art software Comput.Sci.Rev. Zhou B, Kockelman K M and Lemp J D 2009 Applications of integrated transport and gravity-based land use models for policy analysis Transp.Res.Rec.2133 123-32 [44] Miller E J, Douglas Hunt J, Abraham J E and Salvini P A 2004 Microsimulating urban systems Comput.Environ.Urban Syst.28 9-44 [45] Mishra S, Ye X, Ducca F and Knaap G J 2011 A functional integrated land use-transportation model for analyzing transportation impacts in the Maryland-Washington, DC Region Sustainability 7 60-69 [46] Wegener M 2021 Land-use transport interaction models Handbook of Regional Science ed M M Fischer and P Nijkamp (Springer) pp 229-46 [47] Waddell P U 2002 Modeling urban development for land use, transportation, and environmental planning J. Am.Plan.Assoc.68 297-314 [48] Waddell P 2011 Integrated land use and transportation planning and modelling: addressing challenges in research and practice Transp.Rev. 31 209-29 [49] Wang D and Yuan C 2018 Modeling and forecasting household energy consumption and related CO 2 emissions integrating UrbanSim and transportation models: an Atlanta BeltLine case study Transp.Plann.Technol.41 448-62 [50] Jin J and Lee H Y 2018 Understanding residential location choices: an application of the UrbanSim residential location model on Suwon, Korea Int.J. Urban Sci.22 216-35 [51] Shafizadeh-Moghadam H, Asghari A, Tayyebi A and Taleai M 2017 Coupling machine learning, tree-based and statistical models with cellular automata to simulate urban growth Comput.Environ.Urban Syst.64 297-308 [52] Chaturvedi V and de Vries W T 2021 Machine learning algorithms for urban land use planning: a review Urban Sci. 5 68 [53] 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 GISci.Remote Sens. 54 283-304 [54] Khan A and Sudheer M 2022 Machine learning-based monitoring and modeling for spatio-temporal urban growth of Islamabad Egypt.J. Remote Sens. Space Sci. 25 541-50 [55] Gamba P and Lisini G 2013 Fast and efficient urban extent extraction using ASAR wide swath mode data IEEE J. Sel.Top.Appl.Earth Obs.Remote Sens. 6 2184-95 [56] Rienow A, Mustafa A, Krelaus L and Lindner C 2021 Modeling urban regions: comparing random forest and support vector machines for cellular automata Trans.GIS 25 1625-45 [57] Aburas M M, Ahamad M S S and Omar N Q 2019 Spatio-temporal simulation and prediction of land-use change using conventional and machine learning models: a review Environ.Monit.Assess.191 205 [58] Shafizadeh-Moghadam H, Asghari A, Taleai M, Helbich M and Tayyebi A 2017 Sensitivity analysis and accuracy assessment of the land transformation model using cellular automata GISci.Remote Sens. 54 639-56 [59] Brown D G, Verburg P H, Pontius R G and Lange M D 2013 Opportunities to improve impact, integration, and evaluation of land change models Curr.Opin.Environ.Sustain. 5 452-7 [60] Tong X and Feng Y 2020 A review of assessment methods for cellular automata models of land-use change and urban growth Int.J. Geogr.Inf.Sci.34 866-98 [61] Saxena A and Jat M K 2020 Land suitability and urban growth modeling: development of SLEUTH-suitability Comput.Environ.Urban Syst.81 101475 [62] Martellozzo F, Amato F, Murgante B and Clarke K C 2018 Modelling the impact of urban growth on agriculture and natural land in Italy to 2030 Appl.Geogr.91 156-67 [63] Hosseinali F, Alesheikh A A and Nourian F 2013 Agent-based modeling of urban land-use development, case study: simulating future scenarios of Qazvin city Cities 31 105-13 [64] Jokar Arsanjani J, Helbich M and de Noronha Vaz E 2013 Spatiotemporal simulation of urban growth patterns using agent-based modeling: the case of Tehran Cities 32 33-42 [65] Myers D and Gearin E 2001 Current preferences and future demand for denser residential environments Hous.Policy Debate 12 633-59 [66] Patel A, Crooks A and Koizumi N 2018 Spatial agent-based modeling to explore slum formation dynamics in Ahmedabad, India GeoComputational Analysis and Modeling of Regional Systems (Advances in Geographic Information Science) ed J C Thill and S Dragicevic (Springer) pp 121-41 [67] Valbuena D, Verburg P H, Bregt A K and Ligtenberg A 2010 An agent-based approach to model land-use change at a regional scale Landscape Ecol. 25 185-99 [68] Chen Y 2022 An extended patch-based cellular automaton to simulate horizontal and vertical urban growth under the shared socioeconomic pathways Comput.Environ.Urban Syst.91 101727 [69] Kim Y, Safikhani A and Tepe E 2022 Machine learning application to spatio-temporal modeling of urban growth Comput.Environ.Urban Syst.94 101801 [70] Xu X, Zhang D, Liu X, Ou J and Wu X 2022 Simulating multiple urban land use changes by integrating transportation accessibility and a vector-based cellular automata: a case study on city of Toronto Geo-Spatial Inf.Sci. 25 439-56 [71] Taylor P and Derudder B 2015 World City Network: A Global Urban Analysis (Routledge) [72] Linard C, Tatem A J and Gilbert M 2013 Modelling spatial patterns of urban growth in Africa Appl.Geogr.44 23-32 [73] Zhou Y, Varquez A C G and Kanda M 2019 High-resolution global urban growth projection based on multiple applications of the SLEUTH urban growth model Sci.Data 6 34 [74] European Commission, Joint Research Centre 2023 Using models for policymaking (LU: Publications Office of the European Union) (available at: https://data.europa.eu/doi/10.2760/545843) [75] Paul N, Silva V and Amo-Oduro D 2022 Development of a uniform exposure model for the African continent for use in disaster risk assessment Int.J. Disaster Risk Reduct.71 102823 [76] Aich V, Liersch S, Vetter T, Fournet S, Andersson J, Calmanti S, van Weert F H A, Hattermann F F and Paton E N 2016 Flood projections within the Niger River Basin under future land use and climate change Sci.Total Environ.562 666-77 [77] Beshir A and Song J 2021 Urbanization and its impact on flood hazard: the case of Addis Ababa, Ethiopia Nat.
rules: a case study in Qingpu-Songjiang area of Shanghai, China Stoch.Environ.Res.Risk Assess.30 1387-400 [36] Aburas M M, Ho Y M, Ramli M F and Ash'aari Z H 2016 The simulation and prediction of spatio-temporal urban growth trends using cellular automata models: a review Int.J. Appl.Earth Obs.Geoinf.52 380-9 [37] Yeh A G O, Li X and Xia C 2021 Cellular automata modeling for urban and regional planning Urban Hazards 109 1167-90