PRISMA on Machine Learning Techniques in Smart City Development

Abstract

This article investigates the innovative role of machine learning (ML) in the development of smart cities, emphasizing the critical interrelationship between ML and urban environments. While existing studies address ML and urban settings separately, this work uniquely examines their intersection, highlighting the transformative potential of ML in urban development. Utilizing the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) methodology, a systematic and reproducible approach was employed to review 42 relevant studies. The analysis reveals four key themes: transportation and traffic optimization, people and event flow tracking, sustainability applications, and security use cases. These findings underscore ML’s ability to revolutionize smart city initiatives by enhancing efficiency, sustainability, and security. This review identifies significant research gaps and proposes future directions, positioning ML as a cornerstone in the evolution of intelligent urban environments.

1. Introduction

The concept of a “Smart City” originated from the need to address common urban challenges by improving the quality of life in current settlements. This is achieved through the strategic utilization of both hardware and software [1] to collect data and gain insights to manage assets, resources, and services efficiently. We include here data collected, processed, and analyzed from citizens, devices, buildings, and assets. Smart cities manage essential service sectors, such as energy, traffic and transportation, water, waste, information systems, schools, libraries, hospitals, and public safety, ensuring their proper operation while maintaining a clean, cost-effective, and secure environment for comfortable living by residents [2,3,4,5].

The emergence of smart cities represents a significant step toward addressing the multifaceted challenges posed by urbanization, including congestion, pollution, resource management, and infrastructure strain. The strategic deployment of machine learning (ML) within these urban environments offers transformative potential, enhancing operational efficiency and sustainability across various sectors [6,7,8].

The application of machine learning techniques in smart cities is key to improving the efficiency of transportation systems, covering a wide range of applications, from traffic prediction to traffic congestion management and route optimization. Researchers have delved into areas like traffic pattern prediction, where ML algorithms, including neural networks and deep learning models, have shown success. These algorithms effectively foresee traffic patterns and recommend dynamic route adjustments to alleviate congestion [9].

The application of machine learning techniques also allows smart cities to consume energy more efficiently and manage it more effectively. Predictive models based on machine learning are used to optimize energy distribution, demand forecasting, and resource allocation. Machine learning algorithms are used to adapt to changing energy demands as they arise in energy-efficient buildings and smart grids [10,11].

Through predictive policing, video analytics, and crime pattern recognition, the literature shows that machine learning is emphasized as an aid to improving public safety. Law enforcement agencies can allocate resources effectively based on ML models, which analyze historical crime data to predict hotspots [12,13].

Due to the detrimental effects of air pollution, it is vital to have a reliable model for predicting Air Quality Index (AQI) levels to improve urban public health and foster sustained social progress. In smart urban settings, machine learning applications in health encompass the prognoses of illnesses, monitoring patients, and optimizing healthcare resources. Predictive analytics and machine learning frameworks support prompt identification of disease outbreaks, customized treatment strategies, and effective delivery of healthcare services [14,15].

Furthermore, the integration of the Internet of Things (IoT) data with advanced ML algorithms is being explored to optimize various urban functions, such as water and energy management. These algorithms are designed to adjust consumption patterns dynamically, considering user convenience and sustainability goals. Given the diverse parameters and variables inherent in each system, the primary challenges lie in achieving scalability and adaptability to diverse urban parameters, which necessitates precise control and simulation methodologies [16,17]. Ultimately, the system analysis and design will incorporate more realistic user consumption profiles.

By developing smart city solutions, machine learning techniques have demonstrated substantial potential to address the challenges modern cities face. Researchers and practitioners will continue to explore innovative applications of machine learning in diverse fields as technology continues to advance, contributing to the development of smarter, more efficient, and sustainable urban environments.

2. Methodology

To align with the research questions and identify key search terms, this study employed the PRISMA 2020 flow diagram methodology. An acronym for Preferred Reporting Items of Systematic Reviews and Meta-Analyses, PRISMA is a set of guidelines for conducting systematic literature reviews and meta-analyses. Research articles conducting literature reviews and meta-analyses in the past did not follow stringent scientific reporting guidelines, resulting in inaccurate evaluations. The Quality of Reporting of Meta-analyses (QUOROM) was formed in 1999 to address this concern, focusing on the reporting of meta-analyses specifically conducted in randomized controlled trials [18]. The statement was updated in 2009 and renamed PRISMA to reflect advancements in systematic review methods [19]. When examining an extensive publication collection, including peer-reviewed academic articles, this approach ensures a transparent, thorough, and precise presentation [20]. This study adhered to PRISMA 2020 guidelines to ensure accurate and comprehensive reporting. Consequently, the aim is to furnish readers with a comprehensive, reliable, and lucid comprehension of both the research techniques and the subject matter.

The methodology employed in this study follows the PRISMA guidelines, which are widely recognized for ensuring a systematic and transparent approach to reviewing literature. The PRISMA methodology includes the following key steps:

−

Database Search: The research was conducted across multiple academic databases.

−

Inclusion and Exclusion Criteria (articles were selected according to their relevance of the intersection between machine learning and smart cities).

−

Screening Process (automated removal of duplicates and irrelevant entries, manual screening).

−

Final Selection: A detailed examination of the remaining articles, applying stringent exclusion criteria.

As we conducted the final stages of the research, where exclusion was of the utmost importance, we consistently addressed these questions: “Is the study relevant to the paper’s purpose?” and “Is the study acceptable for the paper?” [21]. These questions were addressed first to the criteria the authors developed in the final stages of the research to identify the main exclusion criteria. Subsequently, we applied the same queries to the paper under review, integrating the exclusion criteria that we had established. Additionally, we focused on the main aspects that should be present in a criterion when validating the exclusion criteria [21]: “(a) study population, (b) nature of the intervention, (c) outcome variables, (d) time period, (e) cultural and linguistic range, and (f) methodological quality” [21]. To ensure that a good number of papers are included in Section 3, papers that did not meet the exclusion criteria were removed. This process was supplemented by principles from the Cochrane Handbook for Systematic Reviews of Interventions, which similarly emphasizes rigorous criteria for inclusion and exclusion, ensuring the relevance and quality of studies selected for review. The integration of both PRISMA and Cochrane methodologies underscores the importance of a meticulous and replicable approach to systematic reviews [22].

To have a comprehensive and integrative literature review on machine learning techniques in smart city development, in January 2024, the research was conducted on academic articles published in Scopus, SSRN, Mendeley, Emerald, Springer LINK, Science Direct, ERIC, MDPI, JSTOR, IEEE Xplore Digital Library, ACM Digital Library, and Google Scholar (first 10 pages). These databases are the most recognized and widely used in academia, containing extensive information on the topic.

Articles chosen for review were required to mention “machine learning” and “smart cities” in their title, abstract, or research focus. Excluded from the study were books, conference papers, commercial reports, opinion pieces, and editorials. We further excluded articles that did not entail “machine learning” or “smart cities” as part of their research or context. To establish validity of the inclusion and exclusion criteria, Robey & Dalebout used relevance and acceptability as criteria for establishing those aspects [23].

In accordance with PRISMA guidelines, we employed specific phrases to locate and sift through academic articles in the databases we used: (1) “machine learning”, (2) “smart cities”, (3) “Machine Learning Techniques in Smart Cities Development”, and (4) “Machine Learning in Smart Cities Development”. These keywords were selected to conduct a comprehensive yet targeted search, effectively capturing pertinent studies on the integration of machine learning within the framework of smart city development. We augmented these queries with synonyms related to each concept, conducting targeted searches within these databases to encompass any relevant papers. We also repeated these search terms in the reference search function to identify papers potentially off-topic. The titles and abstracts of the articles were manually scrutinized to select appropriate papers based on our inclusion and exclusion criteria, culminating in a comprehensive reading of the remaining articles to evaluate their suitability. The PRISMA methodology is depicted in Figure 1, with the detailed procedure outlined as follows.

The initial database search yielded 4808 article titles, which were imported into Zotero 7 software. Through this process, we removed 456 duplicates, 334 articles flagged as irrelevant by automated tools, and 514 articles lacking the keywords “machine learning”, “smart cities”, “Machine Learning Techniques in Smart Cities Development”, or “Machine Learning in Smart Cities Development” in their titles (identified using Zotero’s search function). This resulted in a pool of 3504 articles. A list of these articles was compiled in MS Excel for manual screening. We assessed these titles against our research questions and criteria, discarding those not relevant to our research topic. This process left us with 401 papers, but only 241 were accessible in full text (PDF). To summarize these 241 articles, we employed two AI tools, Wordtune Read and ChatGPT, to assist in summarizing the articles for a more efficient review process. Wordtune Read condensed the articles into approximately 2000-word versions, and ChatGPT further distilled these summaries into single-page overviews, using the command ‘create a one-page summary of this text’:

This approach was carefully chosen to maintain consistency and comprehensiveness in the review. While AI tools can introduce biases, we mitigated this risk by cross-verifying AI-generated summaries with the original texts and ensuring that the core content and context were preserved. This methodology allowed us to efficiently handle a large volume of literature without compromising on the depth and accuracy of the review, enabling us to swiftly go through the 241 papers, pinpointing those pertinent to our study by examining their overall content. The Artificial Intelligence (AI)-generated summaries were more comprehensive than the abstracts, providing a more in-depth examination of the articles’ contents. Subsequent to reviewing these summaries and applying our exclusion criteria to further narrow down the selection, we pinpointed 42 papers (originating from the database searches) for full analysis. These papers are catalogued in

Table 1. List of selected articles.

1. Abdul, M. et al. (2023) [25]	22. Navarro-Espinoza, A. et al. (2022) [26]
2. Al-Eidi, S. et al. (2023) [4]	23. Ngabo, D. et al. (2021) [27]
3. Ali, H.M. (2021) [28]	24. Niyogisubizo, J. et al. (2023) [29]
4. Ashwini, B. et al. (2022) [30]	25. Prakash, J. et al. (2024) [31]
5. Dorosan, M. et al. (2024) [32]	26. Quasim, M.T. et al. (2023) [33]
6. Founoun, A. et al. (2022) [34]	27. Rashid, M.M. et al. (2020) [35]
7. Furuya, M.T.G. et al. (2023) [36]	28. Roslan, R. et al. (2023) [37]
8. Gillani, M. et al. (2023) [38]	29. Saleem, M. et al. (2022) [39]
9. Hodorog, A. et al. (2022) [40]	30. Sanghami, S.V. et al. (2022) [41]
10. Hurbean, L. et al. (2021) [42]	31. Sharma, Amit et al. (2023) [43]
11. Iskandaryan, D. et al. (2020) [44]	32. Shi, Y. et al. (2023) [45]
12. Islam, N. et al. (2023) [46]	33. Shouaib, M. et al. (2023) [47]
13. Javaid, N. et al. (2022) [48]	34. Shulajkovska, M. et al. (2023) [49]
14. Kah Phooi Seng et al. (2023) [50]	35. Śmiałkowski, T. et al. (2022) [51]
15. Khan, M.A. et al. (2022) [52]	36. Soumana, A.N.H. et al. (2024) [53]
16. Kokane, C.D. et al. (2023) [54]	37. Sousa, B. et al. (2023) [55]
17. Liu, F. et al. (2023) [56]	38. Tao, X. et al. (2024) [5]
18. Lu, H. et al. (2023) [57]	39. Ullah, A. et al. (2023) [58]
19. Maktoof, M.A. et al. (2023) [59]	40. Wai, K.-M., Yu, P.K. (2023) [60]
20. Malik, M. et al. (2023) [61]	41. Yao, B. et al. (2023) [62]
21. Mishra, K.N. et al. (2022) [63]	42. Zekić-Sušac, M. et al. (2021) [64]

.

The findings from this conclusive analysis will be detailed in Section 3 of the paper. As indicated in Figure 1, the 199 articles that were omitted were excluded for the following reasons: (1) Articles were excluded if they focused exclusively on machine learning applications or techniques without contextual relevance to smart cities, resulting in the exclusion of 59 papers. (2) An additional 50 articles were excluded because they primarily centered on smart cities without providing substantial emphasis on the application or development of machine learning techniques. (3) Another 45 articles were excluded for not being data-driven or failing to showcase the integration of machine learning in smart cities. (4) Last, 45 papers were excluded because, although they mentioned machine learning in smart city development, the emphasis was more on other concepts like IoT, AI, or Deep Learning rather than machine learning as the main focus in the development strategy. This criterion ensured that studies lacking substantial focus on machine learning techniques were excluded, thereby maintaining the review’s emphasis on machine learning’s impact on smart city development. Our filtering criteria included relevance to the application of machine learning in smart city contexts, publication within the last ten years, peer-reviewed sources, and significant emphasis on the application or development of machine learning techniques.

3. Results

In the dynamic landscape of urban development, the emergence of smart cities has become a solution to address the challenges posed by rapid urbanization, environmental sustainability, and resource management. A comprehensive review utilizing the PRISMA integrative literature approach sheds light on the interplay between machine learning techniques and various aspects of smart city infrastructure optimization. Through the analysis of the 42 studies that resulted at the end of the PRISMA chart, four major themes have emerged, each contributing to the overarching goal of creating eco-friendly and efficient urban environments.

(1) Transport Systems and Traffic Flow (13 papers): Efficient transport systems and traffic flow are crucial for alleviating traffic congestion, improving air quality, and boosting the overall quality of life in urban areas. By optimizing transportation, cities can minimize travel times, reduce fuel consumption, and decrease greenhouse gas emissions. Numerous studies highlighted the application of machine learning in predicting traffic patterns, managing transportation networks, and improving road safety. These studies demonstrate the potential of smart technologies to transform urban mobility.

(2) Flow Tracking System (13 papers): Flow tracking systems are essential for monitoring and managing the movement of people and goods within a city. These systems help in ensuring the efficient use of infrastructure and resource allocation, enhancing public safety, and supporting effective urban planning. This theme was identified due to the significant focus on the use of machine learning to analyze and predict human flow patterns, event dynamics, and other movement-related aspects in urban environments. These insights are vital for optimizing city operations and improving the quality of urban life.

(3) Green Innovation (Energy) in Smart Cities (12 papers): Sustainable energy solutions are a cornerstone of smart city initiatives. Green innovation in energy management helps in reducing the environmental footprint, promoting energy efficiency, and supporting the transition to renewable energy sources. The PRISMA analysis revealed this theme as a major area of research due to the numerous studies on smart street lighting, energy waste management, and air and water quality monitoring. These studies underscore the role of machine learning in advancing sustainability and environmental stewardship in urban settings.

(4) Security/Cybersecurity (4 papers): Ensuring the security of urban infrastructures and protecting them from cyber threats is fundamental for the resilience and reliability of smart cities. Robust security measures are needed to safeguard sensitive data, maintain public safety, and prevent disruptions. This theme emerged because of the growing concern over cybersecurity threats in increasingly connected urban environments. The literature highlighted the application of machine learning in detecting and mitigating security breaches, thus ensuring the safe operation of smart city systems.

These four themes emerged from the PRISMA analysis as they represent the most critical and frequently addressed areas in the literature on machine learning applications in smart cities. The comprehensive review highlighted these themes as pivotal to achieving the overarching goals of efficiency, sustainability, and resilience in urban environments. Each theme addresses specific challenges faced by cities and showcases the transformative potential of machine learning technologies in creating smarter, more sustainable urban spaces.

3.1. Transport Systems & Traffic Flow

The first identified theme revolves around the transformative role of smart mobility solutions within the context of smart cities. This theme includes a range of elements, including sustainable transportation systems, active transportation modes, efficient public transportation, pedestrian-friendly infrastructure, network traffic management, and the integration of low-emission vehicles [5,54].

The paradigm shift from traditional reactive approaches to traffic management toward proactive measures is emphasized by the infusion of machine learning algorithms and data analytics into urban planning frameworks [26]. This transition not only emphasizes the significance of predictive models in accurately forecasting traffic flow but also provides valuable insights for decision-makers to enhance road accident management within the smart city context.

Moreover, these proactive measures extend beyond traffic optimization; they contribute to broader environmental goals. Studies by Saleem et al. [39] and Tao et al. [5] highlight the potential of machine learning techniques in reducing fuel consumption, carbon dioxide emissions, and the overall carbon footprint. The findings also suggest interventions such as early alarms, speed governors, and traffic calming measures as effective means to mitigate the impact of influential factors on crash occurrences [29]. This integrated approach not only improves road safety but also aligns with sustainability objectives within smart city planning.

The integration of advanced technologies, particularly Graph Neural Networks (GNNs), also plays a crucial role in real-time traffic-speed estimation for smart cities. GNNs leverage the inherent graph structure of urban road networks to capture spatial and temporal dependencies. A notable example is the Spatio-Temporal Graph Generative Adversarial Network (STGGAN), which stands out for its integration of spatio-temporal data, graph structures, edge features, and a residual structure. This results in highly accurate real-time traffic-flow prediction [31,37,43].

Transport Systems and Traffic Flow is a major theme that encompasses the crucial role of smart mobility solutions within the context of smart cities. This theme has been further divided into three subthemes: vehicle speeds, traffic congestion, and improvement in existing traffic forecasting studies.

3.1.1. Vehicle Speeds

Accurately predicting vehicle speeds is essential for developing efficient traffic management systems. This involves creating models that can forecast the speed of vehicles in various conditions, which is critical for optimizing traffic flow and minimizing delays. Accurate speed predictions contribute to smoother traffic operations, reduce fuel consumption, and decrease emissions. The models that focus on the need for precise and reliable speed prediction models often employ advanced machine learning techniques to handle the complexity of urban traffic environments [5]. The ability to predict vehicle speeds helps in implementing real-time adjustments to traffic signals and managing speed limits dynamically, thereby improving overall traffic efficiency. By isolating vehicle speeds as a subtheme, researchers can specifically address the unique challenges and technological requirements needed to improve speed prediction models. This focused approach allows for the development of specialized algorithms and tools tailored to enhance vehicle speed forecasting.

One example is the development and application of machine learning models to predict vehicle speeds in Cauayan City, Philippines. This accentuates the challenges faced by developing countries in implementing smart city initiatives. The study utilizes gradient-boosting tree models, incorporating road properties, day and time, and land use data as features. The machine learning model aims to predict a 1-min average probe vehicle speed based on various factors, including land use patterns, road properties, day-of-week, and hour-of-day features [32]. The findings of the study not only provide valuable insights for urban planners regarding impactful predictors of traffic behavior but also underline the significance of considering road features, land use, and time-related factors in effectively modelling vehicle speeds in developing cities.

Machine learning algorithms also play a pivotal role in predicting traffic flow patterns based on historical and real-time data, weather conditions, and other influencing factors, thereby enhancing urban transportation management [5].

One study focuses on the utilization of machine learning strategies, including decision trees, random forests, extra trees, and XGBoost (2.0.0.), to classify Internet of Vehicles (IOVs) traffic in the road system. This research analyzes the link between the characteristics of the machine learning dataset and its accuracy in classifying real-world IOV traffic. The study also proposes an Intelligent Intrusion Detection System (IDS) for IOV-dependent Vehicle-to-Network Communication (VNT) using the tree-based machine learning technique KNN (K-Nearest Neighbours). Therefore, the study emphasizes the effectiveness of tree-based machine learning models in classifying IOV traffic and intrusion detection in smart cities [31].

Another study introduces a lightweight machine learning-based data collection protocol named ML-TDG to address the challenges of data collection and communication in real-time traffic environments with millions of vehicles. This protocol utilizes real-time traffic data, distinguishes indexes based on assigned criteria, and collects data with minimal sources. The study focuses on the integration of Machine Learning into Vehicular Ad Hoc Networks (VANETs) and Intelligent Transport Systems (ITS) to enhance traffic management, data processing, and communication efficiency [26,38].

Data from Tumakuru city service buses are also used for model development. The non-linear model employs extreme gradient boosting (XGBoost) to estimate BAT based on factors like land use patterns, signalized intersections, and section length.

The proposed Dynamic BAT Estimation Model considers preceding trip data, estimating BAT for individual bus stops. Adjusted Travel Time is calculated as the weighted average of forecasted and estimated travel time. The results and discussion indicate that the proposed model outperforms the XGBoost model in predicting travel time, but, overall a hybrid model considering live traffic flow, incidents, and delays for better predictions is suggested [30].

3.1.2. Traffic Congestion

Urban traffic congestion is a major problem, resulting in longer travel times, higher fuel consumption, and increased pollution. Effective congestion management is vital for enhancing the quality of life in smart cities. Machine learning techniques can analyze patterns and predict congestion, enabling proactive measures to mitigate traffic jams. Studies have shown that machine learning models can effectively predict congestion patterns by analyzing historical traffic data and real-time inputs from various sensors and devices. Focusing on traffic congestion as a subtheme allows researchers to delve into specific strategies for congestion prediction and management. This includes developing algorithms that can optimize traffic signal timings, suggest alternative routes, and manage traffic in real time to alleviate congestion.

In the context of addressing traffic congestion in large urban cities, a significant factor is the quest for parking spaces. This subsection delves into the intricacies of traffic contingency and parking issues, aligning with the broader theme of traffic flow. One paper aligning with this theme introduces a framework utilizing diverse machine learning models to predict long-term parking space occupancy specifically in Los Angeles. The study meticulously evaluates over 10 machine learning models, incorporating both parking data and external factors like meteorological data. The primary objective is to enhance prediction accuracy and subsequently improve traffic flow within urban areas, as emphasized by Saleem et al. (2022) [39].

Building upon previous work that predominantly focused on short-term predictions and limited parking lots, other papers emphasize the efficiency of various machine learning models in tackling urban mobility challenges. The need to predict parking space availability is underscored, and the integration of multisource data and deep learning techniques is explored for reducing traffic congestion [29]. Furthermore, the integration of Federated Learning with Machine Learning Algorithms (FL-MLA) is discussed as a key element in reducing traffic congestion and enhancing travel safety [50].

To achieve efficient traffic flow and road safety, low-latency communication between vehicles and Roadside Units (RSUs) is emphasized. Machine learning algorithms are proposed to address safety, communication, and traffic issues associated with Vehicular Ad Hoc Networks (VANETs). The article also explores the combination of meta-heuristics methods with machine learning techniques for IoT-based smart monitoring applications.

A multi-stage prediction approach using attention-based convolution neural networks with long short-term memory (CNN-LSTM) is also discussed for enhancing energy consumption, traffic estimation, parking occupancy, efficiency ratio, and accident detection. The integration of FL-MLA in traffic management is highlighted for its potential to significantly reduce traffic congestion. Therefore, the use of fuzzy logic control systems in traffic management and smart parking emphasizes the role of expert knowledge in decision-making. It also explores the application of fuzzy models in traffic light management systems to minimize waiting times at red lights. The use of artificial neural networks (ANN) is presented to forecast traffic flow, especially in adverse weather conditions [25,53].

3.1.3. Improvement in Existing Traffic Forecasting Studies

Continuous improvement in traffic forecasting is crucial for adapting to the ever-changing dynamics of urban traffic. Enhancing existing forecasting models ensures that they remain effective and relevant in providing accurate predictions, which are critical for efficient traffic management. One can observe the need to refine and upgrade traffic forecasting models to incorporate new data sources and advanced machine learning techniques. Researchers have been working on improving the accuracy and reliability of these models to better serve the needs of smart cities. By highlighting this subtheme, researchers can systematically evaluate and enhance the performance of current models. It encourages a continuous feedback loop where models are regularly tested, validated, and updated to incorporate the latest advancements in machine learning and data analytics.

In considering the threat landscape within transportation flow and networks, the backdrop highlights the vulnerability of in-vehicle communications to cyber-attacks, posing potential risks to overall communication structures and human safety. The absence of robust security mechanisms for vehicles is identified as a potential source of chaos in urban settings. This subsection delves into the exploration of cybersecurity measures in traffic, highlighting the imperative of Intrusion Detection Systems (IDS) and the utilization of machine learning (ML) techniques for anomaly detection.

Various works in the field have proposed IDS solutions using ML algorithms, addressing a spectrum of cyber threats. Sousa et al. (2023), however, critically note the oversight in many of these works regarding non-trivial ML issues such as data distribution and class balance, tending to neglect simpler and more interpretable ML models [55]. The research employs simulated scenarios in a virtual machine using the NS-3 network simulator, containing different maps of Lisbon, Portugal, generated with SUMO (Simulation of Urban Mobility). Datasets include crucial information like simulation time, packet transmission/reception tags, identifier, and packet size. The article stresses the importance of data diversity in training sets for accurately classifying flooding behavior. Notably, the results highlight the effectiveness of training with multiple datasets compared to a single dataset.

In comparison with similar works, the proposed approach demonstrates superior results in terms of F1 scores despite employing simpler ML methods. The conclusion presents the significance of generating vehicular network datasets through simulations and the importance of data diversity for training ML algorithms to detect malicious behaviors, particularly flooding attacks in 5G-enabled vehicular networks [55]. Addressing the challenge of reliable test data for network intrusion detection and prevention systems, the article introduces supervised machine learning techniques, specifically decision tree, random forest, and XGBoost algorithms. Performance indices such as accuracy, precision, and recall are meticulously calculated for these algorithms, revealing high precision and recall rates.

The paper written by Shouaib et al. (2023) also acknowledges skepticism and a call for improvement in existing traffic forecasting studies [47]. Highlighting limitations, it proposes a novel approach using a sliding window technique to capture time-dependent features, integrating both machine learning and deep learning models for traffic prediction. Evaluation metrics include RMSE (Root Mean Square Error), MAE (Mean Absolute Error), and R² (R squared), with the LSTM (long short-term memory) model identified as the best-performing in terms of accuracy. The proposed method is validated by applying it to a different dataset, showcasing its potential for broader applications [47].

Results from different articles in the PRISMA review indicate that accurate prediction models significantly enhance transportation system efficiency in smart cities. The studies highlight the significance of traffic flow prediction in reducing congestion, implementing dynamic traffic routing systems, and overall urban planning. It discusses the application of predictive models in optimizing public transportation systems, reducing waiting times, and lowering carbon emissions. The emphasis should be put on the promising role of traffic flow prediction in achieving sustainable transportation systems in smart cities, urging future research to refine predictive models and explore socio-economic impacts. By leveraging traffic flow prediction, sustainable transportation strategies can be tailored to each city’s needs, resulting in more efficient and eco-friendly transportation systems [5,25,33,44].

3.2. Flow Tracking System (13 Papers)

The second theme of this PRISMA analysis delves into the escalating trend of urbanization. The efficient management of human and vehicular flow within urban environments is critical for the functionality of smart cities. Flow tracking systems leverage machine learning to monitor and predict movement patterns, enabling better infrastructure utilization and enhancing public safety. This section explores how various studies have addressed these challenges and the innovative solutions proposed to optimize flow tracking systems in smart cities.

With 55% of the global population currently residing in urban areas and projections indicating a rise to 68% by 2050, it becomes crucial to comprehend the flow of people and events in this urban landscape. This understanding is made possible through technologies like machine learning and edge computing, which play pivotal roles in shaping sustainable smart cities.

As urbanization intensifies, the movement of people in smart cities leads to challenges such as congestion and security risks. Real-time understanding of people flow statistics is essential to managing these challenges effectively. This not only aids in efficient city management but also elevates service levels in public areas.

The theme explores various studies and technologies related to people flow statistics, with a particular focus on machine learning and deep learning. These technologies offer practical insights into predicting and optimizing the dynamic movements within urban environments. The narrative aims to unravel the complexities of urban life, where data-driven insights pave the way for responsive and intelligent city planning.

The authors suggest a smart system for detecting and recognizing unusual human flow patterns in smart cities [46,49,62]. They create a people flow statistics system employing machine vision and deep learning, implementing the YOLOv3 algorithm for pedestrian detection and a Mean Shift tracking algorithm for improved tracking. Experimental simulations validate the system’s practicality.

In the context of population growth and urbanization, the focus also falls on utilizing machine learning for enhancing urban mobility planning, particularly in Bilbao, Spain, through the URBANITE project. The URBANITE project is described as an open-source ecosystem for testing mobility policies, providing valuable insights and a considerable acceleration in simulations using the ML module. The discussion concludes with plans to extend the ML approach to other pilot cities within the URBANITE project, enhance user-friendliness, and upgrade the research prototype into a comprehensive, open-source system for European cities [49,62]. In this case, the machine learning approach uses data from microscopic traffic simulators to optimize city mobility policy proposals. Multiple machine learning algorithms, such as logistic regression, decision tree, random forest, linear support-vector regression, and gradient-boosting regressor, are also employed in other articles for policy prediction [45,46,49,62].

As a method for enhancing smart city surveillance, the application of Machine Learning in Motion Detection and Pedestrian Tracking-assisted Intelligent Video Surveillance Systems (ML-IVSS) also plays a pivotal role. The ML-IVSS model aims to improve abnormal behavior detection in surveillance systems by utilizing optical flows to calculate various parameters like acceleration, speed, direction, trajectory, and motion amplitudes. The model relies on a dual-stream Convolutional Neural Network (CNN) to address the lack of hidden movement data and uses a weighted sum of squares algorithm to minimize optical flow in the neighborhood.

The model integrates adaptive Gaussian mixture modelling, spatial-temporal convolutional neural networks, and object detection with visual tracking approaches to enhance abnormal behavior analysis. It highlights the importance of surveillance video services (SVS) in smart cities and how ML techniques can improve data analysis for large-scale surveillance models [42,59,61].

Smart cities, incorporating the concept of fuzzy logic for handling uncertainties, are also a crucial focus. The study conducted by Liu et al. (2023) utilizes machine learning algorithms, including logistic regression, K-Nearest Neighbours (KNN), support vector machine (SVM), and artificial neural network, to categorize cities as leaders or followers [56]. The SVM with a sigmoid kernel emerges as the most accurate model for future smart city predictions. Models are evaluated based on various parameters, leading to the prediction of 50 cities as either smart city leaders or followers. Validation involves comparing internet infrastructure performance improvements, revealing better progress in the leader group [56].

Also, implementation of regression algorithms with Gradient-Boosting Regressor (GBR) are discussed. The models are evaluated using t-tests, revealing GBR’s accuracy in predicting closure start times and durations. The ML approach demonstrates advantages in reducing simulation time and providing advanced decision-making insights [34,49].

It is also important to discuss the human rights implications of flow systems and surveillance in smart city technology. For example, in China, it is highlighted that the swift digital transformation of cities and the widespread adoption of advanced surveillance systems like facial recognition could become harmful to human rights, addressing the potential impact of omnipresent sensors on citizen activities and the ethical challenges posed by smart city technology. It is important to stress the significance of patent analysis as a tool for comprehending emerging technologies and their implication on citizens’ lives [56].

3.3. Green Innovation in Smart Cities (12 Papers)

Green innovation is a cornerstone of sustainable urban development. Smart cities aim to minimize their environmental impact through the adoption of energy-efficient technologies and sustainable practices. This section delves into how machine learning is being utilized to drive green innovation in energy management, focusing on areas such as smart street lighting, energy waste management, and the monitoring of air and water quality. [58]. The studies reviewed provide insights into the transformative potential of these technologies in creating more sustainable urban environments. This theme highlights the transformative impact of Machine Learning (ML) and deep learning on the implementation of green innovations, particularly in energy management and sustainability within smart cities.

Challenges in implementing deep learning in smart cities are acknowledged, and the feasibility of results in implementation and computation is considered. The information derived from the PRISMA analysis, for instance, underscores an advanced early warning system known as BARCS (Bacterial Risk Controlling System), which is engineered for the surveillance and management of hazards related to bacteria and disinfection byproducts within a drinking water distribution network. BARCS utilizes an artificial intelligence (AI) approach, specifically employing Support Vector Regression (SVR) for data-driven prediction. The system centers on the concentration of total chlorine (TCl) as a key factor for identifying and managing risks. The research involved a case study in a highly urbanized city in Southern China utilizing data from regular manual sampling and laboratory analysis provided by the local water supply company. The proposed framework involves a closed-loop monitoring-prediction-warning-advising system aiming to control the dual risk of bacteria and disinfection byproducts in urban drinking water distribution systems. The SVR model demonstrated effective prediction of TCl concentration throughout the distribution system, and BARCS showed comparable performance to other machine learning methods. The system’s robustness under data error conditions was also evaluated. BARCS is presented as a flexible and expandable framework with potential applications in smart cities, emphasizing real-time monitoring and online prediction for improved drinking water safety. Overall, BARCS is deemed a promising solution for the safe management of drinking water in future smart cities, with potential for further enhancements and applications [57]. The issue of particulate pollution causing health problems in Southern China is also brought up, particularly in Hong Kong. The need for a practical urban air pollution [52] simulation model is emphasized, and the challenges of plume dispersion in the urban canopy layer (UCL) are discussed. The article highlights the use of machine learning (ML), specifically an Artificial Neural Network (ANN) model, to predict neighborhood-scale PM10 concentration levels in a densely populated urban environment. The performance of the ML model is compared with a Computational Fluid Dynamics (CFD) model [28,58].

Studies in Hong Kong feature a roadside air quality monitoring station, major roads, sparse vegetation, and varied building heights. The ANN model, which comprises interconnected neurons, is used for pollution transport modelling. The ENVI-met CFD model simulates PM10 dispersion in the UCL using inputs such as wind speed, wind direction, air temperature, and relative humidity. Results indicate that the ANN model captures diurnal cycles and trends well, demonstrating good agreement with measurements for NO2 but less for particulate levels. The CFD model, while underestimating measurements, performs similarly to the ANN model for PM10 predictions. However, the ANN model requires significantly less computational time. The article concludes that the ANN approach is a useful tool for air quality predictions, especially in smart city applications, despite being a “black-box” model with limited contribution to knowledge development [60].

Another machine learning technique is the approach for mapping the Surface Urban Heat Island (SUHI) in a medium-sized Brazilian city with a focus on mitigating the impacts of climate change. The study uses environmental and socio-economic variables to estimate Land Surface Temperature (LST) and employs the decision tree method. The research aims to investigate the spatial distribution of SUHI throughout the year in the city of Presidente Prudente, Brazil.

Results indicate that the decision tree algorithm has a high correlation coefficient (r = 0.96), low mean absolute error (MAE = 1.49 °C), and low root mean square error (RMSE = 1.88 °C). Thematic maps show the spatial distribution of predicted LST and SUHI areas, with the model successfully classifying sectors during different seasons. The study discusses the correlation between attributes and LST, highlighting the influence of seasons, impervious surfaces, and socio-economic factors [36,51].

In conclusion, the dynamic relationship between smart cities, machine learning (ML) techniques, and their practical applications highlights a transformative paradigm in urban development. The studies navigate the complexities of urbanization, aligning with Sustainable Development Goals (SDGs) and leveraging ML to enhance smart city services. By addressing current challenges and presenting a model with real-world applications, the research contributes to the ongoing discourse on sustainability in smart cities. As the world grapples with unprecedented challenges, including the impact of the COVID-19 pandemic, the studies advocate for the role of ML in optimizing essential services such as electric vehicle charging and distribution network stability [37]. Moreover, the emphasis on quality data emerges as a cornerstone for the success of smart city projects.

3.4. Security/Cybersecurity (4 Papers)

The increasing reliance on digital infrastructure in smart cities brings with it significant security challenges. Ensuring the cybersecurity of smart city systems is paramount to protecting sensitive data, maintaining public safety, and preventing disruptions. This section examines the role of machine learning in enhancing cybersecurity measures within smart cities, highlighting key studies that address the detection and mitigation of cyber threats. The insights gained from these studies are crucial for developing robust security frameworks that safeguard urban infrastructures.

First, we introduce the challenges in deploying smart cities, emphasizing issues like sensor data integration, fraudulent activities, and security. One can observe the threat of surveillance capitalism to individual privacy, and a study proposes a novel consensus protocol for a permissioned blockchain system to address these challenges [41].

The dataset that emerged from the PRISMA analysis highlights the rising incidence of fraud in debit/credit card transactions, categorizing it into different types based on the perpetrators’ characteristics. PVarious fraud detection methods, including decision trees, rule-induction techniques, support vector machines, logistic regression, artificial neural networks, and meta-heuristics, are explored [27,63].

Different approaches, such as meta-classifier-based models and neural network-based techniques, are examined for fraud detection in cloud and IoT-based environments. It is discussed that the application of Dempster-Shafer theory, Bayesian learning, and rule-based filtering for fraud detection would be suitable. Also, the data compare different deception detection techniques, highlighting their pros and cons in cloud and IoT settings.

A novel LR-n-fold intelligent machine learning technique is introduced for detecting fraudulent transactions in smart societies. The proposed method involves data collection, reformation, and feature mining, with a focus on tracing online cheating cases. Logistic regression, along with n-fold model selection, is employed for analysis, and the performance is compared with other existing techniques.

The study uses datasets from European banks and addresses challenges like dataset imbalance through techniques like synthetic minority over-sampling. The proposed LR n-fold intelligent machine learning approach demonstrates high reliability and accuracy in detecting fraud in smart societies. Performance comparisons with other methods, including random forest-based decision tree models, are discussed, acknowledging certain limitations, especially in handling abrupt increases in datasets [27,63].

Urgency is also emphasized for transaction prioritization to manage network traffic effectively. The protocol utilizes a permissioned blockchain with dynamic leader election, a modified true random number generator, and a sophisticated machine learning algorithm [41]. Moreover, the study delves into the intricacies of transaction prioritization mechanisms, the blockchain-based smart city design, overall procedures of the consensus algorithm, leader election, block creation, and peer prediction-based feedback collection. Crucially, the security analysis comprehensively addresses potential threats such as majority attacks, DDoS attacks, replay attacks, and empty block attacks. The experimental evaluation section further validates the protocol’s prowess through simulated data analysis involving 1000 nodes, shedding light on the algorithm’s accuracy and its responsiveness to various parameters, including the number of peers, trustworthiness score, number of blocks generated, and history of vote-outs on leader candidate prediction values [41].

The dataset also discusses the growing influence of the Internet of Things (IoT) in the development of smart cities, accompanied by an increased risk of cybersecurity threats. Smart city applications utilizing IoT devices face security challenges such as zero-day attacks, resource constraints, and limited security functionalities. The proposed approach is a machine learning-driven method for detecting attacks and anomalies specifically tailored for IoT-enabled smart cities, with an emphasis on distributed fog networks. Ensemble methods combining models from different sources are explored for enhanced performance compared to single classifiers. The proposed anomaly detection model, using the UNSW-NB15 and CICIDS2017 datasets, tracks network traffic through fog nodes to identify cyber-attacks. The conclusion underscores the significance of automation and smart city systems while acknowledging the increasing threat of cyber-attacks, endorsing the use of ensemble models for improved cybersecurity in such contexts [35].

On a separate front, the battle against electricity theft in smart grids takes center stage in another research endeavor. The study puts forth a Machine Learning Boosting Classifiers-based Stacking Ensemble Model (MLBCSM) as a robust solution for detecting non-technical losses (NTLs) [48,64]. Recognizing the pivotal role of machine learning and deep learning algorithms in enhancing NTL detection accuracy in smart grids, the researchers navigate through challenges like detection accuracy and false positive rates associated with existing models. The proposed MLBCSM, a fusion of five boosting classifiers and an Adaptive Synthetic Sampling technique (ADASYN) to tackle data imbalance, outshines standalone models. The classification approach employs a stacking ensemble strategy, demonstrating remarkable performance in terms of ROC-AUC and PR-AUC values. The research concludes by underscoring the MLBCSM’s advantages, including reduced false negative rates, low false positive rates, and improved performance compared to baseline models like AdaBoost [48].

Shifting the focus to real-time event detection in smart cities, a pioneering framework emerges, leveraging sentiment analysis and event detection data from social media platforms like Twitter [40]. This framework, motivated by the escalating real-time reactions of social media users in smart city contexts, offers a nuanced approach to detecting events and conducting citizen satisfaction analysis. Employing machine learning techniques, including natural language processing (NLP), the authors establish a risk taxonomy for smart cities, proposing responses to different event categories. A case study rooted in Cardiff City Council’s smart city roadmap provides practical insights into scenarios related to smart street lighting, parking, transportation, and the environment. The methodological approach, characterized by an optimized data pipeline, data sourcing, classification methods, and statistical significance evaluation, enriches the understanding of relationships between variables, thereby contributing valuable insights for smart city decision-making.

The comprehensive analysis of the selected literature underscores the revolutionary role of machine learning in advancing smart city development. By categorizing the research into key themes such as transportation systems, flow tracking, green innovation, and security, this review elucidates the multifaceted applications and benefits of machine learning techniques. These insights demonstrate how machine learning not only addresses urban challenges but also enhances the efficiency, sustainability, and safety of smart cities.

Furthermore, the incorporation of cutting-edge technologies into urban planning and management highlights the ongoing evolution of smart city concepts. The insights from this review offer a solid basis for future research and development, emphasizing the need for continued innovation and interdisciplinary collaboration. By leveraging data-driven approaches and fostering adaptive city planning, smart cities can achieve their full potential in creating sustainable, livable urban environments for the growing global population.

4. Limitations

One limitation of this study pertains to the exclusion of articles published in languages other than English, acknowledging the strength of academic research in Chinese, French, Japanese, German, etc. This exclusion could be a notable constraint. Another limitation involves excluding non-academic publications, recognizing that despite lacking rigorous blind peer review, commercial research can offer practical and timely insights. Third, the review’s reliance on published articles means that unpublished works, such as conference presentations, technical reports, or dissertations, may not have been included. These sources can sometimes provide valuable insights and preliminary findings that precede formal publication. Consequently, the exclusion of such gray literature might result in a less comprehensive understanding of the current research landscape.

Another limitation is influenced by the evolving nature of the field of machine learning in smart city development, which is relatively young and lacks established institutions and research patterns, facing challenges such as an inadequate vocabulary and research frameworks. The field suffers from an inadequate vocabulary and standardized terminologies. As machine learning applications in smart cities are interdisciplinary, involving elements from computer science, urban planning, engineering, and environmental sciences, there is often confusion and miscommunication among researchers and practitioners due to the lack of a unified language. This fragmentation hampers effective collaboration and slows down the progress of research and implementation. Also, the research frameworks and methodologies in this domain are still underdeveloped. There is no agreement on the optimal practices for integrating machine learning with urban systems, which leads to inconsistencies in study designs and evaluation metrics. This inconsistency makes it difficult to compare results across different studies, hindering the ability to draw generalizable conclusions and develop comprehensive models. The rapid pace of technological advancement in machine learning means that research findings can quickly become outdated. This constant state of flux requires continuous updating of knowledge and skills, which can be resource-intensive for researchers and practitioners. Additionally, the integration of cutting-edge technologies into existing urban infrastructure poses practical challenges, including compatibility issues and the need for substantial financial investments. Moreover, the field’s relative youth means that there is a scarcity of longitudinal studies that can provide insights into the long-term impacts and sustainability of machine learning applications in smart cities. Most current studies focus on short-term outcomes, leaving a gap in understanding how these technologies will perform and evolve over time.

The fourth limitation is associated with the constraints imposed by the selected articles in this review. We recognize that our search parameters might have overlooked significant articles, and it is possible that relevant studies were published after our search. The selection criteria, such as specific keywords, databases, and time frames, inherently restrict the scope of the review. While we employed a comprehensive and systematic approach using the PRISMA methodology to identify relevant studies, there is always a possibility that some pertinent research was excluded. This exclusion could be due to the use of different terminologies, variations in indexing across databases, or publication in less accessible journals.

Despite these constraints, we have endeavored to present a thorough and balanced overview of the application of machine learning techniques in smart city development. Our analysis highlights critical themes and discoveries, offering a detailed examination of the existing literature. This comprehensive review provides important perspectives on the present state of research, identifying gaps and suggesting future research directions.

In summary, while mindful of these constraints, this review serves as a beneficial resource for scholars and professionals aiming to deepen their knowledge of machine learning applications within the realm of smart city development. By pinpointing critical themes and discoveries, we hope to contribute to the ongoing discourse and encourage further exploration and innovation in this interdisciplinary field.

5. Conclusions and Discussions

The fragmented nature of current research on machine learning (ML) applications within smart cities necessitates an integrative approach to fully understand their transformative potential across diverse urban domains. Existing studies often focus narrowly on specific subareas such as transportation optimization, energy management, or security applications. While these focused studies contribute valuable insights, they do not offer a holistic view of how ML can revolutionize urban systems as a whole. By providing a comprehensive overview, this review bridges the gap between isolated research efforts, highlighting the interconnectedness of various urban domains. This integrative perspective is essential for creating smart cities that are technologically advanced, cohesive, and efficient at the same time.

The interplay between machine learning techniques and smart city development is an emerging area of study that holds significant promise. While there is a growing body of research on ML and urban computing, few studies explicitly analyze how ML can drive smart city growth and sustainability. This review delves into this critical intersection, demonstrating how ML can enhance urban planning, infrastructure management, and service delivery. By focusing on the synergies between ML and smart city initiatives, this review uncovers new avenues for innovation and highlights the significance of adopting ML-driven strategies to tackle intricate urban issues. The findings suggest that ML is not just an add-on to existing systems but a foundational element that can redefine urban living.

A critical aspect often overlooked in the discourse on ML and smart cities is the human and social dimension. Most studies emphasize technical capabilities and efficiency gains, neglecting the potential of ML to create citizen-centric intelligent urban environments. This review brings attention to the ways in which ML can enhance the quality of life for urban residents by improving public services, fostering community engagement, and ensuring equitable access to resources. By centering the discussion on citizens’ needs and experiences, this review advocates for a more inclusive approach to smart city development. It emphasizes that technological advancements should not only serve operational goals but also contribute to the overall well-being of the population.

Despite the technical advancements in ML applications for smart cities, there is a noticeable gap in the literature regarding practical implications for policymakers. Municipal authorities play a pivotal role in the adoption and implementation of ML technologies, yet much of the existing research fails to translate technical findings into actionable policy recommendations. At the same time, identifying the challenges and limitations associated with the integration of ML in smart cities is essential for guiding future research.

This research utilizing the PRISMA methodology provides valuable insights into the current state of research on machine learning techniques and their applications in smart city development. The analysis of 42 studies reveals four major interrelated themes that demonstrate the transformative potential of machine learning across diverse urban domains.

There are a few key research gaps that this article aims to address:

• Lack of a comprehensive overview integrating research on machine learning applications across diverse smart city domains. Many existing articles focus narrowly on specific subareas like transportation or energy without providing a big-picture view. This article provides an integrative review spanning key urban systems.
• Scarce literature examining the interplay between machine learning techniques and smart city development. While studies on machine learning and urban computing exist, few works specifically analyze the role of ML in enabling smart city growth and sustainability. This review contributes insights situated at the intersection of these areas.
• Limited synthesis of implications for policymakers regarding the adoption of machine learning in city governance and planning. Much research focuses on the technicalities without delineating practical takeaways for municipal authorities. This review extracts policy and planning-level insights.
• Few works systematically identify challenges, limitations, and gaps to guide future research at the confluence of machine learning and smart cities. This review concludes by outlining recommendations to advance work in this nascent domain.
• Lack of reviews centered on machine learning’s capacity to create citizen-centric intelligent urban environments. This work highlights the human and social dimension, rather than solely technical capabilities.

This article addresses the research gaps of lack of a multi-disciplinary integrative overview, implications for urban policymakers, and directions for future research at the intersection of machine learning and smart city development. It provides a holistic lens into this emerging field.

The literature highlights the pivotal role of machine learning in optimizing transport systems and traffic flow within smart cities. Advanced algorithms demonstrate significant promise in traffic prediction, congestion management, route optimization, accident reduction, and lowering emissions. Graph neural networks are noted for effectiveness in real-time traffic speed estimation leveraging spatial and temporal data. Studies also showcase applications in predicting parking availability to mitigate urban congestion issues.

Another prominent theme focuses on utilizing machine learning techniques to track the flow of people and events within smart city environments. This contributes to responsive mobility planning, security, surveillance, and overall efficient city management. Regression algorithms prove useful in predicting road closures, while CNN-LSTM models show potential in detecting abnormal crowd behaviors through surveillance systems.

Sustainability is a crucial theme, with papers underscoring machine learning’s contribution to energy efficiency, renewable integration, and environmental impact reduction in smart infrastructure. Use cases include drinking water quality monitoring, air pollution forecasting, managing electric vehicle charging loads, and mapping urban heat islands to mitigate climate change effects.

Machine learning is noted for its increasing role in the security and cybersecurity aspects of smart cities. Studies discuss techniques like intrusion detection systems, network traffic analysis, credit card fraud detection, and electricity theft prevention in smart grids.

5.1. Implications for Researchers

The insights from this review carry valuable implications for researchers aiming to advance the interdisciplinary field of machine learning and smart cities. Traffic prediction studies emphasize the need for hybrid models combining spatial, temporal, and external data like weather to boost accuracy, underscoring the importance of data fusion. Crowd flow tracking research indicates semantic segmentation algorithms may outperform detection-based approaches, suggesting exploring advanced computer vision techniques could be beneficial.

Energy studies point to a lack of standardized public datasets to evaluate and compare machine learning models. Efforts toward open benchmark datasets could aid progress in this area. Security studies highlight simpler interpretable ML models frequently outperforming complex deep nets, indicating that calibrated use of sophisticated vs simple algorithms should be pursued. Several papers call for real-world validation of models beyond simulation studies, implying that increased use of field studies and pilot deployments should be pursued by researchers.

5.2. Implications for Policymakers

The integration of machine learning within urban systems also carries crucial implications for smart city policymakers and administrators. Traffic optimization use cases make a strong case for investment in intelligent transportation systems and data-driven mobility planning. Environmental sustainability applications indicate the value of supporting energy analytics and emission monitoring systems. Crowd-tracking projects provide the rationale for surveillance and public safety systems with appropriate safeguards.

Discussions around data privacy and ethics emphasize the need for policies protecting individuals’ rights within smart cities. Studies reveal replicable frameworks for pilot projects can ease adoption by city officials, suggesting developing such frameworks should be prioritized. Several papers underscore the need for policies ensuring open data sharing between public and private entities to maximize innovation.

This review synthesizes key insights and provides clear, evidence-based recommendations for city governance and planning. It highlights the need for robust policy frameworks that support the integration of ML in urban systems, emphasizing issues such as data privacy, algorithmic transparency, and ethical considerations. By doing so, the review aims to facilitate the adoption of ML technologies in a way that is both effective and socially responsible.

5.3. Recommendations for Future Research

While the reviewed studies provide valuable insights, further research is required to realize the full potential of machine learning in urban environments. More pilot studies translating proofs-of-concept to field deployments across diverse cities could enhance model robustness and utility. Improving the interpretability of complex models should be prioritized to increase user trust and enable tweaking. Exploring the integration of simulation and empirical studies could lead to more robust evaluation.

Assessing machine learning model biases through an ethical lens is crucial to avoid marginalization. Comparing the ecological impacts of machine learning systems vs traditional approaches could reveal sustainability benefits. Analyzing the socio-economic implications of algorithms in urban life should be emphasized. Frameworks evaluating the return on investment of ML integration can assist adoption by city authorities.

This article highlights the tremendous promise of machine learning techniques in enabling the development of intelligent, sustainable, and human-centric smart cities of the future. It provides key insights for researchers to advance this nascent interdisciplinary domain and for policymakers to leverage its potential in serving citizens. The discussion also delineates recommendations to address research gaps through impactful studies that demystify, validate, and extend the role of machine learning in tomorrow’s urban environments.