Advanced optimization and data modeling techniques to improve accesibility and thermal comfort in urban plannning

University of the Basque Country UPV/EHU Doctoral Thesis Advanced Optimization and Data Modeling Techniques to improve Accessibility and Thermal Comfort in Urban Planning Author: Iñigo Delgado Enales Supervisors: Prof. Dr. Javier Del Ser Dr. Patricia Molina Costa A Thesis submitted in fulfillment of the requirements for the degree of Doctor of Philosophy in the Department of Communications Engineering October 28, 2024 (cc) 2024 Iñigo Delgado Enales (cc by-nc-nd 4.0) iii “The more I learn, the more I realize how much I don’t know.” Albert Einstein v UNIVERSITY OF THE BASQUE COUNTRY UPV/EHU Abstract Bilbao Faculty of Engineering Department of Communications Engineering Doctoral Degree Advanced Optimization and Data Modeling Techniques to improve Accessibility and Thermal Comfort in Urban Planning by Iñigo Delgado Enales In the 21st century, urban areas have established themselves as the most common form of human settlement, far outnumbering rural areas in terms of population. The trend for cities is that their population will continue to grow, until they account for 70% of the world’s population in 2050. The strategic role that cities play and will play for society in the future is so great, that since 2015, this importance is reflected in one of the 17 Sustainable Development Goals (SDGs) of the United Nations, which aims to make cities more inclusive, safe, resilient and sustainable. The impact of cities and the decisions made when planning them, not only affect the urban area in question but also have a global impact. One example of the wide impact of urban planning is its contribution to pollution in urban areas, further exacerbating climate crisis. Achieving a more sustainable city, that is both resilient and accessible to its inhabitants, depends on proper urban planning. Urban planners usually face complex problems, as several factors come into play. This is why decision support systems play an important role in facilitating the planning and design of more sustainable and livable cities. These systems, which are based on digital tools such as Geographical Information System (GIS) or 3D modeling, can help to analyze and visualize data, evaluate different scenarios and strategies, and foresee the possible consequences of urban planning decisions. In the last decades, within the new digital era with more access to loads of data sources, Artificial Intelligence (AI) is postulated as the new leading support tool in urban-decision making processes. This Thesis explores the possibilities of different AI techniques for urban planning from different perspectives. In the first part of the Thesis, accessibility and urban comfort are addressed. Specifically, the Thesis proposes a new decision-support system based on multi-criteria optimization combined with a georeferenced graph model. The framework makes it possible to design urban projects that improve accessibility and reduce exposure to noise and/or air pollution vi through the installation of urban elements (ramps, escalators and so forth) taking into account the total economic cost of the installation. In this way, urban planners are provided with a new tool that takes into account the simultaneous consideration of different objectives in an urban planning problem, thus facilitating the decision on where to undertake cost-effective actions in public spaces. The second part of the Thesis examines modeling problems that focus on the estimation of heat stress in cities due to climate change. In this second research contribution, the usefulness of image segmentation techniques for correlating spatial and meteorological data with street-level air temperature is explored. The ability to estimate temperatures with a high degree of accuracy allows for the identification of the highest priority areas in cities where urban improvements need to be made to reduce thermal discomfort. The air temperature at street level is estimated both spatially and temporally for a specific use case, and compared with existing, well-established numerical models. Based on the obtained results, neural networks are postulated as a faster and less computationally expensive alternative to numerical modeling counterparts. Overall, the outcomes derived from this Thesis contribute to the development of an embryonic research area –AI-based urban optimization and modeling– that in the future will enjoy relevant importance. The contributions herein reported expose with evidence the suitability of AI techniques for effectively solving urban planning problems that underly the design of safer and friendlier city ecosystems. vii Acknowledgements When I finished my physics studies at the UPV/EHU, back in 2017, I never imagined that I would end up writing a Thesis on Artificial Intelligence, a field that was totally unknown to me at those days. Thanks first of all to TECNALIA for trusting me, and in particular, to my two supervisors who have accompanied me during these 4 years, Prof. Javi Del Ser and Dr. Patricia Molina. Thanks also for their continuous support, patience, and guidance throughout my research. Their insightful feedback and encouragement have been invaluable. Special thanks also to all my work colleagues I have had in these four years of Thesis. First of all, to my colleagues at JRL-A2I, who helped me in the first two years of my Thesis when Covid still made social interaction difficult. Later, when I moved to TECNALIA I found a wonderful working group. Thanks to all the Core IA colleagues for providing support, collaboration and help when I needed it. Your camaraderie has made achieving this challenge easier and more enjoyable. Finally, thanks to all my family. Thanks to my grandparents, specially to my grandparents aitite y amama for taking care of me when I was little as if they were my parents. Thanks to my aunts and uncles. Thanks especially to my parents Elisa and Salvador and to my sister Ane. Thanks to my parents for instilling in me the values of effort, sacrifice and work, without which I could never have reached where I am. Thanks to my sister for always taking care of me, putting up with me and supporting me. Thanks also to all my friends, which are part of my family too. And last, but not least, thanks to my life partner Garazi which has always been there in good and bad moments and has being a source of energy and happiness for me. Love you all. ix Contents Abstract v Acknowledgements vii 1 Introduction 1 1.1 Motivation ........................... 3 1.2 Objectives ............................ 4 1.2.1 Outline and Contributions of the Thesis ....... 5 Chapter 2: Theoretical Framework and Literature Review .................... 5 Chapter 3: Improving Urban Accessibility and Environmental Conditions using Graph Modeling and Multi-objective Optimization . . 5 Chapter 4: Efficient Estimation of Ground-Level Air Temperature in Urban Areas using Machine Learning ................... 6 Chapter 5: Concluding Remarks and Future Research 6 1.2.2 Reading this Thesis .................. 6 2 Theoretical Framework and Literature Review 9 2.1 Theoretical Framework ..................... 9 2.1.1 Optimization Problems ................ 11 Multi-Objective Optimization ............. 13 Non-dominated Sorting Genetic Algorithm II . . . . 14 Non-dominated Sorting Genetic Algorithm III . . . . 17 Strength Pareto Evolutionary Algorithm 2 ...... 18 2.1.2 Modeling Problems ................... 20 Convolutional Neural Networks ............ 20 Encoder-Decoder Architecture ............ 22 2.1.3 Simulation Problems .................. 24 2.2 Literature Review ....................... 25 3 Improving Urban Accessibility and Environmental Conditions using Graph Modeling and Multi-objective Optimization 31 3.1 Related Work .......................... 33 3.1.1 Accessibility and Environmental Quality for Agefriendly Cities ...................... 33 3.1.2 Multi-criteria Design and Multi-objective Optimization ........................... 34 xvi 4.6 Time series of Tafor different LWT days, for all the test locations. The values measured by the meteorological stations are represented by the black line. The Tavalues obtained by the UrbClim model are in red, whereas the ones estimated by U-Net model in green. ................... 88 4.7 Temporal progression of 24 hours for the relative temperature (T𝑟𝑒𝑙 𝑎) of the area of Galindo (top). Location of the hotspots in the area (bottom). ................ 90 4.8 Temporal progression of 24 hours for the relative temperature (T𝑟𝑒𝑙 𝑎) of the area of Arrigorriaga (top). Location of the hotspots in the area (bottom). .............. 91 xvii List of Tables 3.1 Values of the parameters configured for the experiments. . . 53 3.2 Statistics (mean±standard deviation) of the quality indicators obtained by every algorithm over use cases 𝐴1and 𝐴2.57 3.3 𝑝-values resulting from a two-tailed Wilcoxon rank sum test performed over the quality indicator values obtained by every pair of algorithms over use cases 𝐴1and 𝐴2........ 57 4.1 Regression metrics computed between the two models (UrbClim and U-Net) for each weather station during the period of study. ............................. 82 4.2 Regression metrics computed for the two models (UrbClim and U-Net) with respect to the real temperature data collected by each weather station during the period of study. Best results for every station and score are highlighted in light blue. ............................ 82 xix List of Abbreviations General Terms UN United Nations DSS Decission Support System PSS Planning Support System AI Artificial Intelligence GIS Geographic Information System IGN Instituto Geografico Nacional DTM Digital Terrain Model Algorithmic approaches ML Machine Learning DL Deep Learning ANN Artificial Neural Network CNN Convolutional Neural Network DNN Deep Neural Network EC Evolutionary Computation EA Evolutionary Algorithm GA Genetic Algorithm MOEA Multi-Objective Evolutionary Algorithm NSGA Non-dominated Sorting Genetic Algorithm SPEA Strength Pareto Evolutionary Algorithm MOCELL Multi-Objective Cellular genetic algorithm ReLU Rectified Linear Unit Performance Metrics GD+ Generational Distance plus IGD+ Inverted Generational Distance plus HV Hypervolume indicator EP Epsilon indicator PCC Pearson Correlation Coefficient MAE Mean Absolute Error MAPE Mean Absolute Percentage Error RMSE Root Mean Squared Error Chapter 3 ABMS Agent-Based Modeling and Simulation ACO Ant Colony Optimization xx GIV Green Index View MCA Multi-Criteria Analysis Chapter 4 TaStreet-Level air temperature UHI Urban Heat Island LST Land Surface Temperature UrbClim Urban Climate LWT Local Weather Types NDVI Normalized Difference Vegetation Index 1 Chapter 1 Introduction Since the beginning of the industrial revolution, the proportion of the population living in urban areas has grown steadily. In the last 50 years, this growth has even accelerated due to rural exodus, particularly on the Asian and African continents [1], [2]. According to the United Nations (UN), 55% of the world’s population lived in cities in 2019, while a share of 70% is expected by 2050 [3]. This trend of urban population growth brings with it various challenges, such as environmental issues and problems with mobility and accessibility due to poorly planned urban areas. Furthermore, cities not only influence their immediate surroundings but also have farreaching global effects, such as their contribution to the current climate crisis [4]. Faced with this strategic global role and the importance that cities are adopting, in recent decades urban planners have come up with proposals for new models of cities. A clear case of this is the so-called age-friendly city concept [5], which refers to urban areas that are designed taking into account the needs of older people, improving their quality of life and independence. The key factor is to keep older people as active as possible in their daily lives. To this end, these cities prioritise improving the health, safety and participation of older people. Another concept for city that has gained relevance in the XXI century is the concept of resilient city [6]. A resilient city is one that possesses the capacity to absorb, recover, and prepare for future shocks, whether they be economic, environmental, social, or institutional [7]. Following this trend, and with a view to the future, the UN has introduced sustainable cities as one of the goals to be achieved under the UN Sustainable Development Goals [8]. Sustainable cities are defined by the UN as cities that are resilient, safe and sustainable. This sustainability has three main backbones, commonly called “the three pillars of sustainability" which are: social equity, economic viability, and environmental protection [9]. These three characteristics are not independent and are intertwined with each other, as depicted in Figure 1.1. This Thesis addresses the sustainability of cities from the environmental and social side, with the challenge of reducing the environmental footprint of cities, as well as improving their resilience while ensuring their social inclusivity. Several critical factors play a pivotal role in determining 2Chapter 1. Introduction Sustainable cities Environment Economy Social Figure 1.1: The three pillars of sustainability. the sustainability of cities. These factors intersect and collectively contribute to creating resilient, livable urban environments. Some of these key aspects are now presented: •Air and noise pollution: Cities are facing environmental challenges derived from the fossil fuel dependent economy, as well as from expanded urban models that require high levels of commuting for daily activities. Designing cities for the use of private vehicles means that public spaces are flooded with large numbers of motor vehicles, leading to accidents, noise and air pollution. Cities currently emit 36% of total annual CO2 emissions [4]. Further along this line, besides CO2, vehicles also emit carcinogenic substances such as nitrogen oxides [10], leading to poor air quality in cities and prejudicial impact on human health. •Accessibility and walkability: For various reasons (e.g., need for building within short periods of time, lack of resources) the rapid and steady growth of urban areas has given rise to poorly planned urban areas that today leads to accessibility problems in cities. Safe cities for pedestrians increase accessibility in cities by encouraging walking to the detriment of motor vehicles. In addition, having an accessible urban area for all citizens promotes social equity among them. •Thermal comfort: Cities face several challenges in their way to be a more livable spaces. Among those obstacles arises the Urban Heat Island (UHI) phenomenon that together with global warming has increased sharply the air temperature and the annual heatwave days of our cities, becoming less livable [11]. The UHI can be defined as the difference of temperature between an urban area and the surrounding rural area, building on the fact that urban areas are usually warmer than rural 1.1. Motivation 3 environments, especially at night [12]. This leads to cities not cooling down during the night, resulting in higher thermal stress during the day. This stress hinders the thermal comfort of their inhabitants. These factors are not always independent of each other, and sometimes solutions must go through multi-criteria planning processes, forcing urban planners to adopt a more holistic perspective when making decisions. With the advent of computational tools, planners now have access to various digital systems that support their decision-making processes, both in the short and the long term. Among these tools, Decision Support Systems (DSS) and Planning Support Systems (PSS) have emerged as valuable assets. They contribute to the planning and management process, enhancing the quality of outcomes, even though their adoption remains relatively low. Dynamic data from diverse sources should inform these decisions, and their impact should be continuously evaluated to rectify any unintended consequences. Over the past few decades, disruptive tools for data processing and acquisition—such as sensors, the Internet of Things (IoT) and other smart technologies—have reshaped the urban landscape. The Smart City concept, as forged by Giffinger in 2007 [13], has gained prominence. As cities become increasingly digitized, there is ample opportunity for the adoption of advanced technologies like AI and DSS tools. In particular, and focusing on AI, the advantages of this technology are yet to be exploited both as long-term (PSS) and short-term (DSS) planning tools. Although there are some scarce examples of their use in DSS (e.g.,[14]–[18]), and PSS ([19], [20]) the exploitation of these technologies for the aforementioned purpose is still in an embryonic stage. Therefore, great potential remains untapped in the use of new technologies to support DSS and PSS and related processes in public policies [21]–[23], including urban planning. In recent decades, both private companies and public entities are already using DSS and PSS frameworks endowed with avant-garde technological functionalities like AI. Nevertheless, there is still a long way to go towards improving further AI application in urban planning and management processes. 1.1 Motivation AI has progressively become an essential part of our lives over the last decades. Indeed, AI prevails as a technological driver that has fueled the evolution and progress of many application domains, yielding manifold benefits for both the economic prosperity and the society as a whole. For instance, in medicine and healthcare the latency and accuracy of diagnostic processes have improved [24], [25]. Physical sciences have also harnessed advances in AI: in the field of quantum physics, AI has reduced calculation times and has offered radically new approaches for solving complex many-body systems [26]. Similarly, traditional productive sectors such as industrial manufacturing have observed disruptive changes in their inspection and maintenance processes as a result of data-based prognosis, 4Chapter 1. Introduction adopting Machine Learning (ML) models at its core [27]. The technological momentum of AI is reflected by the strategic technological development of important companies such as Google, Apple or Microsoft, all of which embrace AI as a nuclear competence of their projects and value offer. Current prospects estimate a global Gross Domestic Product (GDP) increase from $75 trillion in 2016 to approximately $114 trillion by 2030: this estimate is expected to be 14% higher as a result of the prevalence of AI in industrial manufacturing and logistics [28]. In contrast to the acknowledged maturity of AI-based solutions in productive processes, the adoption of this key technology in urban planning and management has gone at a significantly lower pace. The potential of AI as an informing tool for experts during urban decision-making processes is yet to be tapped. Nonetheless, urban planners are starting to use AI-supported DSS [21], [29] and PSS [30], [31] in order to find the most suitable solutions to the complex problems undergone by cities in modern times. However, despite the progressive incorporation of AI in urban planning, there remains a notable gap in methodological approaches that apply AI to urban planning while considering criteria such as accessibility, pollution, and thermal comfort. It is precisely this niche that this Thesis aims to fill, by addressing the objectives set out in the following section. 1.2 Objectives AI models or techniques can be used in different areas of urban planning that, carried out with software planning tools lacking intelligent functionalities, can be unproductive, inefficient, slow or costly. Consequently, the general objective of this Thesis is to explore the different capabilities of AI in solving urban related issues. AI techniques will be explored and exploited to improve urban comfort in its different branches. Specifically, the focus will be on enhancing two crucial elements that significantly contribute to the sustainability of cities: pedestrian accessibility and thermal comfort. Different techniques will be addressed for each of the challenges. As milestones to be completed, the following specific objectives have been established: •Improving walkability through intelligent optimization techniques: To this end, a new framework will be proposed to improve both the pedestrian accessibility of cities and their perception of the environmental quality. Special emphasis will be placed on improving accessibility for the elderly and people with reduced mobility, taking into account their demands. The aim is also to make a contribution to the scarce literature on the improvement of walking routes with AI. •Exploring AI-based modeling techniques for thermal comfort: Modeling is a very early field in this area and has a lot of potential for development for urban applications. In this scenario, the emphasis is on image-to-image modeling to estimate air temperature in cities. Deep Neural Networks (DNNs) can be a time-saving way to detect heat stress in different cities. 1.2. Objectives 5 The two specific objectives of this thesis are technologically independent. However, from an urban planning perspective, these factors cannot be considered in isolation, as they involve distinct decision-making elements that may conflict with one another, ultimately influencing the overall decision-making process. 1.2.1 Outline and Contributions of the Thesis This Thesis begins by providing some preliminaries on the technology adopted during the course of research, focusing on the specific techniques and sub-tasks that will be used in the two technical contributions of the Thesis: meta-heuristic optimisation and deep learning (DL). Then, a brief review of the general state-of-the-art of the field will be made, focusing its coverage on the application of AI-based approaches to resilient cities. Hereafter, the two main technical contributions of this thesis will be presented, delving into the methodology proposed in each of them and the results validating their applicability to real-world urban problems. The last chapter will outline and conclude the Thesis with a summary of the main conclusions drawn from the Thesis. A short description of each of the chapters of the Thesis is given next: Chapter 2: Theoretical Framework and Literature Review This chapter provides the essential context for a proper understanding of the technical contributions presented in the following chapters. First, an introduction of the methods used in this work and an explanation of the basic concepts involved will be done. The discussion will revolve around AI and the three main problems covered by this technological paradigm: optimization, modeling and simulation. Additionally, a state-of-the-art section is included, showcasing different works that display the applicability of AI in the field of resilient cities. Chapter 3: Improving Urban Accessibility and Environmental Conditions using Graph Modeling and Multi-objective Optimization This chapter presents a novel framework based on georeferenced graph modeling combined with evolutionary optimization for the improvement of accessibility and path conditions for older people. An urban area is represented by a multi-weighted graph and attached to a multi-objective optimization algorithm. This multi-criteria problem is tackled with multiobjective evolutionary algorithms (MOEAs), resulting in a set of Paretooptimal set of solutions that refer to the set of possible urban interventions. This chapter manifests the usefulness of these foundations as a DSS for urban planners, who usually have to decide on an urban intervention on the basis of different factors such as investment and impact. 12 Chapter 2. Theoretical Framework and Literature Review the constraints are all linear functions. That is, they can be expressed in its scalar form. A linear objective function can be expressed as: 𝑓(x)=A·x+𝑏, (2.1) where A∈R1×𝑁and 𝑏∈R. On the other hand, non-linear optimization problems are a set of problems where either the objective function or the constraints or both contain nonlinear parts. This means that the problems have multiple variables and the relationship between these variables is not linear. For solving non-linear problems there are several techniques. For example, when the problem involves a convex objective function, techniques such as gradient descent can be applied. Gradient descent is an iterative optimization algorithm for finding the minimum of a function. It works by taking steps proportional to the negative of the local gradient (or approximate gradient), i.e. the gradient of the function at the current point. The algorithm is said that it converges when it reaches a minimum. The same technique can be applied to a concave function, using the gradient ascent method instead. However, not all optimization problems have these properties. Some problems are non-linear, non-convex, or involve complexities where no mathematical expression can adequately represent the system or model to be optimized, making it a black box problem that classical methods struggle to solve. For these cases intelligent optimization, a branch of AI that creates algorithms capable of learning by themselves to solve optimization problems, is applied. Intelligent optimization is also generally called metaheuristics [32]. Meta-heuristics, a term coined by Fred Glover in [33], are modern nature-inspired algorithms that outperform basic search heuristics (e.g. Tabu search algorithm, Greedy algorithm or A* algorithm). Unlike heuristics, meta-heuristics are not problem-specific. Meta-heuristics are in general stochastic algorithms that leverage a balance between explorative and exploitative (local) search, also known as diversification and intensification. Finding a good balance between these two search behaviors through the design and configuration of meta-heuristic solvers is imperative to explore the solution space of an optimization problem efficiently, find promising areas early during the search, and focus the search in such areas [34]. These algorithms can find high-quality solutions to challenging optimization problems in a reasonable computing time, but they do not guarantee optimal solutions. They are expected to work most of the time and are well-suited for global optimization. The selection of the best solutions ensures convergence to the optimum, while diversification increases the diversity of solutions and allows escape from local optima. Among the most renowned meta-heuristic algorithms are the family of Evolutionary Computation techniques (EC), although others such as Simulated Annealing [35] can also be found. Inside the family of EC the most prominent branch is that of Evolutionary Algorithms (EAs), having Genetic Algorithms (GAs) [36] as the most frequently encountered type of EA. GAs are inspired by the biological evolution of selection, reproduction and mutation. The general algorithmic flow for a GA starts with 2.1. Theoretical Framework 13 a population of candidate solutions (the decision variables) and evolve it over generations. Each solution has properties that can be mutated and altered. The objective function of each solution, which is also known as fitness in the context of EC, is evaluated and the fittest ones are selected for the next generation, where their decision variables (genes) are again modified via the application of search operators (namely, crossover and mutation) that imprint changes to their numerical values. The process continues until a predefined stopping criterion is met (e.g. a maximum number of generation or a lack of improvement of the best solution over generations). So far, optimization with a single objective function has been described. In the next subsection, problems driven by several objectives will be described, along with a brief presentation of the EC-based solvers capable of tackling these problems that have been considered in the Thesis. Multi-Objective Optimization Normally, in different areas of real life several objectives must be taken into account, which can be opposing each other. For instance, risk versus profitability in betting games or cost versus reliability in industrial maintenance. This is the case also in urban planning, where decision makers often face the dilemma of accomplishing urban renewal while reducing the cost of urban actions. This is why techniques to address multi-objective problems are of great interest. Building upon the example provided in the section on single-objective optimization (TSP), the definition is extended to illustrate what a multi-objective optimization problem stands for. In the single-objective TSP, the objective is to minimize the total distance 𝐿covering all cities once. However, let us now assume that the transportation company not only wants to minimize the distance but also maximize the profits. This problem is known as the Traveling Salesman Problem with Profits (TSPP) [37]. As an illustrative example, an open variant of the TSPP, referred to as the Open TSPP, is selected. In this open version, the traveler is not required to visit all cities but can choose a subset to optimize the overall cost. Specifically, two key conditions are assumed: (1) the problem is open, allowing flexibility in city selection, and (2) there is a distinct cost associated with visiting each city, meaning that the benefits of the journey depend not only on the number of cities visited but also on the specific cities chosen. This setup leads to a Pareto front that emerges from the intuition that the cost is inversely related to the distance traveled and the particular subset of cities selected for the route. Hence, in this case, the problem becomes from having a single objective function to minimize, the total distance 𝑓(x)=𝐿to having two objective functions: the total distance 𝑓1(x)and the profit 𝑓2(x). The goal is to find the set of Pareto-optimal solutions that minimize 𝑓1(x) (distance) and maximize 𝑓2(x)(profits). In multi-objective optimization, a Pareto optimal solution (also known as a non-dominated solution) is a solution for which it is impossible to improve one objective without worsening at least one other objective. The set of non-dominated solutions composes the sought Pareto front, represented in Figure 2.4 for a min-min 14 Chapter 2. Theoretical Framework and Literature Review multi-objective optimization problem. The challenge for multi-objective optimization solvers is to efficiently compute this Pareto front for a given set of objectives, or at least provide a set of solutions that approach this Pareto optimal set, both in terms of convergence and diversity. Optimal Pareto front f1(x) f2(x) Dominated solution Non-dominated solution Figure 2.4: Pareto dominance in a multi-objective optimization problem comprising two objectives 𝑓1(x)and 𝑓2(x)that must be minimized. Points x are categorized as Dominated solutions and Non-dominated solutions. The Pareto front curve represents the optimal trade-off between objectives 𝑓1(x) and 𝑓2(x). The goal is to produce a set of solutions that best approximate this optimal Pareto front. Approximating the optimal Pareto front of a multi-objective optimization problem can become a complicated task when the dimensionality of the problem makes an exhaustive evaluation of solutions computationally unaffordable, and/or the objectives do not possess a tractable mathematical formulation. In such cases, meta-heuristic optimization solvers can be extended to retain solutions based on different criteria related to Pareto optimality. This gives rise to the family of MOEAs. Once again, as in single-objective problems, GAs are the most widely used class. Within GAs a wide range of algorithms can be found that are conducive to multiobjective optimization. In what follows we focus on three of the most exploited GAs in particular and explain their most differential features: Non-dominated Sorting Genetic Algorithm II (NSGA-II), Non-dominated Sorting Genetic Algorithm III (NSGA-III) and Strength Pareto Evolutionary Algorithm 2 (SPEA2). Non-dominated Sorting Genetic Algorithm II The NSGA-II is an improved version of the NSGA algorithm, proposed by Deb et al. [38]. It is considered as the EA of reference when applying 2.1. Theoretical Framework 15 multi-objective optimization due to its widely proven robustness and efficiency. NSGA-II has a computational complexity of O(𝑀𝑁2), where M is the number of objectives and N is the population size, giving it lower computational times compare to other MOEAs. Before exploring the algorithmic flow, some key features will be presented. There are three aspects that make the NSGA-II different from other multi-objective algorithms: •Non-dominated sorting: Individuals are ranked based on Pareto dominance. An individual is said to dominate another if it is no worse in all objectives and better in at least one objective. The population is sorted into different fronts (F1,F2,F3, ...), which are shown in Figure 2.5a. The first front (F1) consists of non-dominated individuals, the second front (F2) is dominated only by individuals in the first front, and so forth. •Diversity preserving mechanism: It is based on the concept of crowding distance (CD). The CD, illustrated in Figure 2.5b, quantifies how close a solution is to its neighboring solutions, represented by the perimeter of the cuboid surrounding the individual in the objective space. To maintain diversity within the population, solutions located in less crowded regions—those with larger crowding distances—are prioritized for selection. In the event that the population limit is exceeded, and multiple solutions share the same rank (e.g., 7 individuals in rank 1 and 8 in rank 2), the algorithm retains all the rank 1 individuals (i.e. nondominated sorting) and selects additional solutions from rank 2 based on their crowding distance, favoring those with the highest values. This ensures a well-distributed set of solutions across the Pareto front while maintaining the population size. •Elitism: In NSGA-II, elitism refers to the preservation of the best solutions from a population, specifically the non-dominated solutions, into the next generation. To implement elitism, it is essential to evaluate the quality of a solution in terms of its proximity to Pareto optimality. This is achieved using a dual criterion: non-dominated sorting, which ranks solutions based on their dominance relationships, and CD, which ensures diversity by measuring the density of solutions in the objective space. Once the basic inherent concepts of NSGA-II have been explained, now the algorithmic procedure will be presented. The Figure 2.6 presents the schematic of the procedure. The first step is to create a initial population Pt, where 𝑡represents the generation number, with 𝑡=0corresponding to the first generation. After applying the usual genetic operators of crossover and mutation operations, an offspring population Qtis created. By fusing both populations together, a new one of size 2N is formed, say Rt. Then, the non-dominated sorting is applied, placing the individuals in the different fronts according to their non-dominance. The best solutions will be placed at the 𝐹1front. Once the first front is full, the solutions will be placed in the second front 𝐹2and so forth. The Rtsize is 2N while the new population Rt+1 size must be N. Hence, the most dominated fronts are rejected. When evaluating the final accepted front (𝐹3in Figure 2.6), 16 Chapter 2. Theoretical Framework and Literature Review it might contain more points than the available spaces in the new population. Rather than randomly eliminating some elements from the last front, the points that maximize the diversity of the chosen points are selected, thanks to the crowding distance. After all the process, the new population Pt+1 is created and the process is repeated again, until an optimal solution is reached. Pareto front Rank 1 Rank 2 Rank 3 f1(x) f2(x) (a) Cuboid i-1 i i+1 f1(x) f2(x) (b) Figure 2.5: The following figures illustrate two key aspects of the NSGA-II algorithm: (A) The non-dominating sorting procedure. The curve closest to the origin is the Pareto front, representing the optimal set of solutions. (B) The crowding distance calculation. i-1, i and i+1 are different solutions of the Pareto front. Non-dominated sorting F1 F2 F3 Rejected Pt+1 Pt Qt Rt Figure 2.6: NSGA-II procedural scheme for a population 𝑃𝑡and a offspring population 𝑄𝑡, which together form the new population 𝑅𝑡. Each of the 𝐹1, 𝐹2, 𝐹3are the different fronts containing solutions of 𝑅𝑡sorted by applying the non-dominated sorting. The horizontal dashed line limits the number of N solutions allowed for the next 𝑃𝑡+1population. 2.1. Theoretical Framework 17 Non-dominated Sorting Genetic Algorithm III The NSGA-III [39] can be seen as an extension of the NSGA-II algorithm, developed to improve the efficiency in optimization problems comprising more than two objectives. NSGA-III shares similarities with its predecessors (NSGA and NSGA-II) since the population is sorted in different fronts according to the non-dominance of the individuals, in the similar way it is done in NSGA-II. However, the diversity preserving mechanism is different. In NSGA-III the crowding distance introduced in NSGA-II is replaced by a set of reference points or directions in the objective space that guide the search towards different regions of the Pareto front. The set of reference points is determined in the beginning of the search and is widely distributed in the objective space and near the Pareto front to ensure diversity. For each set of solutions, the perpendicular distances between each solution and the different reference directions are calculated, keeping the shortest distance to one of the reference directions. For all solutions that fall within this region, i.e. for all solutions sharing the same reference direction, niche preservation is applied. This means that only the solution with the smallest distance is saved for the next population. This procedure is applied in all the different regions that are divided by the different reference directions. This ensures diverse and uniform solutions are kept within the target space in the next population. Figure 2.7 shows the concepts of reference point and directions and the niche preservation. f1 f2 1 2 3 4 5 6 Reference line Ref. point Normalized hyperplane Figure 2.7: NSGA-III algorithm working procedure. Points numbered from 1 to 6 are plotted, showing potential solutions. Reference lines and the Normalized hyperplane intersect at a Reference points, indicating reference directions for the optimization process. In 2D for clarity. NSGA-III has yielded consistent results on multi-objective problems of between three and 15 objectives [39]. The problems considered in the extensive benchmark performed in this latter work are heterogeneous and 18 Chapter 2. Theoretical Framework and Literature Review comprise different objective functions ranging from convex, concave, disjointed and multi-modal ones, involving multiple local fronts. In all of them, NSGA-III managed to outperform other multi-objective solvers with superior Pareto front approximations in terms of convergence and diversity. Strength Pareto Evolutionary Algorithm 2 The SPEA2 [40] is an improved version of the SPEA algorithm. The upgrades hinge on three main pillars: i) improved fitness assignment strategy, ii) a density estimation technique and iii) a new archive truncation method. In the following a description of the three most relevant features of SPEA2 is presented: •External archive: In both SPEA and SPEA2, there is a external archive where non-dominated solutions encountered during the search process are stored. The archive is updated every generation to maintain the best solutions so far. However, while for SPEA the size varies in each generation, SPEA2 operates with a constant size for both the population and the external archive. •Fitness assignment: The fitness of an individual x𝑝in the population is composed by the sum of its raw fitness 𝑅(𝑝)and density 𝐷(𝑝). Starting from a population 𝑃𝑡that serves as the existing pool of potential solutions and an external archive ¯ 𝑃𝑡, let assign to each individual x𝑝 inside that population and archive with a strength value 𝑆(𝑝), which represents the number of solutions it dominates: 𝑆(𝑝)=|{𝑝′|𝑝∈𝑃𝑡+¯ 𝑃𝑡∧𝑝≻𝑝′}|,(2.2) where | · | denotes the cardinality of a set, + stands for multiset union and the symbol ≻corresponds to the Pareto dominance relation. The raw fitness 𝑅(𝑝)of an individual x𝑝is calculated based on the 𝑆(𝑝): 𝑅(𝑝)=∑︁ 𝑝′∈𝑃𝑡+¯ 𝑃𝑡, 𝑝≻𝑝′ 𝑆(𝑝′).(2.3) In other words, the raw fitness is calculated based on the strengths of those that dominate it in both the archive and the population, unlike in SPEA where only members of the archive are taken into account in this regard. The goal here is to minimize fitness, meaning that 𝑅(𝑝)=0 signifies a non-dominated individual, while a high 𝑅(𝑝)value indicates that the individual 𝑝is dominated by a large number of individuals, as depicted in Figure 2.8. Finally, the density is computed using the k-th nearest neighbor [41] method: 𝐷(𝑝)=1 𝜎𝑘 𝑝+2,(2.4) 2.1. Theoretical Framework 19 f1 f2 0 0 0 Non-dominated Dominated 2 5 9 19 14 Figure 2.8: The raw SPEA 2 fitness values for a maximization problem. where 𝜎𝑘 𝑝is the euclidean distance between the individual 𝑝and its k-th nearest neighbor in objective space. The final fitness value for the x𝑝 individual is: 𝑄(𝑝)=𝑅(𝑝) + 𝐷(𝑝).(2.5) •Environmental selection: ensures both the diversity and convergence of the population. The environmental selection process copies all the non-dominated solutions (𝑄(𝑝)<1) to the next generation archive. If the size of the non-dominated solutions fits the pre-defined size of the archive, denoted as ¯ 𝑁, the environmental selection is fulfilled. If there are less non-dominated solution than ¯ 𝑁, the best dominated solutions are added until filling all the archive. However, if the non-dominated solutions exceed the size ¯ 𝑁, an archive truncation is done in an iterative removal process where the individual which has the minimum euclidean distance to another individual is removed. The algorithmic flow is presented as in [40]. First an initial population 𝑃0is generated under an initialization criterion, while the archive ¯ 𝑃𝑡=∅ is initially empty. Then the fitness computation procedure is performed for all the individuals in 𝑃𝑡and ¯ 𝑃𝑡(for the first archive no fitness computation is done, since it is empty). Next, the environmental selection is applied. At that point, if the maximum number of generations has been reached or another stopping criteria is satisfied the ¯ 𝑃𝑡+1is set as the output non-dominated set. Then a binary tournament selection with replacement on ¯ 𝑃𝑡+1is performed on P in order to fill the mating pool. Next, recombination and mutation operators are applied to the mating pool and the resulting population is assigned to ¯ 𝑃𝑡+1. Finally, the generation counter is incremented (𝑡=𝑡+1) and the fitness assignment is recalculated again. Like NSGA-II, SPEA2 is a widely used algorithm in the field of optimization. 20 Chapter 2. Theoretical Framework and Literature Review 2.1.2 Modeling Problems Within modeling problems, several types can be encountered depending on the data format to be processed. Given the large modeling capability of modern ML techniques (e.g. DNN), most modeling problems tackled via AI are formulated on complex data modalities, including text, audio and image data. Among them, image modeling is arguably one of the most widely considered data modalities tackled via AI-based techniques. In image modeling problems, several modeling tasks can be distinguished, including: •Image Classification: Image classification modeling consist on classifying images with a unique label. A typical example of image classification modeling is the dog-cat classifier, where the model is trained with several dog and cat images and learns to classify them into their respective categories. •Image Segmentation: This type of modeling can be considered as a step forward in image classification modeling. Image segmentation models work at pixel-level and learn to label pixels according to their class, dividing the image in different segments. There are two types of segmentation: semantic segmentation and instance segmentation. In semantic segmentation the pixels are labeled and grouped based on the categories they belong to (e.g. trees, persons, cars...). In instance segmentation, this partitioning is more exact and the different categories are also segmented depending on the individual objects (e.g. tree one, tree two, human one, human two...). •Image Regression: Unlike the two previous models, image regression works with continuous values instead of labels. The model is trained to translate the pixels of an image, or the whole image, to continuous numerical values. The model learns to map the features extracted from the image to the target regression values. Image regression models have a wide range of applications, including age estimation [42], [43], facial landmark detection [44], [45], and bounded box regression in object detection [46]. In recent years, modeling problems formulated on image data as those exemplified above are mostly solved by using DNNs. In order to account for the spatial relationships within image data that are inherent to such problems, DNNs used in image classification are built on top of convolutional neural networks (CNNs), which were initially proposed by Lecun [47] inspired by early work on the neocognitron by Fukushima [48]. In what follows we briefly revisit the fundamentals of this particular flavor of DNNs. Convolutional Neural Networks CNNs are a type of feed-forward neural networks primarily used for image processing and object recognition [48]. The reason that CNNs are so 2.1. Theoretical Framework 21 extensively used in this area is that they are designed to autonomously and dynamically learn spatial positions of features from the input data. In CNNs the spatial information is preserved, which provides the context and relationship between different parts of the image. This can be achieved thanks to the convolution operation that this neural network apply to the input data, which downsamples the input data while maintaining the spatial relationships between the pixels of the image. The superficial architecture of CNNs does not differ from those of other types of ANNs and consists of three large building blocks, as can be seen in Figure 2.9 . First, an input layer holds the input pixel matrix of the image. Then, several hidden layers can be found. The hidden layers are the core part of the CNNs and the building block that differentiates them from other types of ANNs. Finally, an output layer which is the final layer that produces the predictions or classifications based on the learned features from the previous layers. The main building block -hidden layersdepicted in Figure 2.9 , can be broken down into three main types of layers: convolutional layers, pooling layers and fully connected layers. •Convolutional layers: The convolutional layers apply the convolution operator to the input image. A convolution is a linear operation that involves the scalar product between the pixels of the image and the different weights of the kernels. A kernel (or filter) is a matrix of weights. These filters are “convolved” over the input matrix to compute a feature map. In other words, the filter is slid over the entire input image, left to right, top to bottom. The convolution operation exhibits several properties that are particularly advantageous for computer vision tasks. First, the patterns learned by convolutional layers are translation invariant, meaning that a pattern recognized at one location can be detected regardless of its spatial position within the image. Second, convolutional layers capture the spatial hierarchies of patterns. This hierarchical learning allows early layers to identify fundamental, small-scale structures such as edges, while deeper layers detect increasingly complex patterns, which are essentially combinations of features extracted by preceding layers. As a result, a CNN is able to progressively comprehend more intricate and abstract visual patterns through its layered architecture. •Pooling layers: Pooling layers are applied immediately after a convolutional layer to process the resulting feature maps. The primary function of a pooling layer is to perform downsampling, thereby reducing the dimensionality and the number of parameters in the model. A commonly employed pooling technique is Max Pooling, which reduces the size of the feature maps by selecting the maximum value within each defined window. This operation not only decreases computational complexity but also helps to retain the most salient features of the input. •Fully Connected layers (FC): FC layers, which typically follow the convolutional and pooling layers, establish connections between every neuron in the current layer and every neuron in the preceding layer. Through this dense connectivity, the FC layers learn patterns and relationships 28 Chapter 2. Theoretical Framework and Literature Review and thus calculate the Green Index View (GIV). Ki et al. [66] go a step further and also applies image segmentation to Google Street View to obtain the GIV and then, they correlate this GIV with the walking time spent on those streets. The work also shows that other factors such as income and car use in a given neighborhood also affect both GIV and walking time. Again, as in the work of Nagata et al. [64], the impact of this work offers development of public policies that encourage pedestrian-friendly environments. In the field of accessibility, Google Street View images have also been used as a database. For instance, Hara et al. [67] propose a new system called Tohme that leverages crowd-sourcing, computer vision, and machine learning to detect curb ramps in Google Street View scenes. Tohme can effectively analyze and identify curb ramps in Google Street View scenes by using image segmentation, contributing to the system’s overall success in detecting accessibility issues. The system’s innovative approach significantly reduces the human time cost associated with manual audits while maintaining high quality in results, making it a valuable tool for urban planning and improving accessibility in cities worldwide. So far, the review of the state of the art has focused on segmentation of Google Street View images, a field with a wide scope. However, other types of examples can also be found within modeling. For example, Wang et al. [68] introduce a geometric consistency enhanced deep convolutional encoder-decoder framework for urban seismic damage assessment. The encoder-decoder architecture used is a U-Net with slight modifications. They apply image segmentation to recognize under diverse weather conditions (rain, fog, darkness) earthquake-damaged buildings. They train the model using unmanned aerial vehicle images from the county of Beichuan achieving reliable, accurate and robust results, increasing the resilience of the urban area. There are also some works where the two techniques merge and the optimization process is made over the parameters of the DL models chosen with the aim of developing more precise modeling techniques. This is the case of Pan et al. [69], who propose a differential evolution (DE) back propagation neural network traffic prediction model for smart cities. The model accurately predicts network traffic trends by mapping the impact factor of network traffic to the actual traffic volume. The model utilizes DE for global search to optimize the connection weights between layers, enhancing the accuracy of predictions. Through training with past traffic data, the model achieves accurate prediction of network traffic trends in smart cities. The predicted traffic volume aligns with the actual traffic volume trend within an small error. It has also proven to be suitable for long-term predictions in complex and heterogeneous smart city network environments. Thanks to quality predictions, urban planners are assisted in managing the traffic network as well as in improving infrastructure. Similarly, Alghamdi in his work [70] employs a combination of EAs and DL models, Recurrent Neural Networks (RNNs) specifically, to predict the urban evolution of future smart cities. RNNs are trained with datasets coming from CityPulse framework and as in the previous case, the EA is applied for the optimization of the weights of the neural network layers. Once the model is trained, they propose two use cases. The 2.2. Literature Review 29 first use case is to creating a new city from scratch, with some constrains in energy, number of inhabitants and vehicles applied. The framework tends to built cities with the maximization of coverage, connectivity, and localization. The second use case predicts the potential developments in the urban growth of a current city. This scheme could help to guide the expansion of a city. The examples provided represent just a few strokes on the vast canvas of modeling for urban problems. There exists a multitude of other examples that further illustrate the depth and breadth of this field. For that, a more comprehensive exploration of the state-of-the-art is presented in the technical contribution of Chapter 4. This preliminary related work reveals that the field of AI applied to urban planning and its challenges is an incipient field of significant practical interest. However, as this initial review evinces, most works concentrate on a purely theoretical approach to the problems. Municipalities and governments, on the other hand, require studies for immediate and practical application. As will be demonstrated later in the chapters of the technical contributions, there is no state-of-the-art for the field of those contributions proposed in this Thesis. These contributions focus not only on the theoretical framework, but also on the applicability of the framework by the institutions. 31 Chapter 3 Improving Urban Accessibility and Environmental Conditions using Graph Modeling and Multi-objective Optimization As commented in the introductory chapter of this Thesis, urban areas are facing multiple challenges that may hinder their social progress in the future. Among the problems that cities suffer from nowadays, accessibility is in the top tier. Accessibility can be defined as the opportunity to access/reach community resources, services and recreational facilities [71]. Some areas in cities however suffer from lack of accessibility. Moreover, other concerning issue is the age of the population: the number of people over 65 years old will double in the next three decades, reaching 16% of the world’s population and exceeding 1500 million people in 2050 [72]. This last population substrate suffers the most from the consequences of the poor accessibility of urban areas due to their physical limitations. Ultimately, their reduced mobility causes serious social problems, including isolation, dependence, and marginalization. Other challenges that modern cities face are the ones related with the environmental aspects of their streets. Today, all major cities suffer from air pollution and high levels of noise in public urban spaces. Apart from the obvious damaging consequences to human health (giving rise to millions of deaths yearly [73]), pollution also brings a significantly negative impact on the economy of countries worldwide. For example, in 2018 alone the costs of air pollution to the National Health Service (NHS) and social care in England were estimated to be £157 million, which could reach an estimate of £18.6 billion by 2035 if no actions are taken [74]. Further along this line, noise pollution represents for Europeans the fifth most significant problem (72%) within their cities, just second after environmental pollution (81%) [75]. Noise 32 Chapter 3. Improving Urban Accessibility and Environmental Conditions using Graph Modeling and Multi-objective Optimization pollution, apart from its inherent annoyance, can cause transitory acute effects (temporary hearing loss) and long-term chronic diseases such as depression, anxiety, and loss of sleep. In the medium-long term, such diseases can entail serious risks for health, such as total hearing loss or cardiovascular diseases [76], [77]. Furthermore, to these health-related risks it must be added that neighbors of noisy urban residential areas may suffer also from socio-behavioral modifications [77]. These effects, in turn, are more likely to appear in vulnerable groups, such as people with health problems and older people. Bearing in mind this scenario, urban planners confront the challenge of making cities friendlier, especially for the older populations, who are most affected by the two problems exposed above. Hence, urban areas should be adapted to account for the needs of older people, making them safer, healthier, and more inhabitable places for this group, i.e., they should become age-friendly [5], [78]. With this objective in mind, in line with the growing integration of computational tools in urban planning, urban planners and designers have begun exploring the potential of AI to enhance decision-making processes and develop urban plans focused on improving accessibility. This reflects a broader shift toward the use of advanced technologies, such as DSS and PSS, which, as mentioned earlier in Chapter 1, contribute to more informed and holistic planning outcomes. However, given the generally restricted municipal public budgets, a widely acknowledged decision driver in urban design decision-making is the cost effectiveness of any proposed decision. For instance, when the process involves the modification or addition of urban elements (e.g., mechanical ramps, escalators, lifts), decisions are often made based on estimations of the improvement in terms of accessibility derived from the intervention. Several practical problems arise along the decision-making process: A) information sources needed for making an informed decision are diverse, heterogeneous and not consolidated into a single point of information; B) there are no technological means to quantitatively gauge the estimated accessibility and/or quality improvements of the urban environment corresponding to a certain decision, nor any guarantee that the decision itself is the best one in terms of cost effectiveness; C) the consequences of the interplay between objectives related to accessibility and urban quality through the cost of the intervention, by which the urban planner is uncertain whether the addition of an element (e.g., a mechanical ramp) is the best option that can be made without impacting negatively on the other objectives. As a result, the process is often approached by resorting to common sense, statistical studies, audits, or the expert knowledge of experienced urban planners [79]–[81]. This chapter addresses this scenario departing from the observation that the simultaneous consideration of different design objectives in an urban planning problem can be regarded as a multi-criteria optimization problem. As such, the goal of this kind of problems is to discover the set of feasible solutions that best approximates the Pareto trade-off between the considered objectives. Within the context of this investigation, a Paretooptimal set of solutions refers to the set of possible urban interventions 3.1. Related Work 33 that best balance the economic cost of the intervention and numerical measurements of its impact on the accessibility, environmental noise, and air pollution of the urban area under study. The technical contribution of this chapter builds upon hypothesizing that the use of a multi-weighted graph model of the urban area, together with multi-objective evolutionary optimization, can be used to solve the issues A, B and C described previously. All these ingredients embody a novel framework that showcases the potential of evolutionary computation for decision-making processes over urban elements, aimed at improving accessibility and path conditions for older people. The framework relies on a geo-referenced graph model to collect and centralize all information of interest over a given urban area. EAs for multi-objective optimization are utilized to compute an estimation of the Pareto front drawn by the cost of installation of different urban elements (i.e., mechanical ramps, escalators, lifts, acoustic panels), and their impact in terms of travel time, air pollution reduction and environmental noise mitigation. Such objectives are computed over key (origin,destination) pairs that represent points of interest (hospital, day-care centers, or drug stores), also considering realistic models for these four objectives. Experimental results in two use cases located in the city of Barcelona (Spain) are discussed to validate the applicability of the proposed framework in real-world settings. 3.1 Related Work Before proceeding with the description of the framework and the underlying optimization problem, this section overviews baseline literature and technical advances reported at the crossroads of urban accessibility and environmental quality (Subsection 3.1.1) and multi-criteria decision-making (Subsection 3.1.2). The section ends by highlighting the contribution done to the state-of-the-art (Subsection 3.1.3). 3.1.1 Accessibility and Environmental Quality for Agefriendly Cities Cities are complex ecosystems where multiple dimensions coexist and interact [82], ranging from physical elements such as buildings, infrastructures, and public spaces to socioeconomic relationships, cultural expressions, environmental (air, noise) and natural elements (blue and green areas, or fauna, to mention a few). Urban quality relies on multiple aspects: one of them is the accessibility to places and services necessary for daily life in a reasonable time, which has been lately framed under the “15-minute city” concept [83]. This is especially true for vulnerable populations, such as older people, which have greater mobility difficulties. Therefore, accessibility is a key aspect of age-friendly cities [5], [84]. Furthermore, it is important not only to be able to get to places, but also to guarantee the quality of the itineraries. One of the goals of agefriendly cities is to foster older people to stay active for as long in life as possible. Therefore, it is key to offer itineraries that are attractive and 34 Chapter 3. Improving Urban Accessibility and Environmental Conditions using Graph Modeling and Multi-objective Optimization safe, to induce people to walk around every day. The capacity of a place to improve people’s well-being relates to the concept of a restorative environment, i.e., environments enhancing or facilitating psychological restoration, and thus contributing to human health and well-being [85]. Among many other factors, acoustic comfort has been proven to greatly contribute to environmental quality [86]. Therefore, it is a dimension that should be considered when designing age-friendly itineraries. Air quality is another key aspect, greatly contributing to urban health indicators [87], with older people being especially sensitive to the effects of breathing polluted air. 3.1.2 Multi-criteria Design and Multi-objective Optimization Multi-criteria analysis (MCA) has been a widely used tool as a support for decision-making when carrying out an intervention in urban contexts [21], [88]–[90]. MCA is used to make a comparative discrimination between heterogeneous projects, measurements to take and/or decisions to be made. To this end, solutions that can be considered are compared and evaluated systematically according to the case. The MCA approach must be nourished by a large amount of data, as well as by the preferences of the decision-makers, which eventually determine the specific weight given to one or another criterion [91]. In some cases, MCA can be combined with simulation engines for a higher degree of robustness in the results [92]. However, this approach is usually very time-consuming, since different possibilities need to be evaluated for diverse scenarios [91], [93], and the approach does not ensure the optimality but rather the feasibility of the plan [94]. According to the review in Chapter 2, the multi-criteria design of urban regeneration plans can be solved with MOEAs which allow efficiently exploring the set of possible solutions (urban interventions), and eventually approximating the Pareto front. As a result, decision-makers are informed with a higher number of possible actions to choose from [94]. MOEAs are used in different fields of research, including chemistry [95], bio-informatics [96], business analytics [97] or environmental sustainability [98], to mention a few. In urban planning several factors generally come into play, which ultimately leads to an optimization problem comprising different objectives. Several examples of the use of MOEAs to make an optimal use of land can be found in the literature. For example, the work in [99] resorts to genetic algorithms to estimate the future use of land in a city, considering three objectives: 1) the minimization of the cost of the intervention; 2) the minimization of traffic congestion; and 3) the minimization of the number of changes to be made over already built-up areas of the city. Masoumi et al. seek in [100] a different use for the land, namely, to create an industrial complex over the outskirts of a city. For that purpose, they formulate different criteria than those in [99], such as the distance to the nearest city, the road accessibility, the availability of resources and the distance to environmentally protected areas and forests. They also 3.1. Related Work 35 utilize genetic algorithms to efficiently deal with the optimization problem stemming from their study. In a different vein, Matias Péres et al. [101] propose to use MOEAs to reduce traffic congestion in Montevideo (Uruguay), optimizing the cycles of traffic lights in the city, leading to a clear reduction of the air pollution in the city. In the next subsection, we focus our review on multi-objective optimization applied to the improvement of urban accessibility for older people, addressing the three issues discussed in the introduction: noise pollution, air pollution and pedestrian accessibility. 3.1.3 Multi-objective Optimization for Urban Accessibility Even though noise, air pollution and accessibility are major issues that modern cities encounter nowadays, very few examples can be found where optimization, – neither multi-criteria nor single-objective – has been employed to make urban decisions to deal with these problems. When considering the specifics of age-friendliness (e.g., reduced mobility), the literature is even more scarce. To tackle the problem of noise pollution, one of the most used techniques is to analyze the propagation of noise through the different streets, through simulations that consider the geometry of the streets and different elements that impact the perception of noise, such as facades, the distance to roads or the presence of vegetation [102], [103]. When it comes to the adoption of multi-objective optimization to deal with noise pollution, Hammad et al. [104] address a multi-objective facility problem, where the aim is to simultaneously minimize noise pollution and improve travel times. To do this, they discriminate between two types of buildings: those that generate noise (industries, shopping malls), and those that are sensitive to noise pollution (hospitals, schools). They use graph theory to apply their optimization, where the nodes are the buildings, and the edges are the roads that connect them. They seek to find the optimal configuration that minimizes noise pollution and travel times between nodes. Likewise, Ning et al. [105] apply a hybrid genetic algorithm –ACO algorithm– to reduce noise pollution in construction sites. The study considers three different objectives to optimize: noise pollution of different construction tasks, the overall transportation cost, and potential risk related to interaction flow between the facilities. Regarding air pollution in urban areas, the trend is to exploit the potential of data-based modeling to forecast pollution, so predictive estimations can inform subsequent decision-making processes [106]–[108]. However, studies have been also reported using optimization to deal with air pollution. An example is the work by Wang et al. in [109], which proposes a multi-objective optimization algorithm to build an air pollution early warning system which forecasts pollution levels and calculates air quality based on that forecast. Other authors have focused on providing different solutions to challenge the growth of air deterioration in urban areas. The aforementioned work published in [101] aims to improve traffic flows 36 Chapter 3. Improving Urban Accessibility and Environmental Conditions using Graph Modeling and Multi-objective Optimization over cities, which is one of the major causes for the increase in pollution. Differently, the work in [110] focuses rather on urban spaces, using multi-objective evolutionary optimization for minimizing urban ecosystem services loss and maximizing the compactness in planning new city developments, using Singapore as a case study. Urban accessibility enjoys a greater bibliographical collection, with several examples of optimization problems formulated to deal with this important driver of age-friendly cities. Unlike in research on noise and air pollution, in urban accessibility issues the most vulnerable groups such as older people or people with reduced mobility must be considered. A widely used technique not only in accessibility problems but also in other urban-related questions is the use of georeferenced graphs. Graphs allow representing the urban area under study and embedding map information and data in a georeferenced manner (pollution, noise, slope of the streets, etc.). For instance, D´Orso et al. [111] propose a tool that helps to decide which urban interventions are of greater importance for pedestrian walkability. To do this, they represent the pedestrian network with arches. A quality index is attributed to each of the arches associated with its walkability. To decide which arcs should be prioritized for their improvement, both the aforementioned index and the usage frequency of the arches are considered. Based on these factors, shortest routes between origin nodes (train stations in this study) and attractors or destination nodes (places with high pedestrian influx) are computed. A well-established methodology to use multi-objective optimization for urban accessibility is to combine MOEAs with geo-referenced graphs. This combination is present, for instance, in the work of Blecic et al. [112], which conceives a support tool for urban planners based on graph model combined with the NSGA-II algorithm. The tool computes a Pareto front approximation between the walkability improvement and the implementation effort. They compute the walkability cost for all the edges of the graph, which represent street segments, based on several parameters that impact on walkability. They aim to reduce costs derived from the application of improvements to such street segments. Hence, the decision maker is offered different alternatives. In the same way, Salcedo-Sanz et al. [113] combine directed graphs that represent the road network of a Spanish city with a MOEA to improve traffic flow when there are punctual traffic cuts in certain zones of the city. They obtain a Pareto front where the different solutions are the possible routes from the input to the output nodes. When it refers to the accessibility of people with reduced mobility (such as wheelchair users or older people), several studies have adopted this graph-based framework. For instance, Bolten et al. [114] utilize graph theory to represent and characterize the streets of an urban area in terms of wheelchair accessibility. Following this trend, Barczyszyn et al.[115] have created a collaborative route planning service for wheelchair users, based on sidewalk-based model. For that, they also use a graph-model to describe an area where the sidewalks and crosswalks are represented by the edges and the junctions on the sidewalks by the nodes of the graph. Those 3.1. Related Work 37 edges are weighted according to some features such as the slope, the existence of curb ramps on crosswalks and the maintenance of the sidewalks. Based on that graph, the rout planner algorithm offers alternative accessible routes for wheelchair users. Similarly, Rahaman et al. [116] devise a Contour-based Accessible Path Routing Algorithm (CAPRA) that hinges on the well-known A∗heuristic to compute several alternative routes that enhance the accessibility of older people. Another work in this direction is the one published by Sasaki et al. in [117]. Specifically, they address the discovery of routes for a walk considering measures of the safety, amenity and walkability of a user traversing them. Algorithmically the study combines the A∗algorithm and a GA to yield subroutes that maximize such measures. Alternative types of urban interventions are also of great importance to create accessible cities, mainly for older people. For example, Zhang et al. [118] propose MOEAs to decide the locations of healthcare centers that maximize accessibility for everyone, covering as many people as possible within an acceptable distance radius from them, and also minimizing the costs of the process. More recently, the work in [119] explores the use of MOEAs to find cost-effective urban interventions that balance accessibility and cost. However, no other objectives are considered in this study, nor does it analyze qualitatively whether the optimized interventions correspond to decisions that actually make sense from the practical point of view. Contribution. In light of the above literature review, to the best of our knowledge no previous work has so far proposed the improvement of urban accessibility and quality considering three major problems that metropolises suffer today. In general, noise, pollution and accessibility problems are treated independently. However, for older people, these three factors are of vital importance when choosing one route or another. Therefore, they should not be tackled exclusively, but jointly. Moreover, the overall cost of the urban intervention also induces an interplay between such objectives, making it unclear whether the decision maker should invest in the installation of an urban element or another. For this reason, we extend the framework in [119] to account for the three factors mentioned previously, towards finding urban interventions that best balance (in the Pareto sense) between the noise and air pollution perceived by a user going through them, his/her accessibility given the topographical characteristics of the paths, and their economic cost. This framework, which is described in the next section, combines georeferenced graphs with MOEAs to determine where to install urban elements aimed to improve noise, air pollution and accessibility (e.g., escalators, elevators, green or acoustic panels), in order to improve urban accessibility and environmental quality and ultimately, give rise to a more age-friendly city. 44 Chapter 3. Improving Urban Accessibility and Environmental Conditions using Graph Modeling and Multi-objective Optimization yielding: 𝑓𝑛𝑜𝑖𝑠𝑒 (𝛀|V𝑜,V𝑑) =1 |V𝑜||V𝑑|∑︁ 𝑣𝑜∈V𝑜∑︁ 𝑣𝑑∈V𝑑∑︁ (𝑣,𝑣′)∈E𝑜,𝑑 𝑓𝑛𝑜𝑖𝑠𝑒 (𝑣, 𝑣′, 𝑡(𝑣, 𝑣′), 𝜓(𝑣, 𝑣′)),(3.19) 𝑓𝑎𝑐𝑐 (𝛀|V𝑜,V𝑑) =1 |V𝑜||V𝑑|∑︁ 𝑣𝑜∈V𝑜∑︁ 𝑣𝑑∈V𝑑∑︁ (𝑣,𝑣′)∈E𝑜,𝑑 𝑓𝑎𝑐𝑐 (𝑣, 𝑣′, 𝑡(𝑣, 𝑣′), 𝜓(𝑣, 𝑣′)).(3.20) Finally, the cost objective follows straightforward from the sum of the costs associated with each asset of the intervention: 𝑓𝑐𝑜𝑠𝑡 (𝛀)= 𝐾 ∑︁ 𝑘=1 𝑓𝑐𝑜𝑠𝑡 (𝑣𝑘, 𝑣′ 𝑘, 𝑡(𝑣𝑘, 𝑣′ 𝑘), 𝜓(𝑣𝑘, 𝑣′ 𝑘)),(3.21) where, clearly, 𝑓𝑐𝑜𝑠𝑡 (𝑣𝑘, 𝑣′ 𝑘, 𝑡(𝑣𝑘, 𝑣′ 𝑘), 𝜓(𝑣𝑘, 𝑣′ 𝑘)) =0if 𝑡(𝑣𝑘, 𝑣′ 𝑘)=∅. Based on these definitions, the multiobjective optimization problem to be solved seeks a set of solutions (interventions) that best balance these 4 objectives. Since several objectives are considered, this set of solutions must be evaluated in terms of their Pareto optimality: solutions to be produced by the framework must delineate the trade-off between the improvement in terms of accessibility and environmental conditions enabled by the interventions (given by the three first objectives) and the economic cost to be invested to implement the intervention over the urban area under study (fourth objective). Such objectives are conflicting with each other: the more assets are to be deployed to enhance the accessibility and enjoyability of older pedestrians in their routes, the higher the economic cost of the intervention will be. Solutions to be discovered by the framework must ensure that an improvement in any of the objectives would require a degradation of the others. Therefore, the optimization problem can be formulated as the discovery of diverse interventions that approximate the Pareto-optimal balance between the four considered objectives. The problem is augmented by inserting technical limitations associated with each asset, so that the length of ramps, escalators, sound and pollution panels are kept below certain limits (𝑙𝑅 𝑚𝑎𝑥,𝑙𝐸 𝑚𝑎𝑥,𝑙𝑆𝑃 𝑚𝑎𝑥, and 𝑙𝑃𝑃 𝑚𝑎𝑥, respectively). Mathematically: min 𝐾𝑝,{𝛀𝑝}𝐾𝑝 𝑘=1 𝑓𝑎𝑐𝑐 (𝛀𝑝|V𝑜,V𝑑), 𝑓𝑎𝑖𝑟 (𝛀𝑝|V𝑜,V𝑑), 𝑓𝑛𝑜𝑖𝑠𝑒 (𝛀𝑝|V𝑜,V𝑑), 𝑓𝑐𝑜𝑠𝑡 (𝛀𝑝),(3.22) subject to: V𝑜(set of origin nodes), V𝑑(set of destination nodes), (3.23) 𝜆𝑎𝑐𝑐 +𝜆𝑎𝑖𝑟 +𝜆𝑛𝑜𝑖𝑠𝑒 =1,(3.24) Ψ(𝑣𝑝 𝑘, 𝑣′𝑝 𝑘) · 𝑙(𝑣𝑝 𝑘, 𝑣′𝑝 𝑘)< 𝑙𝑡(𝑣𝑝 𝑘,𝑣′𝑝 𝑘) 𝑚𝑎𝑥 ,∀𝑝∈ {1, . . . , 𝑃} and ∀𝑘∈ {1, . . . , 𝐾𝑝}:𝑡(𝑣𝑝 𝑘, 𝑣′𝑝 𝑘) ∈ {𝑅, 𝐸, 𝑆𝑃, 𝑃𝑃},(3.25) 3.3. Search Methodology 45 where index 𝑝in 𝛀𝑝,𝑣𝑝 𝑘and 𝑣′𝑝 𝑘refers to the 𝑝-th intervention of the Pareto-approximating set of interventions sought in the above problem formulation. Decision variables to be discovered include, therefore, the number of assets 𝐾𝑝involved in every intervention 𝛀𝑝of the Paretoapproximating set, the edges (𝑣𝑝 𝑘, 𝑣′𝑝 𝑘)in which the 𝑘-th asset in 𝛀𝑝is installed, its type 𝑇(𝑣𝑝 𝑘, 𝑣′𝑝 𝑘)and characteristics (relative length) Ψ(𝑣𝑝 𝑘, 𝑣′𝑝 𝑘). The mixture of realand integer-valued decision variables involved in the above formulation motivates the use of multi-objective evolutionary algorithms to efficiently explore the space of possible solutions. Details on how this search for Pareto-optimal interventions is accomplished by the framework are given in the next section. 3.3 Search Methodology To efficiently explore the search space spanned by the decision variables of the problem formulated previously, the framework resorts to MOEAs. As discussed in Chapter 2 MOEAs comprise a branch of meta-heuristic optimization research that focuses on the exploitation of concepts from the theory of evolution, including breeding, mutation, and survival of the fittest [121], [122]. The emulation of these natural processes as search operators within an iterative computer program gives rise to approximate gradient-free solvers that do not require any specific mathematical knowledge of the objectives to be minimized. In doing so, the search algorithm featuring evolutionary operators iteratively tailors and checks for the fitness of candidate solutions (individuals) to the problem at hand, retaining the most promising ones in a population. By repeatedly applying operators to the population of individuals, more refined solutions are produced over iterations (generations). Once a stop criterion is met (e.g., a fixed number of generations), the best solution in the population is returned as the solution to the problem. As in any other multi-objective problem tackled with MOEAs, critical design choices of the algorithm are the solution encoding (namely, how solutions to the problem are numerically represented within the evolutionary search) and the search operators (i.e., how individuals are refined and retained in the population during the evolutionary search). These elements of the MOEAs considered in the proposed framework are detailed below. 3.3.1 Solution Encoding The solution encoding (genotype) is designed in close resemblance with the phenotype imposed by the problem. It is recalled that every individual to be refined during the evolutionary search must represent an intervention 𝛀𝑝, given by: 𝛀𝑝=[𝑣𝑝 𝑘, 𝑣′𝑝 𝑘, 𝑡(𝑣𝑝 𝑘, 𝑣′𝑝 𝑘), 𝜓(𝑣𝑝 𝑘, 𝑣′𝑝 𝑘)]𝐾𝑝 𝑘=1,(3.26) namely, a structure divided in 𝐾𝑝assets, each indicating the edge (𝑣𝑝 𝑘, 𝑣′𝑝 𝑘) on which the 𝑘-th asset is installed, its type and characteristics. To ease 46 Chapter 3. Improving Urban Accessibility and Environmental Conditions using Graph Modeling and Multi-objective Optimization the application of search operators and effectively avoid variable-length genotypes, a maximum number of assets per intervention 𝐾𝑚𝑎𝑥 is assumed and an additional type of asset (i.e., 𝑡(𝑣𝑝 𝑘, 𝑣′𝑝 𝑘) ∈ {𝑅, 𝑆, 𝐸, 𝑆𝑃, 𝑃𝑃, ∅}) is inserted to account for the case where no asset is installed. In this way, the effective number of assets of a given intervention can be computed as: 𝐾𝑝= 𝐾𝑚𝑎𝑥 ∑︁ 𝑘=1 I(𝑡(𝑣𝑝 𝑘, 𝑣′𝑝 𝑘)≠∅),(3.27) where I(·) is an auxiliary indicator function taking value 1 if its argument is true (and 0 otherwise). It is important to note that this solution encoding strategy permits to define several different assets over a given edge (𝑣𝑝 𝑘, 𝑣′𝑝 𝑘), which should be allowed only when the type of assets to be deployed on the edge are compatible with each other. For instance, sound panels (𝑇(𝑣𝑝 𝑘, 𝑣′𝑝 𝑘)=𝑆𝑃) and ramps (𝑇(𝑣𝑝 𝑘, 𝑣′𝑝 𝑘)=𝑅) can be simultaneously installed on the same edge. However, a ramp and mechanical stairs (𝑇(𝑣𝑝 𝑘, 𝑣′𝑝 𝑘)=𝑆) on the same given edge make no sense from the practical point of view. For this reason, each newly produced individual is checked for these incompatibilities, retaining (when needed) the asset implying a lower installation cost. 3.3.2 Search Operators Following the algorithmic principles of MOEAs, the iterative refinement of the solutions by the proposed framework requires the definition of several operators to 1) select individuals from the population for breeding (parents); 2) mix their encoded representations to yield new offspring solutions (crossover); 3) mutate the offspring to enforce genotypical diversity with respect to their parent individuals (mutation); and 4) retain in the population those individuals that are better as per the fitness of the optimization problem at hand (survivor selection). Since dealing with a multi-objective problem, the fitness of individuals must be assessed in terms of Pareto optimality. The definition of the criterion by which one individual is better than another in a space comprising several objectives has spurred a flurry of MOEA proposals in the related literature, thoroughly reviewed in several surveys on the matter. This being said, the framework considers four different MOEAs that have been widely proven to adapt and perform competitively in multiobjective optimization problems with mixed-type decision variables: NSGAII [38], NSGA-III [39], SPEA2 [40], and MultiObjective Cellular genetic algorithm (MOCELL [123]). For the sake of readability, details on their differential characteristics are briefly revisited here, although they were detailed in Chapter 2. However, MOCELL is a new introduction in this context. •NSGA-II evaluates the Pareto optimality of newly produced solutions by sorting them together with individuals already contained in the population as per their Pareto dominance (non-dominance rank) and diversity 3.3. Search Methodology 47 inside every rank. In doing so, old and new solutions are sorted according to an ascending level of non-domination in the space of objectives. If there is no room in the population to include all individuals within the last utilized rank, an estimator of the density of other solutions around every solution within this rank is used to select those individuals that are located in less dense areas of the objective space (hence, they are more diverse). The seminal version of the NSGA-II embraces the socalled crowding distance function as a density estimator: however, many other definitions for this estimator can be adopted instead. •NSGA-III can be regarded as an extension of the NSGA-II algorithm that is adapted to provide a better performance in problems comprising more than two objectives. The main difference lies on the population diversity, which is guaranteed by means of reference directions, i.e., a set of points used to guide the optimization process towards a specific region of the objective space. Reference directions divide the space of objectives in a number of sub-regions (divisions), so that each division is associated with a specific reference direction. These divisions are used by NSGA-III to drive its search towards regions of the objective space that are most promising. Hence, crowding distance is not used as a density estimator to retain individuals in the population. Instead, a dual criterion based on non-dominated sorting and environmental selection is used, the latter aimed to maintain diversity among the solutions maintained in the population. Since solutions evenly distributed along the Pareto-optimal front should be preferred, NSGA-III prioritizes solutions that are close to the boundary between sub-regions or divisions, as these are more representative of the different parts of the Pareto-optimal front as per the configured reference directions. •SPEA2 works in the same line as NSGA-II but associates a new measure of fitness (strength) with every individual 𝛀𝑝in the population. The strength indicates the number of other individuals in the population that are dominated by the individual at hand considering their values of the defined objectives, normalized by the size of the population 𝑃. This defined strength is then used to compute and assign a raw fitness to every solution based on the strength of solutions that dominate it. To discriminate between individuals featuring the same raw fitness value, a measure of local density is calculated in terms of the distance of every individual to their neighbors. Furthermore, SPEA2 utilizes an external repository of solutions (archive) where non-dominated individuals found during the search process are stored. The archive size is kept fixed throughout the search by virtue of a truncation operator. •MOCELL defines the population as a regular grid that drives the process of selecting individuals for breeding: only solutions that belong to the same neighborhood in the grid can be recombined to yield a new offspring. Different topologies can be established for the neighborhood, among which the original MOCELL version proposed in [123] used the Moore topology (the neighborhood of a solution located in a cell inside the grid is itself and the solutions of the eight cells surrounding it in 48 Chapter 3. Improving Urban Accessibility and Environmental Conditions using Graph Modeling and Multi-objective Optimization a two-dimensional grid). As in SPEA2, MOCELL employs an external archive to store non-dominated solutions discovered during the search process. However, such archived individuals are fed back to the population, replacing other solutions selected at random. Crowding distance is used as the density estimator for both the population (for selection) and the archive (for truncation). As can be inferred from the above description, the algorithmic differences between such MOEAs do not affect the evolutionary operators used to recombine and mutate newly improvised solutions along the search. Instead, such differences concentrate mainly on the criteria used to select individuals for breeding and to retain good solutions over the search. Therefore, the same procedures for recombination and mutation can be adopted for the three of them, favoring a fair algorithmic comparison between them. An adapted version of the Simulated Binary Crossover (SBX) operator [124] with probability 𝑃𝑐and distribution index 𝐼𝐷are used. For integer variables (𝑣𝑝 𝑘,𝑣′𝑝 𝑘,𝑡(𝑣𝑝 𝑘, 𝑣′𝑝 𝑘)) a random mutation with probability 𝑃𝑖𝑛𝑡 𝑚is utilized, such that the value of every variable is drawn at random from its corresponding alphabet (either the existing edges (𝑣𝑝 𝑘, 𝑣′𝑝 𝑘) in the graph Grepresenting the urban area under study, or the types of assets that can be deployed in the edge at hand). For real-valued variables (Ψ(𝑣𝑝 𝑘, 𝑣′𝑝 𝑘)) a polynomial mutation with probability 𝑃𝑟𝑒𝑎𝑙 𝑚applied to every part of the solution. Finally, all MOEAs utilize a binary tournament selection method to select which individuals to breed from the population (NSGA-II, NSGA-III, SPEA2) or from the neighborhood of the given cell in the cellular grid (MOCELL). The framework can seamlessly accommodate other evolutionary search operators suited for the mixed type of decision variables defined in the problem statement. 3.3.3 Overall Algorithmic Flow of the Framework Once the search procedures have been defined, now all the steps needed for the application of the proposed framework to the improvement of the accessibility and environmental conditions for older pedestrians in a new target area 𝐴can be defined. Such steps are described in Algorithm 1. To begin with, the starting point is the definition of the area under study, which can be specified in several ways by the user of the framework (e.g., as a bounding box delimited by geographical coordinates). It is necessary that air and noise pollution measurements over the area under study are available so as to have a rough estimation of the pollution present on every edge of its graph. The user should also specify several points of interest (both origin and destination) for older pedestrians in 𝐴, so that urban routes connecting such points of interest are utilized for the computation of the objectives to be optimized. Other requirements to be met beforehand also include the technical specifications of the assets that can be installed in the area (including attenuation of the noise/air pollution intensity achieved by panels, nominal speeds of mechanical accessibility assets, their maximum installation length and economical cost models). Such requirements can be queried to technological suppliers of the assets. 3.3. Search Methodology 49 The user of the framework must also indicate the relative importance of accessibility (𝜆𝑎𝑐𝑐) and the environmental conditions (𝜆𝑎𝑖𝑟 and 𝜆𝑛𝑜𝑖𝑠𝑒) in the discovery of optimal routes between the established origin and destination points of interest. This can be easily introduced by the user by using a triangular system of barycentric coordinates, ensuring automatically that any point inside the triangle satisfies 𝜆𝑎𝑐𝑐 +𝜆𝑎𝑖𝑟 +𝜆𝑛𝑜𝑖𝑠𝑒 =1. Once these requirements have been fulfilled, the framework starts by computing the graph G={V,E} corresponding to the area 𝐴(line 1 in Algorithm 1). The simplest version of this graph can be constructed from GIS databases by devising vertices or nodes Vas street intersections/junctions, and by representing streets themselves as edges E. The use of GIS databases can allow for the straightforward annotation of UTM and elevation data of every vertex in V. More detailed versions of this graph can be achieved by fusing spatially collected fine-grained geographical data from other information sources, such as LiDAR (Light Detection and Ranging) measurements. Once graph Ghas been constructed, the user inserts the origin and destination nodes of the routes to be later optimized (line 2), and establishes the nominal noise and air pollution levels of every edge in G(line 3). For this latter purpose, measurements taken by public transport services and/or fixed air/sound quality monitoring stations can be used, together with methods to extrapolate spatially the pollution levels records to non-monitored regions of the urban area. As a result of this extrapolation, nominal pollution levels of every edge in the graph are computed. After initializing at random a 𝑃-sized population of interventions (line 4), the framework evaluates the objective values associated with these initial guesses of the Pareto-optimal interventions. To this end, assets within each intervention in the population are deployed on graph G(line 6), so that weights reflecting the accessibility and the environmental conditions of every edge resulting from the installation of such assets can be refreshed (line 7). The graph with updated edge weights can be used to compute the optimal route between every pair of origin-destination nodes (𝑣𝑜, 𝑣𝑑) ∈ V𝑜× V𝑑via the Dijkstra algorithm [125] (line 9), and evaluate the value of the accessibility, air and noise pollution associated to every route (line 10). By averaging such objectives over all routes (line 11), each intervention 𝛀𝑝is evaluated in terms of the objectives defined for the problem. The fourth objective (economic cost of the intervention) is computed as in Expression (3.21) (line 12). The rest of the steps in Algorithm 1(lines 13 to 20) comprise the refinement of the initial set of interventions by the chosen MOEA algorithm. This evolutionary search consists of the iterative application of the aforementioned selection (line 15), crossover and mutation operators (line 16) on the individuals in the population over 𝑚𝑎𝑥𝐺𝑒𝑛 generations. Before filtering and retaining good individuals in the population (line 18), newly produced interventions are evaluated (line 17) based on the defined set of origin and destination points of interest, following the procedure explained previously (lines 6-12). Once 𝑚𝑎𝑥𝐺𝑒𝑛 generations of this evolutionary search process have been completed, the framework returns the 50 Chapter 3. Improving Urban Accessibility and Environmental Conditions using Graph Modeling and Multi-objective Optimization Algorithm 1: Algorithmic flow of the proposed framework Input: Urban area under study 𝐴,[𝜆𝑎𝑐𝑐, 𝜆𝑛𝑜𝑖𝑠𝑒, 𝜆𝑎𝑖𝑟 ](weights indicating the relative importance of accessibility and environmental conditions when routing pedestrians), [𝐶𝑓 𝑖𝑥 (𝑡), 𝐶𝑣𝑎𝑟 (𝑡)] ∀𝑡∈ {𝑅, 𝑆, 𝐸 , 𝑆𝑃, 𝑃𝑃}(cost model for every asset type), 𝐾𝑚𝑎𝑥 (maximum number of assets per intervention), 𝑃𝑐,𝑃𝑖𝑛𝑡 𝑚,𝑃𝑟𝑒𝑎𝑙 𝑚,𝑃(MOEA parameters), 𝑚𝑎𝑥𝐺𝑒𝑛 (maximum number of search generations), [𝜏𝑛𝑜𝑖𝑠𝑒, 𝜏𝑎𝑖𝑟 ](reduction of noise and pollution per unit length when panels installed), [𝑆𝑅 𝑎𝑐𝑐, 𝑆𝐸 𝑎𝑐𝑐](nominal speed of a ramp/escalator), 𝑙𝑡 𝑚𝑎𝑥 ∀𝑡∈ {𝑅, 𝐸, 𝑆𝑃, 𝑃𝑃}(maximum length of ramp/escalator/sound panel/pollution panel) Output: Set of Pareto-optimal interventions balancing accessibility, air pollution, noise pollution and cost 1Extract graph of the target area: 𝐴↦→ G // From external databases 2Define points of interest: origin (V𝑜) and destination nodes (V𝑔) 3Retrieve nominal noise (𝑓𝑛𝑜𝑚 𝑛𝑜𝑖𝑠𝑒 (·)) and air pollution levels ( 𝑓𝑛𝑜𝑚 𝑎𝑖𝑟 (·)) of all edges Ein G 4Initialize at random a population of interventions: {𝛀𝑝}𝑃 𝑝=1 5foreach 𝛀𝑝in the population do 6Reset Gand deploy assets [𝑣𝑝 𝑘, 𝑣′𝑝 𝑘, 𝑡(𝑣𝑝 𝑘, 𝑣′𝑝 𝑘),Ψ(𝑣𝑝 𝑘, 𝑣′𝑝 𝑘)]𝐾𝑝 𝑘=1 7Update weights 𝑓𝑎𝑐𝑐 (·),𝑓𝑎𝑐𝑐 (·) and 𝑓𝑎𝑐𝑐 (·) of edges in G 8foreach (𝑣𝑜, 𝑣𝑑) ∈ V𝑜× V𝑑do 9Compute min-cost route E𝑜,𝑑 from 𝑣𝑜to 𝑣𝑑through G based on 𝑓𝑇(·) given in Expression (3.1)// Dijkstra 10 Evaluate 𝑓𝑎𝑐𝑐 (·),𝑓𝑛𝑜𝑖𝑠𝑒 (·) and 𝑓𝑎𝑖𝑟 (·) of route E𝑜,𝑑 11 Average the objective values over all routes (Expr. (3.18)-(3.20)), yielding 𝑓𝑎𝑐𝑐 (𝛀𝑝|V𝑜,V𝑑),𝑓𝑎𝑖𝑟 (𝛀𝑝|V𝑜,V𝑑)and 𝑓𝑛𝑜𝑖𝑠𝑒 (𝛀𝑝|V𝑜,V𝑑) 12 Compute cost 𝑓𝑐𝑜𝑠𝑡 (𝛀)of 𝛀𝑝as per Expression (3.21) 13 Initialize generation counter: 𝑔𝑒𝑛 =0 14 while 𝑔𝑒𝑛 < 𝑚𝑎𝑥𝐺𝑒𝑛 do 15 Select parent individuals from the population // Binary tournament 16 Crossover and mutation to produce offspring // SBX, random mutation 17 Evaluate objective values of offspring solutions (lines 6 to 12) 18 Retain fittest individuals in population // As per the selected MOEA 19 Update number of generations: 𝑔𝑒𝑛 =𝑔𝑒𝑛 +1 20 Return the set of non-dominated individuals in the population subset of interventions in the population that are not dominated in the space of objectives, declaring it to be its best approximation of the Pareto 3.4. Experimental Setup 51 front between the defined objectives (line 20). 3.4 Experimental Setup In order to verify the practical applicability of the devised framework in real-world scenarios, the performance will be evaluated over two use cases settled on the city of Barcelona (Spain). The choice of this city for the definition of the use cases finds its rationale in the public availability of data-sets with real air and noise pollution measurements, which is a requirement to be input to the framework (Figure 3.1). The two use cases are areas with a high percentage of older people in their population, a hilly topography and different types of walkable streets. In particular: •The first use case (hereafter denoted as 𝐴1) is located in the neighborhood of Can Baró and the north of Baix Guirnardó, comprising an area bounded by the latitude and longitude coordinates [41.4191◦,41.4105◦] and [2.1734◦,2.1510◦], respectively. This area is known for its steep slopes, especially in the hill of Can Baró, which pose a great obstacle for older people and citizens with reduced mobility. As for the northern area of Baix Guirnardó, it is full of noisy avenues undergoing heavy traffic all day long. •The second use case (correspondingly, 𝐴2) covers the area bounded by [41.4439◦,41.4342◦](latitude) and [2.1748◦,2.1621◦](longitudes), known as La Guineueta and Can Peguera. It is a less hilly area with lower traffic intensity than the previous use case. However, this neighborhood suffers from population aging, being among the ones with highest percentage of people above 65 years with respect to the population of the entire neighborhood (as of 2018, between 26.79% and 30.50% [126]). Once these areas were selected for experimentation, elevation data was retrieved from the National Geographic Institute (IGN) of the Spanish Government [127]. This repository provides public access to Digital Terrain Model (DTM) tiles featuring a resolution equal to 2 meters. Tiles corresponding to the use cases were retrieved, yielding an accurate 3D model of the area covered by the use cases. The graph Gof each use case was created by using the OSMnx library [128], which automates the process of querying and downloading data from the well-known Open Street Maps mapping service [129]. Elevation data 𝑧(𝑣)for every node 𝑣∈ V was computed based on the information in the retrieved DTM tiles, assigning it as the elevation value of the DTM point geographically closest to the coordinates (𝑥(𝑣), 𝑦(𝑣)) of the node. Figures 3.2 (left) and 3.2 (right) depict the finally composed graphs of the use cases 𝐴1and 𝐴2(respectively), overlaid with a simplified map of the city of Barcelona. Air pollution and acoustic noise data were collected from the public data repository made available by the city council of Barcelona at https://opendata-ajuntament.barcelona.cat. The air pollutant considered for the two use cases was NO2, as it was the substance with highest concentration values measured among the three types of pollutants 52 Chapter 3. Improving Urban Accessibility and Environmental Conditions using Graph Modeling and Multi-objective Optimization data (PM2.5, PM10 and NO2) available in the repository. Nevertheless, air quality indices combining different pollutants can be also produced and used within the proposed framework. Use Case A2 Use Case A1 Barcelona Spain Europe 2.155 2.160 2.165 2.170 Longitude 41.411 41.412 41.413 41.414 41.415 41.416 41.417 41.418 41.419 Latitude 2.162 2.164 2.166 2.168 2.170 2.172 2.174 Longitude 41.436 41.438 41.440 41.442 41.444 Latitude 75 100 125 150 175 200 225 250 Height Figure 3.2: (Top) Overall geographical location of the considered use cases; (bottom) geo-referenced graphs of (left) use case 𝐴1; (right) use case 𝐴2. For both cases, nodes have been colored as per the height of their locations in the real world. In the case of 𝐴1a wider range of colors can be observed due to the steep hill of Can Baró. Following the procedure described in Algorithm 1, origin and destination nodes are selected by inspecting the urban areas under study and detecting points of interest over the area. For example, when targeting accessibility for older people, points of interest may include the addresses of nursing homes or senior centers, entrances to a park, drug stores or an ambulatory. Origin and destination nodes are assigned to the vertices 𝑣∈ V whose coordinates (𝑥(𝑣), 𝑦(𝑣)) are closer to the selected points of interest. For both cases of study |V𝑜|=5origin nodes and |V𝑑|=6destination 3.4. Experimental Setup 53 nodes were finally established, giving rise to 30 routes over which to compute the objective values of the interventions optimized by the framework. The number of nodes, and hence of routes, was established accordingly to urban planning experts, considering population statistics and how the older population is distributed spatially in the neighborhoods and near the points of interest. Such points of interest include: •Use case 𝐴1:Residencia Vil La Salut (a home for the elderly), Parque Guinardo (a park), Residència Activa Parc de les Aigües (another nursing home), Hospital Sant Pau,Farmàcia Busquets Balsells /Hospital Hestia Gracia and Hospital de la Esperanza / Farmacia (the last two points of interest include a drugstore in addition to a hospital). •Use case 𝐴2:Farmacia Xavier Solani /El taller de costura (a drugstore), Clinica Phar (a dentist) / Farmacia Izquierdo Ridorsa (a second drugstore), Oficina de Atencion Ciudadana del Distrito Nou Barris (a citizen service office), Parque Canyelles (a park), Mercado de Canyelles (a grocery market) and Parc Central (another park). Table 3.1 summarizes the values of the parameters that are required for the application of the framework. Without loss of generality, the same values have been used for the two different cases of study, which follows from the fact that both are formulated over the same city. It should be noted, however, that such values could vary when applying the framework in other cities/countries, with supply companies imposing other cost models and offering different technological assets than the ones considered in this experimentation. Nevertheless, the design flexibility of the proposed framework can accommodate such variations. Parameter Value Description 𝛽𝑎𝑐𝑐 1.1m/s at slope −2◦Maximum speed of pedestrian 𝜃𝑚𝑎𝑥 8◦Maximum walkable slope for an older pedestrian 𝜃𝑅 𝑚𝑎𝑥 12◦Maximum slope for the installation of a ramp 𝜃𝐸 𝑚𝑎𝑥 20◦Maximum slope for the installation of an escalator 𝑆𝑅 𝑎𝑐𝑐 0.4m/s Nominal speed of a ramp 𝑆𝐸 𝑎𝑐𝑐 0.5m/s Nominal speed of a escalator 𝑆𝐿 𝑎𝑐𝑐 1m/s Nominal vertical speed of a lift 𝜏𝑛𝑜𝑖𝑠𝑒, 𝜏𝑎𝑖𝑟 ,0.7Attenuation factor for noise and air pollution 𝑙𝑡 𝑚𝑎𝑥 100 m (𝑡∈ {𝑅, 𝐸}), 500 m (𝑡∈ {𝑆𝑃, 𝑃𝑃})Maximum installable length of asset 𝐶𝑓 𝑖𝑥 (𝑡)25ke(𝑡=𝑅), 40ke(𝑡=𝐸), 100ke(𝑡=𝐿), 100e(𝑡∈ {𝑆𝑃, 𝑃𝑃})Fixed cost model for assets 𝐶𝑣𝑎𝑟 (𝑡)6.8ke/m (𝑡=𝑅), 7.5ke/m (𝑡=𝐸), 10ke/m (𝑡=𝐿), 40e/m (𝑡∈ {𝑆𝑃, 𝑃𝑃})Variable cost model for assets Table 3.1: Values of the parameters configured for the experiments. Regarding the search process, the four different MOEA described in Section 3.3, namely, NSGA-II, NSGA-III, SPEA2 and MOCELL have been chosen. Implementations available in the jMetalPy library [130] have been used. The population size is set equal to 𝑃=100, whereas the iterative search process is stopped after 𝑚𝑎𝑥𝐺𝑒𝑛 =2·104generations. The maximum number of assets per intervention is set to 𝐾𝑚𝑎𝑥 =60. Crossover 60 Chapter 3. Improving Urban Accessibility and Environmental Conditions using Graph Modeling and Multi-objective Optimization between the selected origin and destination nodes. Together with other saturated streets (e.g., Carrer de les Camélies or Carrer de Tenerife), this street arrives at the Plaça de la Font Castellana, a roundabout with typically high intensity of traffic all day long. Therefore, the selection of these two edges for the deployment of air pollution panels can be thought to match suitably the traffic intensity patterns of the roads traversing these streets. Figure 3.7: Qualitative inspection of the intervention with highest air pollution improvement among those optimized by SPEA2 over use case 𝐴1. The interpretation of the plots included in this figure can be done as in Figure 3.6. Heatmaps of the noise intensity levels of the area in which the assets are suggested to be deployed are included in each case. Now the scope of the discussion is shifted towards the solution best improving the noise pollution of paths between the selected origin and destination nodes of use case 𝑆1. This solution is shown in Figure 3.8, in the same format and interpretation protocol as in the previous figures.The solution is observed to comprise 3 different edges (streets) of the area under study for the deployment of noise attenuating panels: i) Carrer de Francesc Alegre, which converges to the aforementioned Mare de Déu de Monserrat avenue and has a petrol station (shown in the figure) as a clear source of ambient noise; ii) Carrer de Cartagena, which, in addition to its high traffic intensity, suffers from a high level of ambient noise due to the presence of a taxi stop and the main entrance of the Fundaci’o Puigvert (a university hospital), which ensures a regularly massive affluence of pedestrians; and iii) Carretera del Carmel with Carrer d’Ana M. Matute Ausejo, an area where a parking zone for motorbikes is allocated. These locations proposed by the framework are actually in areas with considerable acoustic stress (>70 dB), hence validating the framework also for this objective. It is also worthwhile noting that the installation of these assets does not affect heavily in the cost, yielding one of the cheapest solutions given by the framework for this use case. This is due to the large differences in terms of 3.5. Results and Discussion 61 installation costs assumed for noise/air pollution panels and accessibility infrastructures, such as elevators. Figure 3.8: Qualitative inspection of the intervention with highest ambient noise pollution improvement among those optimized by SPEA2 over use case 𝐴1. Qualitative Assessment of Interventions: Use Case 𝐴2 A similar line of reasoning can be followed when inspecting in depth the interventions evolved by SPEA2 over the second use case under study in the experimental setup. As such, Figure 3.9 depicts the intervention leading to the largest accessibility improvement among those composing the Pareto front approximation found by this solver. This intervention comprises 4 assets concentrated in two different parts of the scenario: the exit of a public parking space close to the metropolitan train station of Canyelles, and the access to the station itself. It is straightforward to note that as in the previous use case, this depicted intervention evinces, on one hand, that these locations in the area actually require a mechanical asset to overcome evident accessibility issues. On one hand, the exit of the parking space is not urbanized and is steep for a pedestrian with mobility constraints. On the other hand, the difference in height between the pedestrian access to the metropolitan train station and the street level is identified as an edge that requires an asset to improve its accessibility. However, the photo corresponding to this location reveals such an asset is already installed in practice (mechanical stairs), thereby validating the interventions optimized by the proposed framework. The discussion on the interventions optimized for the use case 𝐴2follows in Figure 3.10, which illustrates two different solutions to the estimated Pareto front. The map and annotated plots on the leftmost part of the figure correspond to the intervention corresponding to the lowest (best) value of the noise pollution objective 𝑓𝑛𝑜𝑖𝑠𝑒 (·). By contrast, the rightmost 62 Chapter 3. Improving Urban Accessibility and Environmental Conditions using Graph Modeling and Multi-objective Optimization Figure 3.9: Qualitative inspection of the intervention with highest accessibility improvement among those optimized by SPEA2 over use case 𝐴2. part of the image depicts the map of suggested assets for the solution featuring the best value of the air pollution objective 𝑓𝑎𝑖𝑟 (·). Regarding the first depicted solution, its genotype surprisingly contains no assets to be deployed, which makes this solution coincide with that implying minimum cost. The reason for this unexpected result is that the framework discovers that when not inserting any assets, the shortest routes between the selected origin and destination nodes under such circumstances already go through urban areas with minimal ambient noise. To shed light on this statement, a plot is annotated in the map showing the optimal routes between the public parking space (orange) and the two other points of interest (blue) corresponding to the case with largest accessibility improvement (the one already discussed in Figure 3.9). It can be seen that if accessibility assets were deployed in the exit of the parking space and the access to the metropolitan train station, shortest paths between these selected nodes would change, making them flow close to the Via Favéncia and the Ronda de Dalt, two of the roads in this area with the highest daily traffic profiles. If such assets were no installed, the shortest paths change to flow through residential areas, becoming longer to be traversed in terms of distance, but much better in terms of acoustic pollution. This unexpected result supports the inherent utility of the framework to discover counter-intuitive yet reasonable interventions. The discussion continues in the same figure, focusing on the solution analyzed on the right (largest improvement in terms of air pollution). In this second case, the genotype suggests the installation of 4 pollution panels in different parts of the urban area: i) the exit of the Plaça de Karl Marx roundabout, which is relevant given its centrality in the road network of the area and its closeness to one of the selected points of interest; 3.5. Results and Discussion 63 Figure 3.10: Qualitative inspection of the intervention with the highest noise improvement (left) and air pollution improvement (right) among those optimized by SPEA2 over use case 𝐴2. It is important to notice that the solution leading to the highest noise improvement does not involve installing any asset, since the framework determines, through its evolutionary search, that interventions to improve the other objectives imply a change in the optimal routes between origin and destination nodes, leading to a higher exposure of the pedestrian to ambient noise. ii) the junction between Rambla del Caçador and Carrer de l’Isard, a residential area close to a parking lot and the Plaça de la República, a roundabout with intense traffic; and iii) a private access to a parking lot whose exit flows into Carrer de Góngora, a street located in a densely populated neighborhood of the area that is close to the Ronda de Dalt main road. It is crucial to understand that since the area covered in 𝐴2is more residential than 𝐴1, and that no accessibility assets are suggested in the solution currently under discussion, shortest paths between origin and destination nodes already traverse areas with relatively low air pollution. Consequently, pollution panels suggested to be deployed by this solution are mostly affected by their proximity to sources of pollution (e.g., parking lots, roundabouts). The qualitative examination of the interventions finishes by analyzing the intervention with the highest economic cost found by the proposed framework over use case 𝐴2, which is depicted in Figure 3.11. Five different accessibility assets are proposed: besides the access to the metropolitan train station and the access ramp to the public parking space (already discussed in Figure 3.9), three new locations are pinpointed by the intervention: i) a long non-mechanical ramp (Carrer de Can Esenya) located close to one of the points of interest; and ii) two consecutive long stairs located in a neighborhood near the public parking lot and the metropolitan train station. Besides further buttressing that the locations suggested by the framework make practical sense, the fact that the solution involving maximum cost is dominated by accessibility assets is a consequence of the 64 Chapter 3. Improving Urban Accessibility and Environmental Conditions using Graph Modeling and Multi-objective Optimization Figure 3.11: Qualitative inspection of the intervention requiring the higher economical cost among those optimized by SPEA2 over use case 𝐴2. In this last depicted case, the higher installation costs assumed for accessibilityrelated assets imply that the most expensive intervention consists of 5 assets to be deployed in physical locations characterized by accessibility issues, as the annotated photographs clearly expose. higher fixed cost model realistically assumed for this type of assets. For non-extreme investment costs, however, it is not straightforward to decide which type of asset to install and where, nor is it simple to ascertain the consequences of such installations in air pollution, acoustic noise and accessibility. There lies the core purpose and inherent utility of the proposed framework: to automate and optimize the interventions accounting for these multiple criteria, as clearly shown by the qualitative examination of the interventions done in this section. 3.5.3 Considerations for the Application of the Framework in New Case Studies Several aspects must be taken into account when deploying this framework in future practical case studies. Quality and representativeness of available data. Firstly, it is critical to thoroughly examine the relevance and quality of the data available for each case. Unfortunately, despite the hype of Smart City and Open Data Platforms in the past decade, not many cities provide high-quality datasets that allow for the application of frameworks as the one proposed 3.5. Results and Discussion 65 in this Chapter 3. In some cases, data is not captured with enough resolution (spatial and/or temporal) to allow for fine-grained analyses. In addition, historical datasets are not properly archived or made publicly available. For instance, air quality and noise data need to reflect typical situations for the areas under study, therefore excluding non-typical data deriving from particular events or circumstances (e.g., temporary road or construction works). If air quality and noise models were available for future case studies, the algorithm could be fine-tuned to better account for specific attenuation factors, therefore including the impact of installing each asset into the model. Other external data sources. Moreover, as mentioned before, other relevant data sources should be fed to the framework when possible, such as the location and characteristics of urban accessibility infrastructure, or the public or private nature of certain paths and access points included in the graph modeling the urban area under study. Otherwise, the optimal routes output by the framework could eventually traverse segments that are actually close to the public. A similar observation can be made in regard to the spatial granularity of pollution and noise data: depending on the sensing equipment used for the measurement campaign, noise and pollution maps could eventually be accurate in certain parts of the urban area, whereas in other areas data interpolation strategies should be enforced. The flow-like propagation nature of pollution and the blockage of noise by buildings could make this interpolation computationally very costly, requiring complex simulation stages. Computational complexity. In partial connection to the above is the computational effort required for the execution of the framework. Experiments performed to yield the results reported in this chapter were run over an implementation of the framework that is not optimized for nearimmediate decision making. The reason is that the characteristics of the assets composing the intervention does not require a real-time operation of the framework optimizing their location. However, it is possible to formulate alternative scenarios related to accessibility and walkability that could require shorter running times of the framework, such as those implying decision variables that can be tuned in an agile fashion (e.g., vehicles for on-demand transport services) or those involving input data flows with short prevalence periods (namely, short-term weather forecasts). Nevertheless, the algorithmic phases of the workflow described in Algorithm 1can be easily implemented in parallel, from the discovery and evaluation of the routes to the application of the evolutionary search operators of the metaheuristic solver at hand. Furthermore, certain design choices such as the number of origin and destination nodes |V𝑜|and |V𝑑|can further be tuned to alleviate the computational burden of the framework and to make it compliant with the latency constraints imposed for decision making. All in all, the computational efficiency of the framework should match the latency requirements imposed by the optimization objectives, 66 Chapter 3. Improving Urban Accessibility and Environmental Conditions using Graph Modeling and Multi-objective Optimization the speed of implementation of the decision variables involved in the problem statement, and the prevalence of the external information flowing into the system. Optimization for other walkability/accessibility scenarios. Despite not mentioned throughout the chapter, the framework here described is also capable of optimizing routes for wheelchair users and people with other mobility issues. As it is a versatile and easily configurable tool, in the case of these users it would be enough to modify the parameters of the optimizer which affect them significantly, such as the maximum slope value, according to the criteria of the experts, or the speed. The urban elements to be inserted could also be adapted to the users, for example, disregarding the option of escalators or giving greater importance to the installation of elevators. When considering such alternative scenarios, incorporating new external sources of data would become rather a necessity than a possibility, such as weather conditions or seasonal solar shadowing at street level. On the generalization of the benchmark results to other use cases. Finally, it is important to bear in mind that in other practical use cases other relationships might exist between the considered objectives. Consequently, one should not expect that the best performing MOEA found in this contribution replicates its outstanding performance in other setups. Therefore, a new algorithmic benchmark like the one shown in the experiments of this chapter should be followed. Furthermore, an interactive interface suitable for the audience of the framework would be convenient to facilitate the assessment of the solutions by users that are not necessarily experts in multi-objective optimization. This being said, enabling a qualitative inspection of the solutions composing the approximated Pareto front estimated by the algorithm would ease this process and redound to the trustworthiness of the users in the platform. 3.6 Summary Cities face unresolved challenges when pursuing to become friendly spaces for their citizens, especially for older people. The creation of age-friendly urban areas requires crucial actions to improve their accessibility and to enhance the environmental quality (noise, pollution) perceived by pedestrians when traversing the urban environment. To technologically support decisions to be made for this purpose, this Chapter 3 has presented a novel framework that embraces multi-objective evolutionary optimization and graph modelling to optimally determine where to install urban assets that help realize the aforementioned objectives, also accounting for the cost efficiency of their installation. Such conflicting objectives delineate a Pareto trade-off comprising different assets/interventions that differently balance the goals sought in the formulated problem: minimum air pollution, minimum ambient noise, best accessibility, and minimum cost. Evolutionary metaheuristics for multi-objective optimization are then adopted 3.6. Summary 67 to efficiently explore the mixed integer-real search space of the problem, determining which assets to deploy, where they must be installed in the area under study, and their specific functional features. The output Pareto front approximation elicited by this new framework can inform experts in urban planning during their decision-making processes related to accessibility and environmental quality, reducing the time, and improving the quality of decisions made in this regard. Moreover, this output can be used to provide objective grounds for planning and investment decisions made in the context of the long-term General Urban Plans, for instance including demographic projections and planned infrastructures as inputs for the model. It has been experimentally showcased the applicability of proposed framework in two real-world use cases located in the city of Barcelona (Spain), considering real air pollution, environmental noise, and topographical data collected over this city. Experiments have confirmed that: •Clear performance gaps exist between different metaheuristic optimization algorithms when used within our framework, as can be deduced from the reported statistics on four quality indicators, and a posterior significance analysis of differences observed among them. •Interventions evolved by the framework conform to intuition and common sense for each of the use cases, proposing the installation of assets to improve all the considered objectives in locations with real issues in terms of accessibility, noise and air pollution. The design flexibility of the framework and the promising results herein discussed stimulate several research directions to be pursued in the near future. To begin with, a major effort will be invested towards overcoming the practical limitations identified in Section 3.5.3: specifically, means for a better extrapolation of air pollution and environmental noise data at the street level will be studied by gauging the impact of the urban topology in the propagation of noise and air flow. Likewise, a wider portfolio of data sources should be considered to enrich the graph model constructed for the area at hand, including public/private access, already installed assets or other parameters from where additional optimization objectives can be formulated. For instance, by including an estimation of the sunlight exposure at street level for the area from a 3D dynamic model, the framework could also choose paths taking shade into account, or seek the location where to install climatic shelters for the sake of the thermal comfort of the pedestrian along his/her path. Furthermore, the inclusion of other external data sources will enable the extrapolation of the framework to optimize other forms of walkability/accessibility, comprising objectives and decision variables that may differ with respect to the assets herein considered (e.g., wheel chair users or on-demand transport services for older people). Finally, use cases in other regions and countries should be selected to validate the framework, considering local particularities (e.g., prevalent regulatory constraints, cultural differences, and weather conditions) that may require a reformulation of the problem and the adaptation of the framework for their citizens to fully benefit from its application. 69 Chapter 4 Efficient Estimation of Ground-Level Air Temperature in Urban Areas using Machine Learning Since the early 19th century, human activities have contributed to a marked intensification of the greenhouse effect, a natural phenomenon in which atmospheric gases trap heat and maintain the Earth’s surface temperature [138]. This human-induced amplification, referred to as the anthropogenic greenhouse effect [139], has led to a pronounced acceleration in global warming, with significant implications for the Earth’s climate system to date. This climate change is clearly reflected not only by the rise of the global average temperature, which has raised 1.2C◦in the last 200 years [11], but also in the increase in the amount of annual heatwaves days [11]. These worsening weather conditions severely affect human activities, ecosystems and health. Individuals that are exposed to high temperatures and heatwaves can have various detrimental effects on their health. These effects may include increased risks of heat stress and heatstroke leading to morbidity and mortality, as well as the degradation of respiratory and cardiovascular conditions [140], [141]. The heat-related mortality rate has experienced a significant increment of 53.7% in the past 20 years for those over 65 years [11]. Not only human health is affected by the high temperatures. The economic impact of the heat waves is reflected, for example, in the 302 billion working hours lost in 2019, which represents an increase of 51.8% compared to year 2000 [11]. A major problem directly affecting sustainability lies in the rising energy consumption of housing and industry. As heat consumption is closely related to temperature variations influenced by climate change, this poses a pressing concern for sustainable practices. For example, 40% of all energy consumed in Europe is consumed by the housing stock, as well as accounting for 36% of all CO2emissions [4]. 76 Chapter 4. Efficient Estimation of Ground-Level Air Temperature in Urban Areas using Machine Learning use of this type of technology gives a new tool to estimate surface air temperature in a much faster way than classical numerical approaches and more effortlessly, with potentials to be more scalable to other cities than the frameworks reviewed above. 4.2 Materials and Methods The method proposed in this chapter aims to estimate the Tawith a spatial resolution of 100 meters in an hourly basis. For this purpose a U-Net architecture is proposed, which correlates the different spatial and temporal datasets with the Tacomputed by UrbClim over the grid on which the problem is defined. In the next subsections, the study area (Subsection 4.2.1) and the different datasets used (Subsection 4.2.2) will be described, and the modelled architecture (Subsection 4.2.3) will be presented. 4.2.1 Study Area The study area, shown in Figure 4.1, covers the hole metropolitan area of the city of Bilbao (Spain) and part of the province of Biscay with an overall area covered of 25 ×25 km2. The area falls within the Cfb (i.e. temperate, no dry season, warm summer) Köppen-Geiger class [200]. The city is characterized by moderate temperatures in both summer and winter, due to the influence of the Atlantic Ocean. This implies a mean temperature of 14-15 ◦C, reaching maximum temperatures of 25-26 ◦C in July and August [201]. The Nervión river flows through the middle of the city and the valley, where the air masses are channeled providing the city with an important ventilation path. The breezes developed by both the topography and proximity of the sea are relevant as well. 4.2.2 Datasets In this subsection, we provide an overview of the datasets used to train the proposed model, which integrates both meteorological and spatial information to predict ground-level air temperature. This subsection introduces the various datasets employed, focusing on their structure, sources, and roles in the model training process. Four different datasets are presented to comprehensively describe the data pipeline: the Target Training Dataset, which details the UrbClim-simulated temperature data used as the primary target for the model; the Meteorological Dataset, derived from ERA5 reanalysis data and used to provide meteorological inputs for training; the Spatial Dataset, which comprises various geographical and morphological features that describe the urban landscape of the study area; and finally, the Validation Dataset, which contains real air temperature measurements collected from local weather stations and is used to evaluate the accuracy of the model’s estimations. These datasets, taken together, form the basis for both training and assessing the model’s capacity to accurately estimate Taunder varying meteorological and spatial conditions. 4.2. Materials and Methods 77 Target Training Dataset The target training dataset of the model was based on the dataset called “Climate variables for cities in Europe from 2008 to 2017” provided by the Copernicus Climate Change Service [202]. This dataset offers air temperature at a fine-grained resolution of 100 meters in an hourly basis from 2008 to 2017 for a set of 100 European cities. The temperature values were obtained through simulations with UrbClim model [143]. Meteorological Dataset Experiments later presented and discussed in this chapter utilize meteorological data coming from ERA5 Reanalysis [203], which is the same input dataset used by UrbClim. The reanalysis data was facilitated by the European Centre for Medium-Range Weather Forecasts through the Copernicus Climate Data Store in NetCDF file format. ERA5 offers global, hourly data spanning from 1940 to the present at a spatial resolution of 0.25° × 0.25°. Data between 1981 and 2017 was collected for the following surface variables: 2m air temperature, precipitation, specific humidity and wind components (uand vat 10m). Considering the coarse resolution of ERA5, grid cell values for each variable that fall inside the study area where averaged to obtain the corresponding 1981-2017 hourly time series. To reduce the climatic variability of the training dataset and focus the target of the model to those days in which the heat stress can be more present, the Local Weather Type (LWT) classification proposed by [204] is applied. The idea is that similar synoptic conditions lead to a similar thermal behaviour of the city and, hence, the model can focus on estimating how the Tabehaves under those particular conditions (i.e. under that particular LWT). To obtain the LWT several daily metrics are obtained from the above mentioned hourly variables for the baseline period 1981-2010: thermal amplitude (Ta,max - Ta,min), total precipitation, average specific humidity, and wind speed and direction. Wind direction was further processed to convert it into a categorical variable by classifying the values based on the following criteria: N (from -45ºto 45º), E (from 45º to 135º), S (from 135ºto 225º), and W (from 225ºto 315º). This classification slightly differs from the one applied by [204], because it follows the recommendations of the Basque Meteorological Agency for the identification of wind components [205]. 12 different LWT that are representative of the different seasons are obtained. Among them, the selected one was the most representative for summer (June, July, August and September) and the one that had the potential to generate adverse thermal situations. This cluster was identified as a sunny day with a weak (1.72 m/s) eastern wind component, high specific humidity (around 11 g/kg). After searching the summer days that belong to that specific LWT within the period 2008-2017, a total of 164 days were obtained. 78 Chapter 4. Efficient Estimation of Ground-Level Air Temperature in Urban Areas using Machine Learning Spatial Dataset The spatial datasets consist of different variables that define the morphological characteristics of the study area. These variables match the ones used by UrbClim to simulate the training dataset [206]: imperviousness and elevation of the terrain and land cover. The imperviousness and the land cover were obtained from the Copernicus Land Monitoring Service portal [207], which provides free European and global geographical information. The DTM was retrieved from the IGN of the Spanish Government [127] for convenience. This public repository distributes a DTM product at 200m, which is similar to the resolution of the Global Multiresolution Terrain Elevation Data (GMTED) that is originally used by UrbClim. Although the year of the versions do not match (2012 IGN vs. 2010 GMTED), the impact of the differences can be considered negligible. Among the variables used originally by UrbClim, two were not considered: the Normalized Difference Vegetation Index (NDVI) and the anthropogenic heat flux. Focusing only on summer days, and considering the NDVI’s seasonal behavior, all the training days are expected to have similar values. In the case of the anthropogenic heat flux, its spatial resolution (5km) was considered too coarse to get the finer spatial patterns derived from the distribution of, for example, traffic and buildings. Validation Against Real TaMeasurements For validation against real Tavalues, point measurements were also obtained from different weather stations that Euskalmet, the Basque meteorological agency, operates over the area under study. Tameasured values from a total of 7 weather stations were collected for validation in the time range from 2008 to 2017. La Arboleda, located in a mountain; Punta Galea, placed in a lighthouse; Galindo, Deusto and Zorrotza in urban areas near a river; and Arrigorriaga and Derio in urbanized rural areas. Figure 4.1 shows the locations of the weather stations, marked with a red dot. 4.2.3 Proposed Model Train and Test Samples Before explaining in detail the architecture used to correlate the Tawith the spatial and meteorological variables, the subsets of training and test data are described. The first step applied to all the datasets described in Subsection 4.2.2 was the normalization of the variables in the range [-1,1] in order to improve the interpretability of the data, using: 𝑥𝑛𝑜𝑟𝑚 =2𝑥−min 𝑥 max 𝑥−min 𝑥−1.(4.1) Once data was normalized, a subset of training and test data was created by dividing the use case area into several patches, as can be seen in 4.2. Materials and Methods 79 3.1°W 3.1°W 3.05°W 3.05°W 3°W 3°W 2.95°W 2.95°W 2.9°W 2.9°W 2.85°W 2.85°W 2.8°W 2.8°W 43.2°N 43.2°N 43.25°N 43.25°N 43.3°N 43.3°N 43.35°N 43.35°N 43.4°N 43.4°N Figure 4.1: The study area divided in patches. The patches shadowed in yellow correspond to the validation ones, while the purple ones denote test patches. The meteorological stations are located in the red dots. Figure 4.1. A total of 64 patches of 3.1 ×3.1 km were created. To ensure the representativeness of all map areas during model training, 53 patches were distributed for training, highlighted in black in Figure 4.1. From those 53 patches, 5 of them were reserved for model validation, highlighted in yellow in Figure 4.1. The locations for the test patches were chosen in accordance with the locations of the meteorological stations. The seven different stations lie into six different test patches, with the stations of Deusto and Zorrotza falling into the same patch. The test patches are highlighted in purple in Figure 4.1. Hence, from the 24 hours of the 164 days of LWT, 3936 hours are obtained to train the U-Net model. Consequently, each of the train and test tiles will have 3936 examples. U-Net Model The model used to estimate Ta, shown in Figure 4.2, is a variation of the well-known U-Net architecture, which consists of an encoder-decoder with skip connections. The input datasets of the encoder consist of the three 80 Chapter 4. Efficient Estimation of Ground-Level Air Temperature in Urban Areas using Machine Learning spatial maps described above. The normalized input tensor representing those variables has the shape of 32 ×32 ×3 for each patch of the map. As explained in Chapter 2 (Subsection 2.1.2) the U-Net is divided into two parts, the encoder and the decoder. The encoder, composed of CNNs, as well as pooling layers and dropout layers, reduces the dimensionality of the input tensor in aim to find different hidden relationships and extract high level features from it. Once the desired dimension is achieved, the decoder reconstructs the image, applying an inverse process. In this upsampling process, skip connections are added between the encoder and the decoder to help in a high-quality reconstruction of the image. For the case described in this chapter, three downsample blocks are used each consisting of double convolutional layer (with ReLu as the activation function), a pooling layer and a dropout layer. After down-sampling the input data the lowestdimensional space is reached, also known as latent space. At this stage, due to its low spatial resolution, the meteorological data is added. Both Ta and the rest of the meteorological variables at time step tare conditioned by the meteorological conditions of the previous hours. That is why, to estimate the Taof an instant t, the meteorological data of the previous 2 hours is taken into account. Hence, the meteorological vectors go from t-2 to t. The meteorological data of each time step is in inserted in a tensor of shape 5 ×1 to the reconstructed image, resulting in the final output map of the estimated Ta. Once the meteorological data is coupled to the spatial data, the upsampling starts. The decoding structure mirrors that of the encoder, comprising 3 up-sample blocks. Each block includes a pair of convolutional transpose layers, a concatenation layer, and a dropout layer and leads to a reconstructed image of shape 32 ×32 ×1. The U-Net contains different parameters that have to be tuned to obtain the optimal performance. For the case exposed here, a learning rate of 10−5, a batch size of 32, and 30 epochs are chosen. As the loss function, Mean Square Error (MSE) is selected: 𝑀𝑆𝐸 =1 𝑛 𝑛 ∑︁ 𝑖=1 (𝑌𝑖−ˆ 𝑌𝑖)2,(4.2) where 𝑌𝑖is the observed value and ˆ 𝑌𝑖the predicted value. 4.3 Results and Discussion In this section the results obtained for the U-Net model over the scenario defined over the city of Bilbao are presented and discussed. The discussed results aim to inform the responses to the three RQs formulated at the beginning of this chapter. Each of these is addressed individually in the following subsections. 4.3. Results and Discussion 81 64 128 256 512 512 256 128 64 32 x 32 x 3 Input data 32 x 32 16 x 16 4 x 4 4 x 4 8 x 8 32 x 32 16 x 16 Output 32 X 32 X 1 2 x 2 2 x 2 1024 1024 +5 x 1 = Temporal meteorological data t-2 t t-1 8 x 8 Figure 4.2: The U-Net architecture proposed for the estimation of Ta. The sum in the latent space corresponds to a concatenation of two flattened vectors. 4.3.1 RQ1: Can an AI-based Data Model Achieve Accurate TaEstimations that are Close to Those Elicited by a Numerical Model? In this first RQ the aim is to test whether the estimated values of Taare spatially consistent and with values that are close to the ones elicited by the UrbClim numerical model. We recall the reader that, as described in Section 4.2.2, the U-Net model is being trained to estimate the Tavalues provided by the Copernicus UrbClim simulations dataset. Hence, a good model will be that reducing the error between the Tavalues obtained from it and those from UrbClim. To assess whether this condition is met by our proposed U-Net architecture, four regression metrics between the two models are computed for each weather station area during the period of study. The metrics chosen are the Pearson Correlation Coefficient (PCC), the Root Mean Square Error (RMSE), the Mean Absolute Error (MAE) and the Mean Absolute Percentage Error (MAPE). The results can be found in Table 4.1. As it can be seen in the above table, for all the seven locations, the Pearson correlation is above 0.95, and for almost all of them is near or above 0.98 value. Considering that the definition of the Pearson correlation coefficient is a value between -1 and 1 that quantifies the linear correlation between two datasets, being 0 no correlation and -1 and 1 full 82 Chapter 4. Efficient Estimation of Ground-Level Air Temperature in Urban Areas using Machine Learning UrbClim - U-Net Station Pearson RMSE MAE MAPE Arboleda 0.98 2.71 2.13 18.83% Arrigorriaga 0.99 2.07 1.58 14.37% Derio 0.98 1.75 1.35 9.0% Deusto 0.98 2.24 1.74 8.95% Galindo 0.98 2.3 1.77 8.58% Punta Galea 0.95 2.51 1.95 12.47% Zorroza 0.98 2.29 1.77 9.0% Table 4.1: Regression metrics computed between the two models (UrbClim and U-Net) for each weather station during the period of study. correlation depending on the slope, it can be seen that the correlation between both is almost absolute. In other words, when in the reference model, Taincreases, in the other one also does it in the same or very similar way, and when in the reference model the temperature decreases, so does the proposed model. The location where the Pearson is lower (0.95, Punta Galea) may not be coincidental. Dealing with a cape in the Atlantic Ocean, temperature changes are mostly affected by the sea, making it potentially more challenging for the model to accurately model the patterns of the target variable Ta. We now steer our focus towards the other metrics. In the same table, we observe that both RMSE and MAE are floating around 1.5◦C to 2.5◦C of difference. As per its definition, RMSE penalizes more the large errors made by the model. Nonetheless, the highest RMSE is still 2.71◦C for Arboleda, an acceptable value. For the MAE, the values are lower than for the RMSE, around 1.7◦C, being the highest 2.13◦C also for Arboleda, and 1.35◦C the lowest, in Derio. Continuing with the quantitative results, Table 4.2 shows a comparison in terms of the same performance metrics between the U-Net and UrbClim models, and the real Tameasurements collected by the chosen meteorological stations. UrbClim - Real U-Net - Real Station Pearson RMSE MAE MAPE Pearson RMSE MAE MAPE Arboleda 0.87 7.14 5.99 33.21% 0.88 5.95 4.94 26.68% Arrigorriaga 0.94 6.01 4.75 25.13% 0.95 5.25 4.28 23.22% Derio 0.90 4.97 3.80 19.60% 0.92 4.70 3.63 18.44% Deusto 0.89 6.27 4.75 22.19% 0.91 5.56 4.37 20.98% Galindo 0.89 6.77 5.10 23.26% 0.91 5.84 4.53 21.10% Punta Galea 0.87 3.95 3.05 16.40% 0.88 4.24 3.25 16.73% Zorroza 0.89 5.75 4.20 20.13% 0.91 4.76 3.58 17.33% Table 4.2: Regression metrics computed for the two models (UrbClim and U-Net) with respect to the real temperature data collected by each weather station during the period of study. Best results for every station and score are highlighted in light blue. As can be noticed in the above table, results for the U-Net model are 4.3. Results and Discussion 83 slightly closer to the real measurements collected by all stations, except from Punta Galea. Nonetheless, the results are similar for both models. The UrbClim simulations exhibit spatial artefacts in their Tavalues, particularly in the form of exaggerated temperature gradients. When using the U-Net model for temperature prediction, its convolutional processing inherently smooths these artefacts. Although this smoothing results in a less accurate approximation of the UrbClim output, it has an unintended positive consequence. By blurring the artificial sharp gradients present in the UrbClim data, the U-Net model produces estimates that, despite being a poorer fit to the UrbClim simulations, align more closely with the actual temperature values observed in reality. This suggests that the UNet’s ability to mitigate spatial artefacts enhances its capacity to generate predictions that are closer to real-world temperature distributions, even though it technically underperforms in replicating UrbClim’s output. The Pearson correlation for the U-Net model is still quite high and encompasses values from 0.88 to 0.95, being most of them around 0.9. The model correlates the Tavalues reasonably well, with MAE for U-Net model with respect to the real Tarecorded by the weather stations oscillating between 3.25◦C and 4.94◦C. It should be noted that the model estimates Tawith a resolution of 100 meters, while the Tavalues given by the meteorological station are single-point measurements, sometimes collected in urban areas, which may not be as representative as the hole area covered by the estimated pixel of Ta. This can be regarded as an intrinsic source of error for the evaluation. There is no significant variations in the error between different locations. However, the highest error is located in la Arboleda, which is a mountain of 329 meters above sea level, while the lowest error is in Punta Galea, a meteorological station surrounded by the ocean. Moving towards qualitative results, Figure 4.3 shows the aggregated spatial distribution of the predicted Tafor 05:00 and 14:00 hours. We begin our discussion on the qualitative results with the map corresponding to 5:00, where a spatial consistency of temperatures can be observed in the tiles and their transitions between them. The first thing that can be observed in this plot is the clear influence of the sea as a large body of water that plays the role of temperature reservoir. The areas close to the sea and the estuary maintain milder temperatures (lighter blue). The effect of the sea is clearly and correctly represented. It can also be observed that the most densely populated areas and where the largest road networks are located are the ones that retain more heat. This is due to the land uses of these areas. The materials that cover these areas are generally asphalt and concrete, which accumulate more heat during the day. Also, the imperviouness of the land plays an important role. This difference of estimated Taas a function of land cover and imperviousness is the expected from the model, since it has been trained taking into account those datasets. Finally, it can be seen that the areas with more vegetation and height such as the mountains are those that reflect lower temperatures (dark blue). Specifically, the Ganekogorta mountain range (1000 meters) is the area where the lowest temperatures are estimated, according to what 84 Chapter 4. Efficient Estimation of Ground-Level Air Temperature in Urban Areas using Machine Learning occurs in practice. Hence, we confirm that the U-Net model learns the relationship between the different spatial datasets fed in the input of the model, such as height, and Ta. Following our qualitative analysis, we now focus on the map for the 14:00, which best reflects the thermal contrasts between different zones. The bright colors represent a higher temperature, and as in the early morning map, heat retention in urban areas is much higher than in green areas and mountains. In addition, the geographical idiosincrasy of Bilbao and its surroundings, nestled in a narrow valley, makes thermal differences be very large within short distances. Once again, the Ganekogorta mountain range stands out as the area with the lowest temperatures. In the same way, both the sea and the mountains are very well delimited by the model. At the cape of Punta Galea, in the upper right part, there is a clear change of estimated temperature between the sea area and the land area. 3.1°W 3.1°W 3.05°W 3.05°W 3°W 3°W 2.95°W 2.95°W 2.9°W 2.9°W 2.85°W 2.85°W 2.8°W 2.8°W 43.2°N 43.2°N 43.25°N 43.25°N 43.3°N 43.3°N 43.35°N 43.35°N 43.4°N 43.4°N 3.1°W 3.1°W 3.05°W 3.05°W 3°W 3°W 2.95°W 2.95°W 2.9°W 2.9°W 2.85°W 2.85°W 2.8°W 2.8°W 43.2°N 43.2°N 43.25°N 43.25°N 43.3°N 43.3°N 43.35°N 43.35°N 43.4°N 43.4°N Figure 4.3: Aggregated spatial distribution U-Net at 05:00 (top) and 14:00 (bottom). Now, we turn the focus towards analyzing in detail the results corresponding to the test tiles, comparing them with the estimations of the UrbClim numerical model. This analysis is done in Figure 4.4, where it is worth noticing the difference in the gradients of Ta. The first thing that comes to eye in all scenarios is the difference in the gradients of Ta. The results of the U-Net model show smooth transitions between temperature differences, with more homogeneous values and no abrupt changes. This does not occur in the numerical model. The reason relies on the nature of the convolutional operations that take part in the architecture of the U-Net. The convolution operation, as explained in Chapter 2 (Subsection 2.1.2), by definition, tends to reduce noise and to smooths gradients spatially. In the 5:00 maps, the spatial results obtained by the U-Net are very similar to those of the numerical model taken as reference. In the DeustoZorroza tile, it is able to discern between the temperature of the estuary, its influence in the margins, and the surrounding heights in a very similar way to the numerical model. In Derio, the colder zones that appear in 4.3. Results and Discussion 85 the numerical model are also localized by the U-Net model even though they have a complex geography. However, U-Net ensures a smoother spatial gradient of the estimated temperature which, as argued previously, guarantees a more plausible spatial distribution of this variable in the area under study. The same happens in the case of Galindo and Arrigorriaga. At Punta Galea, the U-Net manages to estimate different temperatures for the small area of land that appears on the right side of the tile. In the 14:00 maps, the same trend continues as in the previous ones. The spatial distribution of the Taare very similar in the two models. The U-Net model is able to differentiate small temperature patterns. For example, in the case of Arrigorriaga, in the center of the image can be seen a colder spot which is the representation of the Malmasin mountain that stands out over the whole area. A similar pattern is observed in La Arboleda, where the spatial distribution of lower Tavalues aligns with areas of higher altitude and greater vegetation coverage. At Punta Galea the protruding land area is still estimated quite accurately in both models. In the cases of Deusto-Zorroza and Galindo, however, due to the homogenization and smoothing process applied by the U-Net, spatial details are lost that do appear in the numerical model, such as the river part that appears in a lighter orange. Nonetheless, it can not be confirmed without real data measurement, that the sharp gradient in the numerical mode is correct in both cases. In conclusion, it is fair to state that the proposed U-Net model is capable of estimating the spatial distribution of Ta, achieving a plausible and accurate spatial distribution of this variable, and recognizing singular spatial patterns. 4.3.2 RQ2: Can an AI-based Data Model Consistently Estimate Temperature Over Time? One of the model’s aims is to be able to estimate the Tafor a resolution of 100 meters and with a temporal frequency of one hour for the 164 days chosen between 2008 and 2017. To answer RQ2, an hourly time series was extracted from the estimated Tavalues for different LWT days. Then, the estimated Tavalues at the same locations as the weather stations were compared both with the weather station data and with the data provided by the UrbClim model. The time series are presented in Figure 4.6. The trends of the day-night cycles closely resemble those of reality. However, in both models (U-Net and UrbClim) a clear overestimation of the temperature is observed during the day cycle. This overestimation, which in some specific cases such as Deusto or Galindo can reach more than 6◦C, generally remains at about 3-4 ◦C, similar to the errors shown above in Table 4.1. During the night cycle, estimations vary depending on the weather station and both cases of overestimation and underestimation can be found. As shown in these plots, the U-Net model effectively reproduces the temperature patterns of the numerical model it was trained to approximate, demonstrating a strong capacity to mirror UrbClim outputs on an 92 Chapter 4. Efficient Estimation of Ground-Level Air Temperature in Urban Areas using Machine Learning these areas do not dissipate heat during the night and are subject to strong heat stress. Other two hotspots (bottom and top-right of the image, shadowed in yellow and orange respectively) are related to heavy industry and industrial parks. The last hotspot (center-left, shadowed in red) is close to the junction of two important highways, which generates a large asphalted surface surrounded by green areas, leading to a high gradient in Ta. 4.4 Summary In the next decade, cities will inevitably have to implement measures that will lead them to become more sustainable and resilient places. As seen throughout this chapter, Taplays a pivotal role in this desired urban sustainability. Urban planners face the challenge of implementing actions that produce tangible impacts and benefits, requiring access to as many resources as possible. In this context, having access to fine-grained, hourly Taestimations for a city represents a valuable resource for their decisionmaking processes. Furthermore, if urban planners have access to tools that provide relevant information in a short time, the efficiency and speed of decision-making processes can be significantly improved. In this context, this chapter has presented a new model for estimating air temperature at ground level, with an accuracy of 100 meters and with a temporal resolution of one hour. This model, based on the U-Net architecture, has proved to be very efficient in estimating the Tain a reasonable time, and with a much smaller amount of variables and data than the numerical models. The three RQs posed in the introduction of this chapter have been answered with satisfactory results. On the one hand, the model has shown that estimates the Tawith a proper spatial resolution and consistency. The results show an acceptable error and a good correspondence not only at the amplitude level, but also in capturing the intra-days patterns of the variable of interest. The temporal results have evinced that the propsed model excels at estimating the temperature variations occurring in a patch. It can be said, that the proposed model learns correctly to correlate the temperature with the input datasets, both spatial and meteorological. Finally, a new formula for detecting hotspots in cities has been proposed and has been proven to be very useful for detecting zones with higher heat stress within urban areas. The promising results reported in this chapter should not conceal the fact that several aspects of the proposed model have still room for improvement. To begin with, the U-Net model herein proposed has been proven to work under certain conditions. The model has demonstrated its validity just for one scenario but it cannot be ascertained that it is scalable to others cities. Also, it has been trained just for a few specific LWT days of summer. For these days U-Net has exposed a reliable and robust performance when eliciting spatially and temporally coherent estimations of Ta. However, for a whole year that robustness has yet to be verified. Additionally, the results obtained in comparison to the numerical model may suggest the possibility of entirely replacing the numerical model with a data-based model. However, numerical models cannot be substituted 4.4. Summary 93 at present, both for the reasons outlined earlier and because the model requires training on an extensive temperature database. At present, the only way to obtain such a large database is through numerical models, since no city has a wide enough sensor coverage to be able to estimate Ta with a high spatial resolution. However, the model can be useful for a what if scenario, since thanks to the computational speed, temperatures can be estimated in a few minutes after applying changes to the input data. For example, it could be used to see the effect on Tathat a specific urban action would cause, before actually implementing the urban intervention. Another option would be to use the tool to compare different urban modifications and their effects, thus allowing decision makers to choose the most appropiate intervention regarding their impact on the urban heat distribution of the city at hand. Besides the aforementioned improvements, the findings discussed here inspire various avenues of research to be explored in the upcoming future. The extension of the temporal range to other seasons should be done in order to see if the model maintains its robustness. Analyzing and broadening the scope of input datasets that are fed into the U-Net model could be another interesting step forward in the research. If the model is to be trained in different seasons apart from summer, NVDI should be considered for example, due to its seasonal variation that could affect its estimations. Another line of research could be to deploy the model in other cities of the same climatic zone to see if it maintains satisfactory results. If the robustness of the model is confirmed in different locations and seasons of the same climatic zone, the logical progression would be to expand the model to other climatic zones, eventually allowing for a total replacement of the numerical model for the zero-shot estimation of the temperature in new urban scenarios. 95 Chapter 5 Concluding Remarks and Future Research As it has been outlined throughout this Thesis, the future of cities relies in some key aspects like sustainability, resiliency and age-friendliness. Cities should be planned to be accessible, inclusive, adaptable to changes and forward-thinking. Moreover, they should also be prepared to respond quickly to both environmental and socio-economic changes leaving no strata of society behind, especially the most vulnerable groups. In short cities must have the capacity to adapt, according to the times, to the needs of their inhabitants. In this context, the role of decision-making processes is evolving to meet the rapid adaptability requirements of 21st century cities. As a result of the new digital technologies and a mass production of data, a shift towards data-driven, intelligent systems is occurring, enabling the analysis of complex urban issues and providing better insights for decision-making. Among these tools, AI has emerged as one of the central actors. As explained in this Thesis, AI’s ability to process vast amounts of data, and hidden or complex relationships from it, allows for more informed and effective decision-making processes. Also, AI can be linked with other innovative or visual technologies such as Digital Twins to enhance its capabilities, and to simulate and predict various scenarios, enabling city planners and decision-makers to test strategies and make better-informed decisions. Summing up, AI-based techniques can be of great use in improving decision-making processes, accomplishing the goal of sustainable and friendly cities easier to achieve. To this end, this Thesis has investigated the capabilities of several AI techniques to solve optimization and modeling problems that urban areas suffer from nowadays, and that, so far have been treated with classical methods. Specifically, two different problems related to urban environments have been tackled by using AI-based methods: 1) an optimization problem to improve age-friendliness of the cities and 2) a modeling problem to detect and estimate thermal stress in urban areas. In what follows we present the main contributions and findings of the Thesis with respect to these two use cases: 96 Chapter 5. Concluding Remarks and Future Research •Chapter 3.Improving Urban Accessibility and Environmental Conditions using Graph Modeling and Multi-objective Optimization. A novel framework for optimization of urban accessibility infrastructure installation taking into account orography and environmental conditions has been proposed, based on MOEAs and graph modeling. The fusion of georeferenced graph modeling describe the urban scenario under consideration and MOEAs has been shown to be useful to determine which urban actuations to perform in regards to accessibility, balancing the trade-off between cost and impact on the accessibility and walkability of pedestrians in their way to points of interest. The main conclusions drawn from the research presented in this chapter can be summarized as follows: •The framework is intuitive and based on common sense. It suggests asset installations in areas with significant accessibility, noise and air pollution problems, thus providing valuable support to urban planners in their decision-making processes. Moreover, the use of MOEAs in our framework permits not only to find optimal locations for urban accessibility based on their contribution to accessibility, but also account for the cost of the installation and the avoidance of noisy and polluted areas over the city. Urban planners often have to take decisions within the constraints of public and private budgets. The Thesis is one of its kind when considering costs of the assets and the installation process together with other factors related to the accessibility and walkability of pedestrians. •From the reported statistics on four quality indicators and subsequent significance analysis of observed differences, there are distinct performance disparities among various MOEAs when implemented within the framework. In our benchmark the chosen algorithm was the SPEA2. •The high-quality and relevance of data are crucial for the application of the proposed framework. The data available is not fine-grained and historical data is often not properly archived or publicly accessible. The availability of high precision data for air quality and noise could enhance the algorithm’s precision in the future. •Chapter 4.Efficient Estimation of Ground-Level Air Temperature in Urban Areas using Machine Learning. As discussed in the state-of-the-art of Chapter 4, the use of AI for ground-level air temperature modeling is still in its infancy, with very scarce contributions reported to date in this research area. The Thesis has built upon this motivating research niche to propose an encoder-decoder DNN architecture to estimate the spatial distribution of air temperature at ground level in urban environments. The proposed model is fed with information about land cover, imperviousness, terrain model, and meteorological variables from the use case area and directly outputs a map with the estimated temperature for the area under study. The goal has been twofold: 1) to yield an estimation of this variable in a more efficient fashion than numerical models used for the same purpose; and 2) to detect urban hotspots and support decision making 5.1. Research Outcomes 97 processes related to the mitigation of the UHI effect in cities. Our results discussed in this chapter have evinced that: •DL models, in particular the U-Net architecture, have proven to be a useful strategy for shaping ground-level air temperatures. The proposed model is capable of estimating Taboth spatially and temporally with a great similarity to numerical models. Additionally, it has obtained a successful performance in the detection of hotspots within urban areas. This detection of hotspots has been validated by expert knowledge, as the identified areas correspond to regions where specific elements or buildings exhibit thermal behaviors conducive to the formation of UHI. This alignment between detected hotspots and real-world conditions demonstrates the reliability of the detection method. •Thanks to its reduced computational time compared to large numerical models, the proposed model can be a core part of a new decision-making tool for urban planners. Indeed, the estimation of the air temperature enabled by the model can be used for the inspection of what-if scenarios, feeding customized spatial dataset that reflect possible actuations over the area under study. The output of the model corresponding to such customized inputs can help the decision maker ascertain how actuations would impact on the estimated air ground temperature. •The target variable modeled by the proposed neural architecture is the output of a numerical model. The use of CNN layers within the architecture allows for a spatial smoothing effect in the estimated air ground temperature maps, which in turn allows for a reduction of the spatial artifacts in the output produced by the numerical model. This byproduct yields an estimation that is closer to real measures of this variable than that of the numerical model. In other words, our proposed architecture provides air ground temperature estimates that are closer to real values collected by weather stations over the city. Nevertheless, while recognizing the existing potential, it remains to be tested in other scenarios, climates and seasons. In conclusion, this PhD has advanced the field of AI-driven modeling and optimization by demonstrating the practical value that these algorithms can bring in addressing critical urban challenges, particularly in the areas of accessibility and thermal comfort. By rigorously validating the effectiveness of AI tools in complex urban environments, the results reported in the Thesis provide solid evidence that AI can be a transformative asset in the decision-making processes of urban planners, enabling more informed, data-driven strategies for creating sustainable, accessible, and resilient cities. 5.1 Research Outcomes The research carried out in the course of this PhD Thesis led to two contributions to international conferences and two journal article, one published 98 Chapter 5. Concluding Remarks and Future Research in a JCR-indexed journal and the other one currently under review. In addition, the framework presented in Chapter 3 has been protected via a patent application. Details of these research outcomes are provided below: •Journal publications: – Iñigo Delgado-Enales, Patricia Molina-Costa, Javier Del Ser, “A framework to improve urban accessibility and environmental conditions in age-friendly cities using graph modeling and multiobjective optimization”, Computers Environment and Urban Systems, Volume 102, 101966, June 2023. JCR: 7.1 (Q1), 18/182, Environmental Studies. – Iñigo Delgado-Enales, Joshua Lizundia-Loiola, Patricia MolinaCosta, Javier Del Ser, “A Machine Learning Approach for the Efficient Estimation of Ground-Level Air Temperature in Urban Areas”, Urban Climate, JCR: 6.0 (Q1), 12/110, Meteorology & Atmospheric Sciences (under review). •Conference publications: – Iñigo Delgado-Enales, Patricia Molina-Costa, Eneko Osaba, Silvia Urra, Javier Del Ser, “Improving the Urban Accessibility of Older Pedestrians using Multi-objective Optimization”, IEEE Congress on Evolutionary Computation (CEC), Padua, Italy, pp. 1-8, 2022. – Iñigo Delgado-Enales, Patricia Molina-Costa, Javier Del Ser, “Spatial Estimation of Ground-Level Temperature for Climate-Sensitive Urban Mobility using Image-to-Image Deep Neural Networks”, IEEE International Conference on Intelligent Transportation Systems (ITSC), Bilbao, Spain, pp. 6206-6212, 2023. •Patent: –“System and Method for Determining Installation of Urban Infrastructures in Urban Environments”, European Patent Application No.: 22382426.9. Inventors: Iñigo Delgado Enales, Patricia Molina-Costa, Javier Del Ser. 5.2 Future Research Directions This Thesis concludes outlining the potential paths for further exploration based on the findings of the current research done in the PhD Thesis. For both the optimization and modeling domains promising research directions will be exposed. Urban graph datasets with increased node/link features. Starting from intelligent optimization of urban accessibility infrastructure, as highlighted before, enriching the georeferenced urban graph model with better and more precise datasets will significantly enhance the accuracy 5.2. Future Research Directions 99 and reliability of urban planning and development strategies. By incorporating high-resolution data such as the locations where urban assets have already been installed, including urban topology or taking into account the pivotal role that street canyons play in the noise and pollution dissipation, a more comprehensive and realistic representation of urban landscapes can be created. With this enriched urban graph model, MOEAs would rely on larger information frames, increasing their precision and facilitating a more informed decision-making process. For that, the rapid digitalization of urban areas is crucial, in order to have access to broader and richer data. Exploitation of urban digital twins. Another research direction worth to be explored in the future focuses on the advent of digital twinning, which has opened up new possibilities for urban planning. Urban digital twins enable new functionalities such as interacting with the virtual urban landscape in real time, manually adding or modifying assets and instantly observing the effects in the model [208]. For example, it is possible to simulate sunlight exposure using a dynamic 3D model [209]. The real-time visualization provides valuable information on the environmental impact of proposed changes to the urban planner. The framework proposed in Chapter 3 of this Thesis could be integrated into local digital twins for planning more accessible and age-friendly cities, like proposed by [210]. In that way, it would be possible, in pseudo-real time, to recalculate and visualize how the interventions proposed by the MOEA reshape the urban landscape. In the same way, another option could allow to change the mode of transport of the user, e.g. from walking to wheelchair, re-configuring all the scenario and ideally the MOEA offering other alternative interventions. In short, the idea would be that the changes could be visualized at the moment, offering a user level experience where the urban planner could evaluate different alternatives. To this end, however, a large improvement of the computational times and better computational resources would be needed. Scalability of the model. When it comes to modeling air temperatures in urban areas (Chapter 4), a promising research path can be to explore whether the estimation model learned from data corresponding to a given city can generalize well to predict the air ground temperature in other cities. This would involve investigating whether there is a correlation of Tabetween cities with similar urban morphology, climatic zones or other attributes. For instance, it would be interesting to prove if data from an specific type of city could be used to estimate Tafor another city that shares the same characteristics. This line of research could provide valuable insights into the applicability of urban planning models across different cities. Uncertainty estimation. Another future research line could be the estimation of uncertainty. This would involve developing a model that, in addition to predicting or estimating the temperature, also provides an 100 Chapter 5. Concluding Remarks and Future Research assessment of the confidence level of its predictions. The model would indicate areas where it is more certain of its results and areas where the uncertainty is higher, allowing for more informed decision-making processes. To this end, techniques for model confidence estimation such as Monte Carlo dropout [211], Bayesian networks [212] or evidential formulations [213] erof the regression loss used for training the encoder-decoder architecture can be considered. This Thesis has investigated the potential of various AI Techniques to support better-informed decision-making processes in the field of urban planning and management. Optimization offers great potential for the location of urban infrastructure of any kind considering multiple factors, including physical, environmental, social and economic ones. Further research could include housing development decisions, public facilities ensuring coverage for all city inhabitants, EV-chargers infrastructure deployment, among others. Modeling urban phenomena can help to predict future scenarios and making decisions accordingly. Apart from temperature, other environment conditions such as Air Quality can be modelled linked to factors such as traffic and weather to predict scenarios and deploy solutions and policies in advance (such as dynamic low emissions zones in city centres) [214], [215]. However, the potential of applying AI techniques to complex urban problems is dependant on the availability of relevant and high-quality data, and therefore it is crucial that city governments advance on their data strategies and governance to be able to grasp the full potential of AI as a decision-support system for enhanced urban planning and management. 101 Bibliography [1] S. V. Lall, H. Selod, and Z. Shalizi, Rural-urban migration in developing countries : A survey of theoretical predictions and empirical findings. The World Bank, 2006. [2] H. X. Wu and L. Zhou, “Rural-to-urban migration in china,” AsianPacific Economic Literature, vol. 10, no. 2, pp. 54–67, 1996. [3] U. N. D. of Economic and S. Affairs, World urbanization prospects: The 2018 revision. United Nations, 2018. [4] A. M. R. Nishimwe and S. Reiter, “Using artificial intelligence models and degree-days method to estimate the heat consumption evolution of a building stock until 2050: A case study in a temperate climate of the northern part of europe,” Cleaner and Responsible Consumption, vol. 5, p. 100069, 2022. [5] W. H. Organization, Global age-friendly cities: A guide. World Health Organization, 2007. [6] M. Alberti, J. M. Marzluff, E. Shulenberger, G. Bradley, C. Ryan, and C. Zumbrunnen, “Integrating humans into ecology: Opportunities and challenges for studying urban ecosystems,” BioScience, vol. 53, no. 12, pp. 1169–1179, 2003. [7] Organisation for economic co-operation and development (oecd),h ttps://www.oecd.org/cfe/regionaldevelopment/resilient-c ities.htm (Date accessed: 13.10.2024). [8] U. N. D. of Economic and S. Affairs, The sustainable development goals report 2023: Special edition. United Nations, 2023. [9] B. Purvis, Y. Mao, and D. Robinson, “Three pillars of sustainability: In search of conceptual origins,” Sustainability science, vol. 14, pp. 681–695, 2019. [10] J. Lewtas, “Air pollution combustion emissions: Characterization of causative agents and mechanisms associated with cancer, reproductive, and cardiovascular effects,” Mutation Research/Reviews in Mutation Research, vol. 636, no. 1-3, pp. 95–133, 2007. [11] N. Watts et al., “The 2020 report of the lancet countdown on health and climate change: Responding to converging crises,” The Lancet, vol. 397, no. 10269, pp. 129–170, 2021. [12] T. R. Oke, G. Mills, A. Christen, and J. A. Voogt, Urban Climates. Cambridge University Press, 2017. 108 Bibliography [89] A. Alkharabsheh, S. Moslem, L. Oubahman, and S. Duleba, “An integrated approach of multi-criteria decision-making and grey theory for evaluating urban public transportation systems,” Sustainability, vol. 13, no. 5, p. 2740, 2021. [90] R. Carli, M. Dotoli, and R. Pellegrino, “Multi-criteria decisionmaking for sustainable metropolitan cities assessment,” Journal of Environmental Management, vol. 226, pp. 46–61, 2018. [91] M. Sibilla and F. H. Abanda, “Multi-criteria decision making optimisation framework for positive energy blocks for cities,” Sustainability, vol. 14, no. 1, p. 446, 2022. [92] I. Ištoka Otković, B. Karleuša, A. Deluka-Tibljaš, S. Šurdonja, and M. Marušić, “Combining traffic microsimulation modeling and multicriteria analysis for sustainable spatial-traffic planning,” Land, vol. 10, no. 7, p. 666, 2021. [93] M. Cieśla, A. Sobota, and M. Jacyna, “Multi-criteria decision making process in metropolitan transport means selection based on the sharing mobility idea,” Sustainability, vol. 12, no. 17, p. 7231, 2020. [94] W. Jiang-Ping and T. Qun, “Urban planning decision using multiobjective optimization algorithm,” in ISECS International Colloquium on Computing, Communication, Control, and Management, IEEE, vol. 4, 2009, pp. 392–394. [95] A. Domenico, G. Nicola, T. Daniela, C. Fulvio, A. Nicola, and N. Orazio, “De novo drug design of targeted chemical libraries based on artificial intelligence and pair-based multiobjective optimization,” Journal of Chemical Information and Modeling, vol. 60, no. 10, pp. 4582–4593, 2020. [96] S.-Y. Shin, I.-H. Lee, D. Kim, and B.-T. Zhang, “Multiobjective evolutionary optimization of DNA sequences for reliable DNA computing,” Transactions on Evolutionary Computation, vol. 9, no. 2, pp. 143–158, 2005. [97] A. S. Yalcin, H. S. Kilic, and D. Delen, “The use of multi-criteria decision-making methods in business analytics: A comprehensive literature review,” Technological Forecasting and Social Change, vol. 174, p. 121193, 2022. [98] W. Zhang, C. Wang, Y. Li, P. Wang, Q. Wang, and D. Wang, “Seeking sustainability: Multiobjective evolutionary optimization for urban wastewater reuse in China,” Environmental Science & Technology, vol. 48, no. 2, pp. 1094–1102, 2014. [99] R. J. Balling, J. T. Taber, M. R. Brown, and K. Day, “Multiobjective urban planning using genetic algorithm,” Journal of Urban Planning and Development, vol. 125, no. 2, pp. 86–99, 1999. [100] Z. Masoumi, J. Maleki, M. S. Mesgari, and A. Mansourian, “Using an evolutionary algorithm in multiobjective geographic analysis for land use allocation and decision supporting,” Geographical Analysis, vol. 49, no. 1, pp. 58–83, 2017. Bibliography 109 [101] M. Péres, G. Ruiz, S. Nesmachnow, and A. C. Olivera, “Multiobjective evolutionary optimization of traffic flow and pollution in Montevideo, Uruguay,” Applied Soft Computing, vol. 70, pp. 472– 485, 2018. [102] A Magrini and A Lisot, “Noise reduction interventions in the urban environment as a form of control of indoor noise levels,” Energy Procedia, vol. 78, pp. 1653–1658, 2015. [103] M. R. Ismail, “Quiet environment: Acoustics of vertical green wall systems of the islamic urban form,” Frontiers of Architectural Research, vol. 2, no. 2, pp. 162–177, 2013. [104] A. W. A. Hammad, A. Akbarnezhad, and D. Rey, “Sustainable urban facility location: Minimising noise pollution and network congestion,” Transportation Research Part E: Logistics and Transportation Review, vol. 107, pp. 38–59, 2017. [105] X. Ning, J. Qi, C. Wu, and W. Wang, “Reducing noise pollution by planning construction site layout via a multi-objective optimization model,” Journal of Cleaner Production, vol. 222, pp. 218–230, 2019. [106] D. Mishra, P. Goyal, and A. Upadhyay, “Artificial intelligence based approach to forecast pm2.5 during haze episodes: A case study of Delhi, India,” Atmospheric Environment, vol. 102, pp. 239–248, 2015. [107] H. Niska, T. Hiltunen, A. Karppinen, J. Ruuskanen, and M. Kolehmainen, “Evolving the neural network model for forecasting air pollution time series,” Engineering Applications of Artificial Intelligence, vol. 17, no. 2, pp. 159–167, 2004, Intelligent Control and Signal Processing. [108] A. Masood and K. Ahmad, “A review on emerging artificial intelligence (ai) techniques for air pollution forecasting: Fundamentals, application and performance,” Journal of Cleaner Production, vol. 322, p. 129072, 2021. [109] J. Wang, L. Bai, S. Wang, and C. Wang, “Research and application of the hybrid forecasting model based on secondary denoising and multi-objective optimization for air pollution early warning system,” Journal of Cleaner Production, vol. 234, pp. 54–70, 2019. [110] S. Wicki, J. Schwaab, J. Perhac, and A. Grêt-Regamey, “Participatory multi-objective optimization for planning dense and green cities,” Journal of Environmental Planning and Management, vol. 64, no. 14, pp. 2532–2551, 2021. [111] G. D’Orso and M. Migliore, “A gis-based method for evaluating the walkability of a pedestrian environment and prioritised investments,” Journal of Transport Geography, vol. 82, p. 102555, 2020. [112] I. Blecic, A. Cecchini, and G. A. Trunfio, “Computer-aided drafting of urban designs for walkability,” in Computational Science and Its Applications, Springer International Publishing, 2017, pp. 695–709. 110 Bibliography [113] S. Salcedo-Sanz, L. Cuadra, E. Alexandre-Cortizo, S. Jiménez-Fernández, and A. Portilla-Figueras, “Soft-Computing: An innovative technological solution for urban traffic-related problems in modern cities,” Technological Forecasting and Social Change, vol. 89, pp. 236–244, 2014. [114] N. Bolten and A. Caspi, “Towards routine, city-scale accessibility metrics: Graph theoretic interpretations of pedestrian access using personalized pedestrian network analysis,” PLOS ONE, vol. 16, no. 3, pp. 1–20, 2021. [115] G. L. Barczyszyn, L. M. D. O. Camenar, D. D. F. D. Nascimento, et al., “A collaborative system for suitable wheelchair route planning,” Transactions on Accesible Computing, vol. 11, pp. 1–26, 2018. [116] M. S. Rahaman, Y. Mei, M. Hamilton, and F. D. Salim, “CAPRA: A contour-based accessible path routing algorithm,” Information Sciences, vol. 385-386, pp. 157–173, 2017. [117] W. Sasaki and Y. Takama, “Walking route recommender system considering saw criteria,” in Conference on Technologies and Applications of Artificial Intelligence, 2013, pp. 246–251. [118] W. Zhang, K. Cao, S. Liu, and B. Huang, “A multi-objective optimization approach for health-care facility location-allocation problems in highly developed cities such as Hong Kong,” Computers, Environment and Urban Systems, vol. 59, pp. 220–230, 2016. [119] I. Delgado-Enales, P. Molina-Costa, E. Osaba, S. Urra-Uriarte, and J. Del Ser, “Improving the urban accessibility of older pedestrians using multi-objective optimization,” in Congress on Evolutionary Computation, IEEE, 2022, pp. 1–8. [120] W. Tobler, “Three presentations on geographical analysis and modeling,” 1993. [121] C. A. C. Coello, G. B. Lamont, D. A. Van Veldhuizen, et al.,Evolutionary algorithms for solving multi-objective problems. Springer, 2007, vol. 5. [122] A. Zhou, B.-Y. Qu, H. Li, S.-Z. Zhao, P. N. Suganthan, and Q. Zhang, “Multiobjective evolutionary algorithms: A survey of the state of the art,” Swarm and evolutionary computation, vol. 1, no. 1, pp. 32–49, 2011. [123] A. J. Nebro, J. J. Durillo, F. Luna, B. Dorronsoro, and E. Alba, “Mocell: A cellular genetic algorithm for multiobjective optimization,” International Journal of Intelligent Systems, vol. 24, no. 7, pp. 726–746, 2009. [124] K. Deb, K. Sindhya, and T. Okabe, “Self-adaptive simulated binary crossover for real-parameter optimization,” in Genetic and Evolutionary Computation Conference, Association for Computing Machinery, 2007, pp. 1187–1194. [125] E. Dijkstra, “A note on two problems in connexion with graphs,” Numerische Mathematik, vol. 1, pp. 269–271, 1959. Bibliography 111 [126] A. de Barcelona, Diagnosis para la estrategia sobre cambio demográfico y envejecimiento: Una ciudad para todos los ciclos de vida (2018-2030), 2018. [127] Organismo Autónomo Centro Nacional de Información Geográfica (CNIG), “Centro de descargas,” in https://centrodedescargas .cnig.es/CentroDescargas/ (Date accessed: 13.10.2024). [128] G. Boeing, “OSMnx: New methods for acquiring, constructing, analyzing, and visualizing complex street networks,” Computers, Environment and Urban Systems, vol. 65, pp. 126–139, 2017. [129] Open street map,https: / /www. openstreetmap .org (Date accessed: 13.10.2024). [130] A. Benítez-Hidalgo, A. J. Nebro, J. García-Nieto, I. Oregi, and J. Del Ser, “jMetalPy: A Python framework for multi-objective optimization with metaheuristics,” Swarm and Evolutionary Computation, vol. 51, p. 100598, 2019. [131] H. Ishibuchi, H. Masuda, Y. Tanigaki, and Y. Nojima, “Modified distance calculation in generational distance and inverted generational distance,” in International Conference on Evolutionary Multi-Criterion Optimization, Springer, 2015, pp. 110–125. [132] H. Ishibuchi, H. Masuda, Y. Tanigaki, and Y. Nojima, “Modified distance calculation in generational distance and inverted generational distance,” in International Conference on Evolutionary Multi-Criterion Optimization, Springer, 2015, pp. 110–125. [133] E. Zitzler and L. Thiele, “Multiobjective optimization using evolutionary algorithms – a comparative case study,” in Parallel Problem Solving from Nature, Springer, 1998, pp. 292–301. [134] E. Zitzler, L. Thiele, M. Laumanns, C. Fonseca, and V. da Fonseca, “Performance assessment of multiobjective optimizers: An analysis and review,” Transactions on Evolutionary Computation, vol. 7, no. 2, pp. 117–132, 2003. [135] Y. Liu, J. Wei, X. Li, and M. Li, “Generational distance indicatorbased evolutionary algorithm with an improved niching method for many-objective optimization problems,” IEEE Access, vol. 7, pp. 63881–63891, 2019. [136] C. M. Fonseca, J. D. Knowles, L. Thiele, E. Zitzler, et al., “A tutorial on the performance assessment of stochastic multiobjective optimizers,” in Third international conference on evolutionary multicriterion optimization, vol. 216, 2005, p. 240. [137] N. Riquelme, C. Von Lücken, and B. Baran, “Performance metrics in multi-objective optimization,” in Latin American computing conference, IEEE, 2015, pp. 1–11. [138] J. F. Mitchell, “The “greenhouse” effect and climate change,” Reviews of Geophysics, vol. 27, no. 1, pp. 115–139, 1989. 112 Bibliography [139] H. Rodhe, “A comparison of the contribution of various gases to the greenhouse effect,” Science, vol. 248, no. 4960, pp. 1217–1219, 1990. [140] M. Székely, L. Carletto, and A. Garami, “The pathophysiology of heat exposure,” Temperature, vol. 2, no. 4, pp. 452–452, 2015. [141] Z. Xu, G. Fitzgerald, Y. Guo, B. Jalaludin, and S. Tong, “Impact of heatwave on mortality under different heatwave definitions: A systematic review and meta-analysis,” Environment International, vol. 89-90, pp. 193–203, 2016. [142] H. H. Kim, “Urban heat island,” International Journal of Remote Sensing, vol. 13, no. 12, pp. 2319–2336, 1992. [143] K. De Ridder, D. Lauwaet, and B. Maiheu, “UrbClim–a fast urban boundary layer climate model,” Urban Climate, vol. 12, pp. 21–48, 2015. [144] T. R. Oke, “The energetic basis of the urban heat island,” Quarterly Journal of the Royal Meteorological Society, vol. 108, no. 455, pp. 1– 24, 1982. [145] M. Mokarram, F. Taripanah, and T. M. Pham, “Investigating the effect of surface urban heat island on the trend of temperature changes,” Advances in Space Research, vol. 72, no. 8, pp. 3150– 3169, 2023. [146] T. Adulkongkaew, T. Satapanajaru, S. Charoenhirunyingyos, and W. Singhirunnusorn, “Effect of land cover composition and building configuration on land surface temperature in an urban-sprawl city, case study in bangkok metropolitan area, thailand,” Heliyon, vol. 6, no. 8, p. 4485, 2020. [147] A. Karimi, P. Mohammad, A. García-Martínez, D. Moreno-Rangel, D. Gachkar, and S. Gachkar, “New developments and future challenges in reducing and controlling heat island effect in urban areas,” Environment, Development and Sustainability, vol. 25, no. 10, pp. 10485–10531, 2023. [148] G.-J. Steeneveld, S. Koopmans, B. Heusinkveld, L. Van Hove, and A. Holtslag, “Quantifying urban heat island effects and human comfort for cities of variable size and urban morphology in the netherlands,” Journal of Geophysical Research: Atmospheres, vol. 116, no. D20, 2011. [149] X. Zhou and H. Chen, “Impact of urbanization-related land use land cover changes and urban morphology changes on the urban heat island phenomenon,” Science of the Total Environment, vol. 635, pp. 1467–1476, 2018. [150] H. Zhang, J. Xu, and J. Wang, “Pretraining-based natural language generation for text summarization,” in Proceedings of the Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, 2019, pp. 789–797. Bibliography 113 [151] L. Lebanoff, K. Song, and F. Liu, “Adapting the neural encoderdecoder framework from single to multi-document summarization,” in Proceedings of the Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, 2018, pp. 4131–4141. [152] Y. Liu and M. Lapata, “Text summarization with pretrained encoders,” in Proceedings of the Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, 2019, pp. 3730–3740. [153] K. Cho, B. Van Merriënboer, C. Gulcehre, et al., “Learning phrase representations using rnn encoder-decoder for statistical machine translation,” in Proceedings of the Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, 2014, pp. 1724–1734. [154] K. Cho, B. Van Merriënboer, D. Bahdanau, and Y. Bengio, “On the properties of neural machine translation: Encoder-decoder approaches,” in Proceedings of SSST, Workshop on Syntax, Semantics and Structure in Statistical Translation, Association for Computational Linguistics, 2014, pp. 103–111. [155] J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic segmentation,” in Conference on Computer Vision and Pattern Recognition, IEEE, 2015, pp. 3431–3440. [156] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille, Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs, 2017. [157] H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia, Pyramid scene parsing network, 2017. arXiv: 1612.01105. [158] R. Mehta and T. Arbel, “3d u-net for brain tumour segmentation,” in International MICCAI Brainlesion Workshop, Springer, 2018, pp. 254–266. [159] W. Chen, B. Liu, S. Peng, J. Sun, and X. Qiao, “S3d-unet: Separable 3d u-net for brain tumor segmentation,” in Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries, Springer International Publishing, 2019, pp. 358–368. [160] A. Saha, Y.-D. Zhang, and S. C. Satapathy, “Brain tumour segmentation with a muti-pathway resnet based unet,” Journal of Grid Computing, vol. 19, pp. 1–10, 2021. [161] Y. Cao, S. Liu, Y. Peng, and J. Li, “Denseunet: Densely connected unet for electron microscopy image segmentation,” IET Image Processing, vol. 14, no. 12, pp. 2682–2689, 2020. [162] H. Cao, Y. Wang, J. Chen, et al., “Swin-unet: Unet-like pure transformer for medical image segmentation,” in European conference on computer vision, Springer, 2022, pp. 205–218. 114 Bibliography [163] X. Yan, H. Tang, S. Sun, H. Ma, D. Kong, and X. Xie, “After-unet: Axial fusion transformer unet for medical image segmentation,” in Proceedings of the IEEE/CVF winter conference on applications of computer vision, 2022, pp. 3971–3981. [164] R. Augustaukas and A. Lipnickas, “Pixel-wise road pavement defects detection using u-net deep neural network,” in International Conference on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications, vol. 1, IEEE, 2019, pp. 468–471. [165] J. Jing, Z. Wang, M. Rätsch, and H. Zhang, “Mobile-unet: An efficient convolutional neural network for fabric defect detection,” Textile Research Journal, vol. 92, no. 1-2, pp. 30–42, 2022. [166] W. Yao, Z. Zeng, C. Lian, and H. Tang, “Pixel-wise regression using u-net and its application on pansharpening,” Neurocomputing, vol. 312, pp. 364–371, 2018. [167] A. Soni, R. Koner, and V. G. K. Villuri, “M-unet: Modified u-net segmentation framework with satellite imagery,” in Proceedings of the Global AI congress 2019, Springer, 2020, pp. 47–59. [168] Z. Wang, P. Yang, H. Liang, et al., “Semantic segmentation and analysis on sensitive parameters of forest fire smoke using smokeunet and landsat-8 imagery,” Remote Sensing, vol. 14, no. 1, 2022. [169] J. McGlinchy, B. Johnson, B. Muller, M. Joseph, and J. Diaz, “Application of unet fully convolutional neural network to impervious surface segmentation in urban environment from high resolution satellite imagery,” in International Geoscience and Remote Sensing Symposium, IEEE, 2019, pp. 3915–3918. [170] M. M. Awad and M. Lauteri, “Self-organizing deep learning (sounet)— a novel framework to classify urban and peri-urban forests,” Sustainability, vol. 13, no. 10, p. 5548, 2021. [171] L. Wang, R. Li, C. Zhang, et al., “Unetformer: A unet-like transformer for efficient semantic segmentation of remote sensing urban scene imagery,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 190, 196–214, 2022. [172] P. Su, T. Abera, Y. Guan, and P. Pellikka, “Image-to-image training for spatially seamless air temperature estimation with satellite images and station data,” Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 16, pp. 3353–3363, 2023. [173] A. C. Lorenc, “Analysis methods for numerical weather prediction,” Quarterly Journal of the Royal Meteorological Society, vol. 112, no. 474, pp. 1177–1194, 1986. [174] P. P. Saha, K. Zeleke, and M. Hafeez, “Impacts of land use and climate change on streamflow and water balance of two sub-catchments of the Murrumbidgee River in South Eastern Australia,” in Extreme Hydrology and Climate Variability, 2019, pp. 175–190. Bibliography 115 [175] G. Flato et al., “Evaluation of climate models,” in Climate change 2013: the physical science basis. Contribution of Working Group I to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change, Cambridge University Press, 2014, pp. 741–866. [176] M. Xue, K. K. Droegemeier, and V. Wong, “The advanced regional prediction system (ARPS)–a multi-scale nonhydrostatic atmospheric simulation and prediction model. part i: Model dynamics and verification,” Meteorology and Atmospheric Physics, vol. 75, pp. 161–193, 2000. [177] A. Sarkar and K. De Ridder, “The urban heat island intensity of Paris: A case study based on a simple urban surface parametrization,” Boundary-layer meteorology, vol. 138, pp. 511–520, 2011. [178] P. Berg, O. Christensen, K. Klehmet, et al., “Precipitation extremes in a euro-cordex 0.11°ensemble at hourly resolution,” Natural Hazards and Earth System Sciences Discussions, pp. 1–21, Dec. 2018. [179] M. J. Lipson et al., “Evaluation of 30 urban land surface models in the urban-PLUMBER project: Phase 1 results,” Quarterly Journal of the Royal Meteorological Society, vol. 150, no. 758, pp. 126–169, 2024. [180] D. Lauwaet et al., “Assessing the current and future urban heat island of brussels,” Urban Climate, vol. 15, pp. 1–15, 2016. [181] N. Souverijns et al., “Urban heat in Johannesburg and Ekurhuleni, South Africa: A meter-scale assessment and vulnerability analysis,” Urban Climate, vol. 46, p. 101331, 2022. [182] C. Reis, A. Lopes, and A. S. Nouri, “Assessing urban heat island effects through local weather types in Lisbon’s metropolitan area using big data from the Copernicus service,” Urban Climate, vol. 43, p. 101168, 2022. [183] D. Hidalgo-García and H. Rezapouraghdam, “Variability of heat stress using the urbclim climate model in the city of Seville (Spain): Mitigation proposal,” Environmental Monitoring and Assessment, vol. 195, no. 10, p. 1164, 2023. [184] D. Hidalgo García and J. Arco Díaz, “Mitigation and resilience of local climatic zones to the effects of extreme heat: Study on the city of Barcelona (Spain),” Urban Science, vol. 7, no. 4, p. 102, 2023. [185] E. J. I. Fernandez, M. A. R. Díaz, and A. P. Morales, “Efectos del sellado del suelo en la temperatura del aire en la ciudad de murcia,” in Geografía: cambios, retos y adaptación: libro de actas. XVIII Congreso de la Asociación Española de Geografía, Logroño, Asociación Española de Geografía, 2023, pp. 375–384. [186] M. Sharifi, M. H. Shamsi, Y. Ma, and D. Lauwaet, “Quantification of the impact of global warming on summer overheating risk in a residential building in urban areas in belgium,” in Journal of Physics: Conference Series, IOP Publishing, vol. 2600, 2023, p. 092016. 116 Bibliography [187] S. Vulova, F. Meier, D. Fenner, H. Nouri, and B. Kleinschmit, “Summer nights in berlin, germany: Modeling air temperature spatially with remote sensing, crowdsourced weather data, and machine learning,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 13, pp. 5074–5087, 2020. [188] M. Varentsov, M. Krinitskiy, and V. Stepanenko, “Machine learning for simulation of urban heat island dynamics based on large-scale meteorological conditions,” Climate, vol. 11, no. 10, 2023. [189] M. Furuya, D. Furuya, L. Y. Oliveira, et al., “A machine learning approach for mapping surface urban heat island using environmental and socioeconomic variables: A case study in a medium-sized brazilian city,” Environmental Earth Sciences, vol. 82, p. 325, 2023. [190] D. Erdemir and T. Ayata, “Prediction of temperature decreasing on a green roof by using artificial neural network,” Applied Thermal Engineering, vol. 112, pp. 1317–1325, 2017. [191] A. Asadi, H. Arefi, and H. Fathipoor, “Simulation of green roofs and their potential mitigating effects on the urban heat island using an artificial neural network: A case study in austin, texas,” Advances in Space Research, vol. 66, pp. 1846–1862, 2020. [192] S. Chan and C. Chau, “Development of artificial neural network models for predicting thermal comfort evaluation in urban parks in summer and winter,” Building and Environment, vol. 164, p. 106364, 2019. [193] F. Tomatis, F. J. Diez, M. S. Wilhelm, and L. M. Navas-Gracia, “Prediction of daily ambient temperature and its hourly estimation using artificial neural networks in urban allotment gardens and an urban park in Valladolid, Castilla y Leon, Spain,” Agronomy, vol. 14, no. 1, 2024. [194] A. Ashtiani, P. Mirzaei, and F. Haghighat, “Indoor thermal condition in urban heat island: Comparison of the artificial neural network and regression methods prediction,” Energy and Buildings, vol. 76, pp. 597–604, Jun. 2014. [195] J. W. Oh, J. Ngarambe, P. N. Duhirwe, G. Y. Yun, and M. Santamouris, “Using deep-learning to forecast the magnitude and characteristics of urban heat island in Seoul Korea,” Scientific Reports, vol. 10, no. 1, pp. 1–13, 2020. [196] K. Gobakis, D. Kolokotsa, A. Synnefa, M. Saliari, K. Giannopoulou, and M. Santamouris, “Development of a model for urban heat island prediction using neural network techniques,” Sustainable Cities and Society, vol. 1, no. 2, pp. 104–115, 2011. [197] G. Assaf, X. Hu, and R. H. Assaad, “Predicting urban heat island severity on the census-tract level using bayesian networks,” Sustainable Cities and Society, vol. 97, p. 104756, 2023. Bibliography 117 [198] S. Liu, J. Zhang, J. Li, Y. Li, J. Zhang, and X. Wu, “Simulating and mitigating extreme urban heat island effects in a factory area based on machine learning,” Building and Environment, vol. 202, p. 108051, 2021. [199] F. Briegel, O. Makansi, T. Brox, A. Matzarakis, and A. Christen, “Modelling long-term thermal comfort conditions in urban environments using a deep convolutional encoder-decoder as a computational shortcut,” Urban Climate, vol. 47, p. 101359, 2023. [200] H. E. Beck, N. E. Zimmermann, T. R. McVicar, N. Vergopolan, A. Berg, and E. F. Wood, “Present and future köppen-geiger climate classification maps at 1-km resolution,” Scientific data, vol. 5, no. 1, pp. 1–12, 2018. [201] J. A. Acero and K. Herranz-Pascual, “A comparison of thermal comfort conditions in four urban spaces by means of measurements and modelling techniques,” Building and Environment, vol. 93, pp. 245– 257, 2015. [202] H. Hooyberghs, J. Berckmans, D. Lauwaet, F. Lefebre, and K. De Ridder, “Climate variables for cities in europe from 2008 to 2017,” Copernicus Climate Change Service (C3S) Climate Data Store (CDS), 2019. [203] H. Hersbach et al., “The ERA5 global reanalysis,” Quarterly Journal of the Royal Meteorological Society, vol. 146, no. 730, pp. 1999– 2049, 2020. [204] J. Hidalgo and R. Jougla, “On the use of local weather types classification to improve climate understanding: An application on the urban climate of Toulouse,” Plos One, vol. 13, no. 12, pp. 1–21, Dec. 2018. [205] Euskalmet,https://www.euskalmet.euskadi.eus/divulgacion /manual-de-estilo/ (Date accessed: 13.10.2024). [206] European centre for medium-range weather forecasts,https://co nfluence.ecmwf.int/display/CKB/Climate+variables+for+c ities+in+Europe+from+2008+to+2017+documentation (Date accessed: 13.10.2024). [207] European Environment Agency, DG Joint Research Centre of the European Commission, “Copernicus land monitoring service,” in https://land.copernicus.eu/en. [208] F. Dembski, U. Wössner, M. Letzgus, M. Ruddat, and C. Yamu, “Urban digital twins for smart cities and citizens: The case study of herrenberg, germany,” Sustainability, vol. 12, no. 6, 2020. [209] G. Schrotter and C. Hürzeler, “The digital twin of the city of zurich for urban planning,” PFG–Journal of Photogrammetry, Remote Sensing and Geoinformation Science, vol. 88, no. 1, pp. 99– 112, 2020.