| Journal of Smart Sensors and Computing
Received: 08 August 2026; Revised: 17 September 2026; Accepted: 19 September 2026; Published Online: 22 September 2026.
J. Smart Sens. Comput., 2026, 2(3), 26213 | Volume 1 Issue 3 (September 2026) | DOI: https://doi.org/10.64189/ssc.26213
© The Author(s) 2026
This article is licensed under Creative Commons Attribution NonCommercial 4.0 International (CC-BY-NC 4.0)
Edge AI and Federated Learning in Smart City
Engineering: A Cross-Domain Review of
Architectures, Applications, and Deployment
Challenges in Transportation, Environmental, and
Building Systems
Mariam Labib Francies
*
Department of Computer Engineering, ElSewedy University of Technology, Cairo-Ismailia Desert Road, 10th of Ramadan City,
Sharqia Governorate, Egypt
*Email: mariam.labib@sut.edu.eg (Mariam Labib Francies)
Abstract
keywordsArtificial Intelligence; Internet of Things; Smart Cities; Edge AI; Federated Learning; Privacy-Preserving
Machine Learning; Intelligent Transportation Systems; Urban Infrastructure; TinyML; Distributed Learning.
The twenty-first century has witnessed an unprecedented acceleration of global urbanization, fundamentally
transforming the landscape of human civilization and presenting profound challenges for urban engineering
systems. According to the United Nations, approximately 55% of the world's population currently resides in
urban regions, with projections indicating this proportion will exceed 68% by 2050.
[1]
This rapid urban
expansion generates substantial pressure on existing infrastructure systems, including transportation
networks, energy grids, water distribution systems, and waste management facilities. Traditional
engineering approaches, characterized by reactive maintenance strategies and siloed operational
frameworks, have proven inadequate to address the complexity and scale of contemporary urban
challenges.
[2]
Urban engineering systems face multifaceted challenges that demand innovative technological
solutions. Transportation networks in major metropolitan areas experience chronic congestion, with average
commuters in developed nations losing approximately 54 hours annually to traffic delays, resulting in
significant economic losses and environmental degradation.
[3]
Environmental infrastructure struggles with
increasing pollution levels, inadequate waste management systems, and inefficient resource utilization. Air
quality monitoring in urban centers reveals that over 90% of the global urban population lives in areas where
air quality exceeds World Health Organization guidelines.
[4]
Building energy consumption accounts for
approximately 40% of global energy use and 36% of carbon dioxide emissions, highlighting the critical need
for intelligent building management systems.
[5]
Water distribution networks suffer from aging infrastructure,
with developed nations experiencing water loss rates of 15-30% due to leakage, while developing nations
face losses exceeding 40%.
[6]
These engineering challenges are further compounded by the dynamic nature
of urban systems, characterized by high variability in demand patterns, complex interdependencies between
infrastructure components, and the need for real-time adaptive responses to changing conditions. Traditional
engineering methodologies, relying primarily on historical data analysis and static optimization models, lack
the flexibility and computational capacity to manage such complexity effectively.
[7]
Moreover, the increasing
frequency of extreme weather events attributable to climate change imposes additional stress on urban
infrastructure, necessitating resilient engineering solutions capable of predictive risk assessment and
adaptive response mechanisms.
[8]
1.2 Artificial intelligence and IoT as enabling technologies
The convergence of Artificial Intelligence (AI) and the Internet of Things (IoT) has emerged as a
transformative paradigm for addressing urban engineering challenges through data-driven decision-making
and autonomous system optimization. AI encompasses a broad spectrum of computational techniques,
including Machine Learning (ML), Deep Learning (DL), Natural Language Processing (NLP), and computer
vision, that enable systems to learn from data, recognize patterns, and make intelligent decisions without
explicit programming.
[9,10]
IoT technology facilitates the deployment of interconnected sensor networks,
actuators, and communication devices that enable real-time data collection, monitoring, and control of
physical infrastructure systems.
[11]
The integration of AI and IoT creates intelligent cyber-physical systems
capable of continuous learning, adaptive optimization, and autonomous decision-making. ML algorithms,
including supervised learning techniques such as Support Vector Machines (SVM), Random Forests, and
Neural Networks, enable predictive modeling of complex urban phenomena, from traffic flow patterns to
energy consumption forecasts.
[12]
DL architectures, particularly Convolutional Neural Networks (CNNs) for
spatial data analysis and Recurrent Neural Networks (RNNs) for temporal sequence modeling, have
demonstrated exceptional performance in processing high-dimensional sensor data and extracting
actionable insights.
[13]
Recent advances in edge computing and distributed AI architectures enable real-time
processing of IoT data streams, reducing latency and bandwidth requirements while enhancing system
responsiveness.
[14]
The deployment of IoT sensor networks in urban environments generates unprecedented
volumes of data, with estimates suggesting that smart cities will produce over 200 exabytes of data annually
by 2025.
[15]
This data encompasses diverse modalities, including traffic camera feeds, environmental sensor
readings, building automation system logs, and mobile device location traces. AI technologies provide the
computational frameworks necessary to transform this raw data into actionable engineering insights,
enabling predictive maintenance, real-time optimization, and evidence-based infrastructure planning.
[16,17]
1.3 Evolution of smart city engineering
The concept of Smart Cities (SC) has evolved significantly since its inception in the early 2000s, transitioning
from technology-centric deployments focused primarily on Information and Communication Technology
(ICT) infrastructure to holistic urban ecosystems that prioritize sustainability, livability, and citizen-centric
services.
[18]
Contemporary SC frameworks recognize that technological innovation must be integrated within
broader urban planning strategies that address social equity, environmental sustainability, and economic
development.
[9,19]
Smart city engineering encompasses six fundamental dimensions that collectively define
urban intelligence: smart governance, smart economy, smart mobility, smart environment, smart living, and
smart people.
[2,20]
Within this multidimensional framework, engineering applications of AI and IoT primarily
concentrate on three domains that form the focus of this review: transportation systems (smart mobility),
environmental infrastructure (smart environment), and building management (smart living). These domains
represent critical engineering challenges where AI-enabled solutions have demonstrated substantial impact
through real-world deployments. Transportation engineering has witnessed transformative innovations
through Intelligent Transportation Systems (ITS) that integrate AI algorithms with IoT sensor networks to
optimize traffic flow, reduce congestion, and enhance safety.
[21,22]
Advanced driver assistance systems and
autonomous vehicle technologies leverage DL models for real-time object detection, path planning, and
decision-making in complex urban environments.
[23]
Environmental engineering applications employ AI for
predictive air quality modeling, optimized waste collection routing, and intelligent water network
management.
[24]
Building automation systems integrate ML algorithms with IoT devices to achieve significant
energy efficiency improvements through predictive HVAC control, occupancy-based lighting optimization,
and fault detection diagnostics.
[25-27]
Despite the proliferation of AI and IoT implementations in urban
engineering, existing literature exhibits several critical gaps that this review addresses. First, prior studies
predominantly adopt narrow disciplinary perspectives, examining AI applications within isolated
engineering domains without exploring cross-domain synergies and transferable methodologies.
[26,28]
Second, while numerous case studies document individual smart-city deployments, comprehensive
comparative analyses examining implementation success factors, performance metrics, and scalability
challenges across diverse geographical and socio-economic contexts remain limited.
[29]
Third, existing
reviews often focus on theoretical frameworks and algorithmic innovations without adequately addressing
practical implementation challenges, including data privacy concerns, system interoperability issues, and
infrastructure integration requirements.
[30]
This review provides a comprehensive synthesis of AI and IoT
applications in smart-city engineering, offering several distinct contributions. First, we present a
multidimensional analysis spanning transportation, environmental, and building engineering domains,
identifying common methodological frameworks and domain-specific optimization strategies. Second, we
synthesize findings from several international case studies across developed and developing nations,
providing empirical evidence of implementation outcomes, performance metrics, and lessons learned. Third,
we systematically analyze technical challenges, including data management, algorithm selection,
computational infrastructure requirements, and integration with legacy systems, proposing evidence-based
mitigation strategies derived from successful deployments. Fourth, we examine ML and DL frameworks
specifically from an engineering implementation perspective, evaluating their suitability for different urban
applications based on computational efficiency, accuracy, scalability, and deployment complexity.This paper
adopts a narrative engineering review approach rather than a systematic or meta-analytic review. The
objective is to synthesize architectural patterns, deployment practices, and engineering trade-offs reported
in recent literature, rather than to statistically aggregate performance outcomes.
1.4 The transversal turn: Edge AI and federated learning
Within this broader AI/IoT trajectory, two methodological developments have emerged since 2020 that cut
transversally across application domains and are reshaping deployment architectures more than any single
algorithmic advance: Edge Artificial Intelligence (Edge AI) and Federated Learning (FL). Edge AI refers to the
local execution of inference and, increasingly, training on resource-constrained devices situated near the data
source, ranging from microcontroller-class TinyML platforms to gateway-class edge servers. FL, introduced by
McMahan et al.
[31]
and matured through subsequent surveys, enables collaborative model training across
distributed datasets without centralizing raw data, supporting privacy preservation, regulatory compliance, and
operational sovereignty.
These two paradigms are transversal in a strict sense: they are not domain-specific algorithms but deployment
patterns that recur across transportation, environmental, and building applications with similar architectural
signatures and similar engineering trade-offs. A traffic-prediction Graph Neural Network trained federatively
across cities,
[26]
an air-quality forecasting model deployed at the sensor edge for low-latency public health
response, and an HVAC controller using on-device reinforcement learning to preserve occupant data sovereignty
exhibit the same underlying methodological shift away from centralized cloud-only deployment toward
distributed, latency-aware, and privacy-preserving urban AI. Reviews that organize the field by application
domain miss this convergence; reviews that focus narrowly on one paradigm miss its breadth. This work adopts
Edge AI and FL as the transversal lens through which the three engineering domains are synthesized.
[32]
1.5 Research gap and contributions
Despite a substantial body of review literature on AI and IoT in smart cities, four specific gaps motivate this
work. First, existing comprehensive reviews.
[1,15,26,30]
organize coverage by application domain or by algorithmic
family, but no recent review systematically synthesizes Edge AI and Federated Learning as transversal
deployment paradigms across the three core engineering domains addressed here. Second, while individual
surveys on FL in smart cities
[26]
and Edge AI for IoT
[33]
have appeared, they treat their respective paradigms in
isolation rather than examining how the two combine in practice (federated edge learning, on-device
personalization, hierarchical edge-cloud-FL architectures). Third, prior reviews predominantly catalog
deployments without proposing taxonomies that support practitioner decision-making over when centralized,
edge, or federated approaches are warranted. Fourth, the absence of a documented review protocol in much of
the existing literature limits reproducibility and makes it difficult for readers to assess corpus completeness.
This review addresses these gaps through four contributions. First, it proposes a transversal taxonomy
classifying smart-city AI implementations along two orthogonal axes computation locality (cloud, fog, edge, on-
device) and learning paradigm (centralized, federated, hybrid) and populates this taxonomy with evidence from
each engineering domain.
[34]
Second, it extends a five-layer smart-city architecture with explicit Edge AI and FL
placement annotations, providing a reference model for architects designing next-generation deployments.
Third, it synthesizes implementation outcomes from over thirty international case studies through
standardized reporting tables that include reported performance, deployment scale, infrastructure dependency,
and methodological maturity, supporting evidence-based comparison across contexts. Fourth, it applies a
transparent PRISMA 2020-compliant selection protocol with documented search strings, screening criteria, and
inclusion flow, enabling reproducibility. These contributions are positioned with respect to six comparator
surveys in a related work comparison table in Section 2.5.
The remainder of this paper is organized as follows. Section 2 documents review methodology, including the
search protocol, inclusion and exclusion criteria, PRISMA-based study selection, and a structured comparison
against six recent comparator surveys. Section 3 presents the technical foundations, including the five-layer
reference architecture with Edge AI and FL annotations, fundamentals of Edge Intelligence and Federated
Learning, IoT sensor networks, and the ML/DL framework landscape relevant to urban deployment. Sections 4,
5, and 6 examine the transportation, environmental, and building engineering domains respectively, each
through the transversal Edge AI and FL lens. Section 7 synthesizes cross-domain patterns through the proposed
taxonomy and a method-by-domain matrix, identifying where each paradigm excels. Section 8 analyzes
implementation challenges and mitigation strategies. Section 9 outlines the research agenda. Section 10
discusses the limitations of this review. Section 11 concludes.
2. Review methodology
2.1 Review protocol and reporting standard
This review follows a structured protocol adapted from the PRISMA 2020 statement for transparent reporting
of systematic reviews. Although positioned as an engineering synthesis rather than a clinical meta-analysis, the
PRISMA flow and reporting checklist are adopted to maximize reproducibility, document corpus boundaries,
and align with current best practice in engineering review literature. The protocol was developed prior to the
formal search and includes a pre-specified research question, search strategy, eligibility criteria, screening
procedure, and data extraction template. No protocol amendments occurred during execution.
The guiding research question is: How are Edge AI and Federated Learning being deployed across
transportation, environmental, and building engineering domains in smart cities, what architectural patterns
and engineering trade-offs characterize successful implementations, and what open problems limit production-
grade adoption? Three sub-questions structure the synthesis: (i) what reference architectures support cross-
domain Edge AI and FL deployment; (ii) what performance, energy, latency, and privacy outcomes have been
reported empirically; and (iii) what implementation challenges and mitigation strategies are evidenced in the
case-study record.
2.2 Search strategy
Literature was identified through systematic searches of five databases: IEEE Xplore, Scopus, ScienceDirect
(Elsevier), ACM Digital Library, and Web of Science Core Collection. The search was conducted in February 2026
and covered publications from January 2021 to February 2026. A Boolean search string was developed
iteratively through three pilot searches and finalized as follows:
("smart city" OR "smart cities" OR "urban computing" OR "smart urban") AND ("edge AI" OR "edge intelligence"
OR "edge computing" OR "TinyML" OR "federated learning" OR "federated AI" OR "distributed learning") AND
("transportation" OR "traffic" OR "mobility" OR "environment*" OR "air quality" OR "waste" OR "water" OR
"building*" OR "HVAC" OR "energy management")
Database-specific syntax adjustments were applied where Boolean operators differ. To capture seminal pre-
2021 work where appropriate, backward citation tracking was applied selectively to the FL foundations and to
the Edge AI taxonomy literature
[33]
Forward citation tracking via Google Scholar was used to identify post-
publication extensions of the six comparator surveys (Section 2.5). Grey literature was excluded to maintain
peer-review provenance, except for authoritative technical reports from the United Nations, World Bank, and
municipal smart-city programmes cited only as contextual background. Vendor documentation was consulted
only where peer-reviewed alternatives describing the same deployment were unavailable and is flagged as such
in the case-study tables.
2.3 Eligibility criteria
Inclusion criteria required that a study (1) addresses a smart-city or urban-engineering application within
transportation, environmental infrastructure, or building management; (2) employs AI, ML, or DL methods,
with explicit relevance to Edge AI, FL, or distributed urban intelligence broadly construed; (3) reports either a
system architecture, an empirical deployment, a benchmark evaluation, or a comparative analysis; (4) appears
in a peer-reviewed journal or established conference proceedings; and (5) was published between January 2021
and February 2026, with backward extensions only for foundational FL and Edge AI works.
Exclusion criteria removed studies that (1) addressed AI methods without urban-engineering context; (2)
focused exclusively on algorithmic novelty without deployment, architectural, or comparative content; (3)
appeared in non-peer-reviewed outlets, predatory venues, or sources flagged by the Cabells Predatory Reports
list; (4) duplicated content already represented by a more comprehensive companion publication from the same
authors; or (5) were not retrievable in full text in English. Pre-prints were included only where the same content
had subsequently been accepted for publication, with the published version cited.
2.4 Study selection and PRISMA flow
Database searches returned an initial pool of records that were de-duplicated by DOI matching and title
similarity. Title and abstract screening was conducted against the inclusion and exclusion criteria, with
uncertain cases retained for full-text review. Full-text screening removed studies that did not meet the
deployment, architectural, or comparative content requirement. Backward and forward citation tracking added
a small number of seminal works that the database queries had missed. The final corpus exceeded 200 included
studies and forms the evidence base for Sections 3 through 9. The PRISMA 2020 flow is summarized in Fig. 1.
Fig. 1: The PRISMA 2020 flow.
Table 1 reports the per-database identification counts underlying Fig. 1. Deduplication was performed on the
pooled corpus across all five databases rather than per database, since a substantial proportion of duplicate
records occur across sources; consequently, screening, eligibility, and inclusion counts in Fig. 1 are reported
only at the pooled level and are not disaggregated by originating database.
2.5 Comparison with related surveys
To position this review within the existing literature, six recent comparator surveys were selected from the 2022
to 2026 window for structured comparison. The comparator set was chosen to span the spectrum from broad
AI-in-smart-cities reviews
[1,15,30]
to paradigm-focused surveys on FL
[26]
and Edge AI,
[33]
with one current high-
profile cross-domain Edge AI survey serving as the most direct methodological reference. Table 2 summarizes
the comparison along seven dimensions: review methodology, PRISMA usage, bibliometric layer, scope breadth,
transversal lens, taxonomy contribution, and figure-and-table density. The comparison establishes that no prior
review combines a documented PRISMA-compliant protocol with a transversal Edge AI plus FL lens applied
across all three engineering domains addressed here.
Table 1: Database-wise identification statistics
Database
Records Identified
IEEE Xplore
412
Scopus
587
ScienceDirect
341
ACM Digital Library
198
Web of Science
463
Citation tracking (backward/forward)
47
Total
2,048
Table 2: Comparison of this review with six recent comparator surveys
Study (ref)
Domain scope
Lens / focus
Protocol
Taxonomy
Figs / Tables
Herath & Mittal
2022
[1]
Broad SC
AI adoption
Narrative
No formal
Moderate
Nguyen et al. 2024
[15]
Broad SC
IoT + AI roles
Narrative
Domain catalog
Moderate
El Ghati et al. 2024
[30]
Broad SC
Visual IoT + AI
Narrative
Application-driven
Moderate
Zheng et al. 2022
[26]
Multi-domain
FL only
Narrative
FL applications
Moderate
Pandya et al. 2023
[35]
Multi-domain
FL only
Narrative
FL challenges
High
This work
Transport + Env
+ Build
Edge AI + FL
transversal
PRISMA 2020
2-axis taxonomy + 5-
layer arch.
7+ / 6+
3.Technical foundations for Edge AI and federated learning in smart cities
3.1 Five-layer smart-city reference architecture
This section presents a five-layer reference architecture (Fig. 2) for smart-city engineering applications,
addressing complex interactions among sensors, communication networks, data processing systems, and end-
user applications. Unlike the canonical five-layer IoT model, this version annotates each layer with explicit Edge
AI and FL placement decisions, providing a deployment-aware reference for architects. The architecture
comprises: Sensing Layer (bottom) for data acquisition, Transmission Layer for communication, Data
Administration Layer for processing and intelligence, Application Layer for service delivery, and Business Layer
(top) for governance and economic models. This layered approach ensures scalability, interoperability, and
security while accommodating the heterogeneous nature of urban infrastructure systems.
Fig. 2: A five-layer framework for smart city architectures.
3.1.1 Sensing layer
The sensing layer constitutes the foundation of smart-city infrastructure, responsible for real-time data
collection from physical environments through IoT devices, sensors, and Wireless Sensor Networks (WSNs).
This layer monitors diverse parameters critical to urban engineering systems, including traffic flow patterns,
and flow rates, structural health indicators, and environmental conditions (temperature, humidity, noise levels,
solar radiation, wind speed, precipitation).
[36]
Modern sensing infrastructures deploy multiple sensor modalities
to capture comprehensive urban data. Traffic monitoring employs inductive loop detectors, radar sensors,
LiDAR systems, and computer vision cameras for vehicle detection and classification.
[37]
Environmental
monitoring utilizes electrochemical sensors for gas detection, optical particle counters for particulate matter
measurement, and acoustic sensors for noise pollution assessment.
[38]
Building automation systems integrate
Microelectromechanical Systems (MEMS) sensors for occupancy detection, temperature regulation, and energy
metering.
[25]
Advanced GPS terminals and mobile device sensors provide location-based data for transportation
planning and emergency response coordination.
[11]
The sensing layer faces several technical challenges,
including sensor calibration drift, data quality assurance, power consumption optimization for battery-
operated devices, and harsh environmental conditions affecting sensor reliability.
[38]
Edge computing
capabilities increasingly deployed at the sensing layer enable preliminary data processing, filtering, and
aggregation, reducing transmission bandwidth requirements and improving system responsiveness.
[33]
Infrastructure monitoring sensors deploy accelerometers, strain gauges, displacement sensors, and corrosion
sensors to assess bridge, tunnel, and building integrity. Recent advances in edge-to-cloud computing
architectures for structural health monitoring demonstrate the effectiveness of hierarchical processing
strategies that combine local edge analytics for real-time anomaly detection with cloud-based comprehensive
structural analysis.
[39]
Fiber optic sensors provide distributed strain and temperature measurement along
kilometers of infrastructure, detecting damage and fatigue.
3.1.2 Transmission layer
The transmission layer facilitates secure and efficient data transfer from the sensing layer to upper processing
layers through heterogeneous communication technologies. This layer must accommodate diverse
requirements including data rate, latency, range, power consumption, and reliability across different urban
applications.
[16]
Short-range communication technologies such as Bluetooth Low Energy (BLE), ZigBee, Z-Wave,
and Near Field Communication (NFC) provide low-power connectivity for localized sensor networks in
buildings and confined spaces, typically operating within ranges of 10-100 meters.
[11]
Long-range
communication technologies enable city-scale connectivity through cellular networks (3G, 4G LTE, 5G) and Low-
Power Wide-Area Networks (LP-WAN) including LoRaWAN, Sigfox, and NB-IoT.
[36]
Fifth generation (5G) cellular
networks represent a transformative advancement for smart-city applications, offering peak data rates
exceeding 10 Gbps, latency below 1 millisecond, and massive device connectivity supporting up to 1 million
devices per square kilometer.
[40]
These capabilities enable real-time applications such as autonomous vehicle
coordination, augmented reality navigation, and instantaneous emergency response systems. Emerging
communication technologies expand connectivity options for specialized applications. Li-Fi (Light Fidelity)
utilizes LED lighting infrastructure for high-speed data transmission, achieving data rates exceeding 100 Gbps
in laboratory conditions while providing simultaneous illumination and communication functions.
[41]
Vehicular
communication protocols including Dedicated Short-Range Communications (DSRC) and Cellular Vehicle-to-
Everything (C-V2X) enable Vehicle-to-Vehicle (V2V), Vehicle-to-Infrastructure (V2I), and Vehicle-to-Pedestrian
(V2P) data exchange for cooperative intelligent transportation systems.
[37]
Network slicing capabilities in 5G
architectures enable customized virtual networks optimized for specific application requirements, allowing
simultaneous support for high-bandwidth video surveillance, ultra-reliable low-latency autonomous vehicle
control, and massive IoT sensor connectivity over shared physical infrastructure.
[40]
Software-Defined
Networking (SDN) and Network Function Virtualization (NFV) provide flexible network management, enabling
dynamic resource allocation and quality-of-service optimization based on real-time demand patterns.
[15]
3.1.3 Data administration layer
The data administration layer functions as the central intelligence of smart-city systems, responsible for data
integration, storage, processing, and decision-making. This layer transforms raw sensor data into actionable
engineering insights through sophisticated analytics pipelines.
[10]
Data integration mechanisms aggregate
heterogeneous data streams from multiple sources, performing data cleaning, normalization, temporal
synchronization, and semantic harmonization to create unified datasets suitable for analysis.
[42]
Cloud
computing platforms provide scalable storage and computational resources for processing massive volumes of
urban data. Leading cloud service providers including Amazon Web Services (AWS), Microsoft Azure, and
Google Cloud Platform offer specialized IoT services with integrated ML capabilities, enabling rapid deployment
of analytics applications without extensive infrastructure investment.
[43]
Hybrid cloud architectures combining
public cloud resources for scalable batch processing with private cloud infrastructure for sensitive data
management provide balanced solutions addressing performance, security, and cost considerations.
[15]
Big data
processing frameworks including Apache Hadoop, Apache Spark, and Apache Flink enable distributed
processing of petabyte-scale datasets generated by urban sensor networks.
[44]
Stream processing engines
facilitate real-time analytics on continuous data streams, supporting applications requiring immediate decision-
making such as traffic signal optimization and emergency detection systems.
[45]
Time-series databases
optimized for temporal data storage and retrieval, including InfluxDB, TimescaleDB, and OpenTSDB, provide
efficient management of sensor data characterized by high ingestion rates and time-based query patterns.
[46]
AI
and ML algorithms deployed at this layer extract patterns, generate predictions, and support automated
decision-making. Supervised learning models predict traffic congestion, energy demand, and air quality levels
based on historical patterns and real-time conditions.
[12]
Unsupervised learning techniques identify anomalies
in sensor data, detecting equipment failures, unusual traffic patterns, and potential security incidents.
[47,48]
Reinforcement learning algorithms optimize control strategies for traffic signal coordination, building HVAC
systems, and energy grid management through continuous interaction with the environment.
[49]
Data
governance frameworks ensure data quality, privacy protection, access control, and regulatory compliance
throughout the data lifecycle. Differential privacy techniques enable statistical analysis of sensitive datasets
while preserving individual privacy.
[50]
Blockchain-based data management systems provide immutable audit
trails and decentralized trust mechanisms for multi-stakeholder data sharing scenarios.
[51]
Federated learning
architectures enable collaborative ML model training across distributed datasets without centralizing raw data,
addressing privacy concerns while leveraging collective intelligence.
[52]
3.1.4 Application layer
The application layer delivers smart-city services to end users, including citizens, city operators, emergency
responders, and infrastructure managers. This layer translates data-driven insights into user-facing
applications and automated control systems.
[13]
Transportation applications provide real-time traffic
information, route optimization, parking availability, and public transit schedule updates through mobile
applications and digital signage.
[3]
Environmental monitoring dashboards visualize air quality indices, noise
levels, and weather forecasts, enabling citizens to make informed decisions about outdoor activities.
[53]
Building
management systems integrate occupancy prediction, energy optimization, lighting control, and HVAC
automation to enhance comfort while minimizing energy consumption.
[5]
Emergency response platforms
aggregate data from multiple sources to provide situational awareness during disasters, coordinating resource
deployment and evacuation procedures.
[54,55]
Utility management applications optimize water distribution,
detect leakage, predict equipment failures, and schedule maintenance activities.
[56]
Application Programming
Interfaces (APIs) enable third-party developers to create innovative services leveraging smart-city data and
infrastructure. Open data initiatives provide access to non-sensitive urban datasets, fostering entrepreneurship
and civic engagement.
[57]
Mobile-first design principles ensure accessibility across diverse device platforms,
while progressive web applications provide app-like experiences without requiring installation. Augmented
reality interfaces overlay digital information onto physical environments, assisting maintenance personnel,
tourists, and emergency responders with contextual information.
[58]
Interoperability standards including Open
Geospatial Consortium (OGC) SensorThings API, FIWARE NGSI-LD, and oneM2M ensure seamless integration
across heterogeneous smart-city platforms and vendor solutions.
[59]
Semantic web technologies and ontologies
provide standardized vocabularies for describing urban data, enabling automated reasoning and cross-domain
analytics.
[15]
3.1.5 Business layer
The business layer encompasses organizational processes, governance structures, and economic models that
enable sustainable smart-city operations.
[60]
Public-Private Partnership frameworks allocate responsibilities
and risks between government entities and private sector technology providers. Total Cost of Ownership
analyses in the smart-city literature generally find that operational expenses dominate lifecycle costs, with
initial capital expenditure commonly representing a minority share of total 10-year cost in municipal
deployments. Sustainability models include user fees, efficiency-based cost recovery, and operational
expenditure subscriptions
[61]
General Data Protection Regulation compliance imposes substantial costs on
European municipalities, and multi-jurisdictional deployments face regulatory fragmentation that further
increases costs and complexity
[50]
Service Level Agreements, performance metrics, and cybersecurity
governance establish accountability frameworks ensuring long-term operational success beyond technical
deployment.
[62]
3.2 Edge AI fundamentals for urban deployment
Edge Artificial Intelligence refers to the execution of AI inference, and increasing training, on devices located at
or near the data source rather than in remote cloud datacenters
[33]
In smart-city deployments, Edge AI spans a
hierarchy of compute classes: ultra-constrained TinyML platforms (microcontroller-class devices with kilobytes
of RAM, watt-level power budgets) suited to always-on sensor analytics;
[63]
gateway-class edge servers (single-
board computers, edge GPUs) handling neighborhood-scale workloads such as camera-based traffic analytics;
and fog or near-edge servers aggregating data across districts. The principal engineering rationales for edge
deployment in urban contexts are four: latency reduction for control-loop applications such as autonomous
vehicle perception and adaptive traffic signaling, bandwidth conservation by transmitting only summaries or
events rather than raw streams, energy and cost reduction over continuous cloud inference at scale, and privacy
preservation by processing identifiable data locally.
[64]
Realizing edge deployment imposes constraints absent from cloud workloads. Model compression techniques
including post-training quantization (8-bit or lower), structured and unstructured pruning, and knowledge
distillation reduce model footprint and inference cost while preserving accuracy
[65]
Hardware-aware neural
architecture search produces models matched to specific edge accelerators such as Google Coral, NVIDIA Jetson,
and Arm Ethos-U. On-device adaptation through transfer learning and continual learning addresses the model-
drift problem characteristic of dynamic urban environments, where deployed models must adapt to seasonal,
infrastructural, and behavioral change without full retraining. The same techniques underpin uncertainty-
aware sensor placement and monitoring frameworks that quantify epistemic uncertainty under sparse edge
coverage.
[66,67]
3.3 Federated learning fundamentals for urban deployment
Federated Learning trains a shared global model across distributed clients (devices, organizations, or
jurisdictions) without centralizing the underlying data
[68]
Each client computes local model updates on its
private dataset; the updates are aggregated, typically through Federated Averaging (FedAvg)
[31]
or related
communication-efficient variants, to produce an improved global model. The data therefore never leaves its
source, addressing the privacy and regulatory constraints that block centralized learning across sensitive urban
datasets such as resident mobility traces, household energy consumption, and clinical records.
[69]
FL in urban systems instantiates along three principal axes. The first is the partitioning of feature versus sample
space: horizontal FL pools clients holding similar features over different entities, suitable for cross-city traffic
prediction or cross-building energy modeling; vertical FL pools clients holding different features over the same
entities, suitable for integrated utility and transportation analytics within a single city. The second axis concerns
coordination topology, ranging from centralized parameter-server aggregation through hierarchical FL across
edge-cloud tiers to fully decentralized peer-to-peer learning
[26]
The third axis covers privacy hardening through
differential privacy noise injection,
[50]
secure aggregation cryptographic protocols, and homomorphic
encryption, each trading communication and computation cost for stronger privacy guarantees. Recent work
has extended FL to handle non-independent and identically distributed (non-IID) urban data through
personalization layers and client-specific fine-tuning, addressing a persistent obstacle in heterogeneous urban
deployments.
[70]
Federated edge learning combines the two paradigms: FL clients execute on edge devices, and the global
aggregator may itself reside at the fog tier rather than in a central cloud. This hybrid architecture is increasingly
the default for smart-city implementations that must satisfy both latency and data-sovereignty constraints
simultaneously, and it is the architectural pattern most directly relevant to the case-study evidence reviewed in
Sections 4 through 6.
3.4IoT sensor networks across urban domains
3.4.1 Environmental Sensors
optical particle counters for PM2.5 and PM10 measurement, and Metal-Oxide-Semiconductor (MOS) sensors for
volatile organic compound detection.
[24]
Low-cost sensor networks supplement reference-grade monitoring
stations, providing high spatial resolution air quality mapping despite reduced accuracy. Sensor fusion
algorithms combine data from multiple sensor types to improve measurement reliability and compensate for
individual sensor limitations. Weather stations integrated into smart-city infrastructure measure temperature,
humidity, barometric pressure, wind speed, wind direction, precipitation, and solar radiation.
[71,72]
These
measurements support weather forecasting models, agricultural management systems, and energy demand
prediction algorithms. Acoustic sensors monitor urban noise pollution, identifying sources and enabling
enforcement of noise regulations. Water quality sensors deployed in distribution networks measure pH,
turbidity, conductivity, dissolved oxygen, and chlorine residual, detecting contamination events and verifying
treatment effectiveness.
[56,73]
3.4.2 Transportation sensors
Intelligent transportation systems utilize diverse sensing technologies for traffic monitoring and control.
Inductive loop detectors embedded in roadways detect vehicle presence and passage through electromagnetic
field disturbances. Magnetometer arrays provide similar functionality with easier installation and maintenance
characteristics. Radar sensors measure vehicle speed and occupancy across multiple lanes simultaneously.
[49]
LiDAR systems generate three-dimensional point clouds enabling precise vehicle classification and trajectory
tracking.
[74]
Computer vision cameras equipped with edge processing capabilities perform real-time vehicle detection,
classification, speed measurement, and license plate recognition.
[75]
Deep learning object detection algorithms
including YOLO (You Only Look Once) and Faster R-CNN achieve real-time performance suitable for traffic
monitoring applications. Pedestrian and cyclist detection systems enhance safety at intersections and
crosswalks through multimodal sensing combining cameras, thermal imaging, and pressure-sensitive
surfaces.
[76-78]
Connected vehicle sensors including GPS receivers, Inertial Measurement Units (IMUs), wheel speed sensors,
and On-Board Diagnostics (OBD) systems provide rich data for fleet management, usage-based insurance, and
collaborative traffic sensing.
[79]
Vehicle-to-Infrastructure (V2I) communication enables vehicles to report
position, speed, acceleration, and warning events to traffic management systems.
[80]
3.4.3 Building automation sensors
lighting sensors, and energy meters for comprehensive environmental control and energy management. Passive
Infrared (PIR) sensors detect occupant presence through thermal radiation, while ultrasonic and microwave
sensors monitor motion patterns.
[81]
Computer vision systems with privacy-preserving analytics provide
occupancy counting without storing identifiable images. Smart meters enable granular energy consumption
monitoring at appliance level, supporting demand response programs and energy efficiency initiatives. Non-
intrusive Load Monitoring (NILM) techniques disaggregate total energy consumption into individual appliance
contributions through analysis of electrical signatures. Building Information Modeling (BIM) integration links
physical sensors with digital twin models, enabling predictive maintenance and optimization.
[82]
3.4.4 Infrastructure monitoring sensors
Structural health monitoring systems deploy accelerometers, strain gauges, displacement sensors, and
corrosion sensors to assess bridge, tunnel, and building integrity.
[83]
Fiber optic sensors provide distributed
strain and temperature measurement along kilometers of infrastructure, detecting damage and fatigue. Acoustic
emission sensors identify crack propagation and material degradation in real-time.
[84]
Water distribution networks utilize pressure sensors, flow meters, and acoustic leak detection systems to
monitor performance and identify losses. Smart water meters enable consumption tracking, leak detection at
customer connections, and demand forecasting.
[6]
Wastewater systems employ level sensors, flow meters, and
chemical sensors for treatment process optimization and combined sewer overflow prevention.
[56]
3.5 Machine learning and deep learning frameworks
Machine learning tools for smart-city applications span three categories with distinct deployment
characteristics. Deep Learning Frameworks (TensorFlow, PyTorch) provide low-level computational
infrastructure for building, training, and deploying neural networks. High-Level APIs (Keras, now integrated as
tf.keras since TensorFlow 2.0) provide simplified interfaces enabling rapid model development. Classical
Machine Learning Libraries (scikit-learn, XGBoost) implement traditional algorithms suited to tabular data,
interpretable models, and scenarios with limited training data characteristics common in resource-constrained
municipal contexts.
[85]
For Edge AI deployment specifically, TensorFlow Lite, TensorFlow Lite Micro, ONNX
Runtime, and PyTorch Mobile dominate; for FL deployment, TensorFlow Federated, PySyft, Flower, and FedML
have emerged as the principal open-source frameworks. Tool selection requires matching framework
capabilities to deployment locality and learning paradigm, not framework loyalty.
[86]
Table 3 presents a
comparative analysis of machine learning tools for smart-city engineering.
3.6 Performance evaluation in distributed urban AI
Evaluation of Edge AI and FL deployments in smart-city engineering requires metrics beyond the standard
classification and regression measures (accuracy, precision, recall, F1, RMSE, MAE, R-squared, AUC) that apply
to centralized AI.
[69,90-92]
For Edge AI, additional engineering metrics include inference latency (typically reported
in milliseconds per inference at a target percentile), energy per inference (millijoules), memory footprint
(model size in kilobytes or megabytes plus runtime activation memory), and accuracy-versus-resource Pareto
frontiers under compression. For FL, evaluation extends to communication rounds to convergence, total bytes
transmitted, robustness under non-IID client data, performance under client dropout, and privacy budget
consumption when differential privacy is applied.
[93]
Reporting these distributed-system metrics alongside
conventional ML metrics is increasingly expected by reviewers in this domain; their absence in much of the
existing case-study literature is a reproducibility gap noted throughout Sections 4 through 6.
Table 3: Comparative analysis of ML/DL frameworks for smart-city deployment, extended with Edge AI and FL frameworks
Framework
Language
Strengths
Edge / FL support
Smart-city applications
TensorFlow
[87,88]
Python, C++
Production-ready, hardware
optimization
TF Lite for edge; TF
Federated for FL
Traffic prediction, air
quality, surveillance
PyTorch
[59]
Python
Research flexibility, dynamic
graphs
PyTorch Mobile, PySyft
(FL)
GNN traffic models, RL
agents
Keras (tf.keras)
[89]
Python
Rapid prototyping, simplified
API
Via TF Lite
POCs, transfer learning,
education
Scikit-learn
Python
Classical ML, well-documented
ONNX export for edge
Baselines, feature
engineering
XGBoost
C++, Python
Gradient boosting,
interpretability
ONNX, Treelite
Tabular prediction,
feature importance
Flower
Python
Framework-agnostic FL
orchestration
Native FL
Cross-silo and cross-
device FL
FedML
Python
Production FL platform
Native FL + Edge
Hierarchical federated
edge learning
4. Edge AI and federated learning in transportation engineering
Transportation engineering is the urban-AI domain where Edge AI delivers the most consistent latency benefit
and where FL is most actively explored as a privacy-preserving alternative to centralized fleet data pooling. This
section examines four implementation areas through the transversal lens: intelligent traffic management,
autonomous vehicle perception, mobility-as-a-service platforms, and predictive maintenance of transportation
infrastructure. Existing work largely organizes around these application sub-areas; the contribution here is to
expose where the Edge AI and FL deployment patterns recur and where they diverge.
4.1 Intelligent traffic management and signal control
Intelligent Transportation Systems integrate AI algorithms with IoT sensor networks to optimize traffic flow,
reduce congestion, and enhance safety.
[94]
Conventional centralized signal-control systems collect detector data
at a central traffic management center, compute timings, and push them back to controllers, incurring round-
trip latencies that limit responsiveness during fast-evolving demand.
[95]
Edge-deployed signal controllers execute reinforcement-learning policies locally, observing nearby detectors
and adjusting phasing on sub-second cycles.
[49,75]
Recent multi-agent reinforcement learning approaches at the
edge coordinate across intersections through limited peer-to-peer messaging, achieving substantial reductions
in vehicle delay relative to fixed-time and centralized adaptive control. Federated coordination across
intersections, neighborhoods, and even cities allows model improvement without exposing intersection-level
traffic counts that may reveal commercial activity or political-event traces. Spatio-temporal Graph Neural
Networks trained with continual-learning regularizers such as Elastic Weight Consolidation address the model-
drift problem characteristic of post-pandemic mobility shifts and seasonal demand change, allowing edge
controllers to adapt without catastrophic forgetting of prior regimes.
[96-98]
4.2 Autonomous vehicle perception and cooperative driving
Advanced driver assistance systems and autonomous vehicle technologies leverage DL models for real-time
object detection, path planning, and decision-making in complex urban environments.
[99]
Edge AI is non-
optional in this application: round-trip latency to a remote cloud is incompatible with the safety envelope of
vehicular perception, which requires inference within tens of milliseconds at worst-case conditions.
[95,100]
Vehicle-mounted edge accelerators run compressed CNN and transformer-based perception stacks; cooperative
driving extends single-vehicle perception through Vehicle-to-Vehicle and Vehicle-to-Infrastructure messaging
using DSRC or C-V2X.
[37,79]
Federated Learning across vehicle fleets enables manufacturers and operators to improve perception models
without centralizing dashcam footage or driving traces, addressing both bandwidth and privacy constraints.
[101]
Cross-silo FL between operators of municipal vehicle fleets (transit, sanitation, emergency services) is an
emerging pattern where each operator maintains sovereignty local data while contributing to a shared model
for urban hazard detection.
[102]
4.3 Mobility-as-a-Service and multimodal routing
Mobility-as-a-Service platforms integrate public transit, shared mobility, ride-hailing, and active modes into
unified booking and payment interfaces. AI components include demand forecasting, dynamic pricing,
multimodal routing, and equity-aware service-area design. Berlin's Jelbi platform and similar deployments
illustrate the integration challenge across operators with heterogeneous data and incentives.
[103]
Edge AI in MaaS appears in two patterns: on-device routing personalization that adapts to individual
preferences without uploading mobility traces, and edge-deployed demand prediction at modal hubs that
informs vehicle repositioning.
[104]
FL across operators allows cross-modal demand modeling without
commercially sensitive trip-record sharing, although standardization gaps remain a barrier.
[105]
4.4 Predictive maintenance of transportation infrastructure
Infrastructure-monitoring sensor networks generate continuous streams from accelerometers, strain gauges,
displacement sensors, and corrosion sensors on bridges, tunnels, and roadways.
[58,59]
Anomaly detection at the
sensor edge filters routine variation, transmitting only candidate events for review. This pattern is well-suited
to TinyML deployment because the underlying signal characteristics often admit relatively small models, and
battery or energy-harvesting power budgets favor on-device summarization over continuous cellular
transmission.
[106-108]
4.5 Case studies in smart transportation
Table 4 presents a structured comparison of transportation case studies drawn from verified technical
deployments and benchmarked studies, extended with dataset, algorithm, evaluation-metric, and key-finding
columns alongside the Edge AI and FL annotations. Continual-learning and edge-deployable architectures
dominate the corpus, reflecting the field's emphasis on handling non-stationary traffic patterns without full
retraining.
[109-111]
Federated approaches appear in two cases spatio-temporal traffic prediction and privacy-
preserving smart parking both framed as research contributions rather than production municipal
deployments, consistent with the maturity gap discussed in Section 7.
[112-114]
Table 4: Smart transportation case studies, annotated with Edge AI and FL deployment patterns
Year /
Author
Context
Algorithm /
Method
Dataset
Evaluation
Metric
Key Finding
Edge
AI?
FL?
2025 /
Francies
et al.
[109]
Highway
traffic
networks
(benchmark
)
STGCN with
continual-
learning
regularizers
PeMSD3,
PeMSD4,
PeMSD7,
PeMSD8
Prediction
error
(MAE/RMSE)
vs. baseline
STGCN
Continual-learning variants
cut catastrophic forgetting
and reduce error significantly
vs. baseline when traffic
patterns shift
Yes
No
2024 /
Kaur,
Grewal &
Jain
[112]
Wireless
sensor
network
traffic data
Fed-STGRU
(Federated
GRU + spatio-
temporal
graph)
Traffic sensor
network data
(WSN-based)
Accuracy, loss
vs.
TGCN/FedAvg/
FedGRU
baselines
FL approach matches
centralized-baseline accuracy
while keeping raw data local
No
Yes
2025 /
Du, Yang
& Liu
[113]
Smart
parking
systems
RDP-FedGAN
(Rényi-DP +
Federated
GAN)
Smart-
parking
sensor/occup
ancy data
Privacy-utility
trade-off under
Rényi-DP
GAN-synthesized data lets FL
train without sharing raw
location/dwell-time data
Yes
Yes
2021 /
Collini,
Nesi &
Pantaleo
[115]
Bike-sharing
stations
Deep learning
(short-term
prediction)
Real bike-
sharing
station
occupancy
logs
Prediction
accuracy, short-
term horizon
DL model gives short-term
available bike counts that are
usable for rebalancing
decisions
Partial
No
2021 /
Chen,
Wang &
Xie
[110]
Streaming
traffic
networks
TrafficStream:
GNN +
continual
learning
Streaming/ex
panding
traffic sensor
network
Forecast
accuracy under
network
expansion
Framework adapts to newly
added sensors/roads without
full retraining
Yes
No
2024 /
Chahal et
al.
[111]
IoT-enabled
smart city
Aquila
Optimizer-
based hybrid
predictive
model
IoT traffic-
congestion
sensor data
Predictive
accuracy vs.
baseline
optimizers
Hybrid metaheuristic-ML
model improves congestion
prediction over standalone
baselines
Yes
No
5. Edge AI and federated learning in environmental engineering
Environmental engineering exhibits a different Edge AI and FL profile from transportation. Edge AI is essential
where dense, low-power sensing must operate over years without manual intervention, particularly in air
quality and water-network monitoring
[116]
FL is most relevant where data sovereignty across municipal
boundaries, environmental regulators, and private operators prevents centralized pooling a recurring obstacle
in regional pollution modeling and watershed management.
[104]
5.1 Air quality monitoring and forecasting
Air quality monitoring networks deploy electrochemical sensors for pollutants and optical particle counters for
particulate matter
[24]
Reference-grade monitoring stations are sparse; low-cost networks supplement them at
higher spatial resolution but with calibration drift. Forecasting models include classical ML (Random Forest,
XGBoost) and deep architectures (LSTM, CNN-LSTM, Graph Neural Networks) that exploit spatial structure
among sensor locations.
[117-122]
Edge AI in air quality appears in two patterns. The first is on-device calibration adjustment: low-cost sensors
learn a local correction function against intermittent reference data and execute it on-device, addressing the
calibration-drift problem without continuous central recalibration. The second is edge-resident forecasting at
neighborhood gateways, supporting real-time public health alerts with sub-minute latency.
[123-126]
Federated
learning across municipal monitoring networks enables shared model improvement without exposing data the
regulator may not be authorized to share with neighboring jurisdictions. Uncertainty-guided sensor placement
frameworks combining stochastic physics-informed diffusion modeling with epistemic uncertainty
quantification address the upstream question of where edge sensors should be deployed to maximize
information gain under sparse coverage.
[127-129]
5.2 Smart waste management
Municipal solid waste management presents complex logistical, environmental, and economic challenges.
[130-133]
Smart waste systems deploy IoT sensors in containers measuring fill levels, temperature, tilt/motion, and
location. Ultrasonic sensors mounted inside container lids measure distance to waste surface, calculating fill
percentage in real-time. Weight sensors integrated into container bases provide complementary information.
Communication technologies transmit sensor data to central management platforms through LoRaWAN, NB-
IoT, or cellular networks. Dynamic route optimization algorithms determine optimal collection sequences
minimizing total travel distance, time, and fuel consumption.
[134-136]
Edge AI in waste management is largely confined to on-device fill-level estimation and tamper detection because
the analytic complexity per bin is modest. The more interesting transversal pattern is in robotic sorting at
material recovery facilities, where edge-deployed computer-vision classifiers running YOLO, Faster R-CNN, and
EfficientDet on conveyor-mounted cameras drive robotic pick-and-place actuators with sub-second
latency.
[134,137]
FL across regional waste-management authorities supports shared model improvement for
sorting and contamination detection without exposing operator-specific data.
5.3 Intelligent water management
Urban water systems face mounting pressures from population growth, aging infrastructure, climate change,
and increasing water scarcity.
[6,43,136,138]
Smart water systems integrate IoT sensors, advanced metering
infrastructure, and hydraulic modeling to enable comprehensive network monitoring and control. Pressure
sensors monitor hydraulic pressures identifying leaks, pipe breaks, and unauthorized connections through
anomalous pressure patterns. Flow meters measure water movement at treatment plant outlets, district
metering areas, and major transmission mains. Acoustic sensors detect leak sounds transmitted through pipe
walls, with correlation analysis of signals from multiple sensors pinpointing leak locations. Smart water meters
at customer connections provide granular consumption data enabling leak detection at property level.
[139]
Edge AI in water networks is most mature in acoustic leak localization, where on-device signal processing
identifies leak signatures locally and transmits only candidate event metadata, preserving the limited
communication budget of battery-operated nodes. LSTM-based demand forecasting and contamination
detection have been studied extensively; their migration to edge deployment is in progress but constrained by
the model-size and energy budgets of typical water-network nodes. FL across water utilities is an active research
area but limited in production by the regulatory and contractual complexity of inter-utility data sharing.
5.4 Climate monitoring and adaptation
Climate change poses escalating risks to urban areas through increased frequency and intensity of extreme
weather events, urban heat island effects, sea-level rise, and disruption of water resources.
[140-144]
Adaptation
strategies supported by AI technologies enable cities to enhance resilience, mitigate impacts, and protect
vulnerable populations. Urban Heat Island mapping uses CNN-based segmentation of satellite thermal imagery
(Landsat, MODIS, Sentinel) combined with ground sensor networks.
[145-147]
Flood prediction surrogate models
accelerate physics-based simulation by learning input-output mappings, enabling ensemble forecasting under
real-time constraints.
[148,149]
5.5 Case studies in smart environmental systems
Environmental monitoring and management case studies (Table 5) span federated, centralized deep-learning,
and non-ML empirical approaches, reflecting the heterogeneous data-governance and infrastructure conditions
described in Sections 5.1 through 5.4.
Federated learning is reserved for the one case where centralizing raw data is itself the obstacle: forest-fire
classification across a distributed network of independently owned surveillance cameras, where FL enables
shared model improvement without pooling footage across jurisdictions.
[150]
Centralized deep learning
dominates where a single institution controls the full data pipeline, as in illegal-dump-site detection from
imagery
[131]
and ensemble air-quality forecasting from a unified sensor network,
[24]
in both cases the barrier to
adoption is data quality and coverage rather than data sovereignty, so centralized training remains the simpler
and sufficient choice. Not every environmental case study in the literature is AI-driven at all: comparative policy
analysis of city-level energy planning
[151]
and empirical documentation of compounding heat and water-
insecurity risk among unhoused populations
[139]
remain valuable evidence for this domain without involving a
trained model, underscoring that Edge AI and FL are tools suited to specific sub-problems within environmental
engineering rather than a universal lens for the domain.
The federated case study also illustrates a governance pattern distinct from the centralized ones: because forest-
fire surveillance cameras are typically owned by different park authorities or private landholders, no single
actor controls enough data to train a useful, centralized model, making FL a practical necessity rather than a
privacy-driven design choice layered onto an otherwise centralizable problem
[150]
This distinction FL adopted
because centralization is infeasible, versus FL adopted because centralization is undesirable recurs across the
transportation and building domains as well and is revisited in Section 7.3.
Table 5: Smart environmental case studies with Edge AI and FL annotations
Year /
Author
Context
Algorithm /
Method
Dataset
Evaluation Metric
Key Finding
Edge
AI?
FL?
2024 /
Siddique,
Alasbali &
Driss
[150]
Forest-fire
monitoring
(Green
IoT)
FL framework
over camera-
equipped IoT
nodes
Forest
surveillance
camera
imagery
(distributed
clients)
Classification
accuracy under
FL vs. centralized
FL enables
collaborative fire-
classification training
across distributed
forest cameras without
centralizing footage
Yes
Yes
2022 /
Shahab &
Anjum
[131]
Municipal
solid
waste,
India
Deep learning
illegal-dump
detection
Dump-site
imagery
Detection
accuracy
DL approach flags
illegal dumping sites
from imagery for
targeted enforcement
Partial
No
2024 /
Addad &
Al-Taani
[151]
Amman,
Jordan
AI-driven
energy-
efficiency case
comparison
Amman
urban
energy/plan
ning data
Comparative
case-study
analysis
AI-assisted planning
shows measurable
energy-efficiency gains
transferable to
similarly resource-
constrained cities
No
No
2024 /
Sabir et
al.
[24]
Urban air-
quality
networks
Ensemble deep
learning
Air-quality
sensor
network data
Prediction
accuracy
(ensemble vs.
single models)
Ensemble DL improves
air-pollution forecast
accuracy over
individual base
learners
Partial
No
2023 / du
Bray et al.
[139]
Phoenix,
Arizona,
USA
Empirical/qualit
ative heat-
emergency
study
Field data,
unhoused-
population
heat
exposure
records
Descriptive/quali
tative findings
Documents
compounding heat and
water-insecurity risk
for unhoused
populations during
extreme heat
No
No
6. Edge AI and federated learning in smart buildings
Buildings account for approximately 40% of global energy consumption and 36% of energy-related carbon
dioxide emissions, representing the largest end-use sector for energy worldwide. In urban environments, this
proportion increases further, with commercial and residential buildings consuming 50-70% of city-wide energy
in developed nations
[121]
Edge AI and FL deliver three concrete benefits in this domain: latency reduction for
tight control loops in HVAC and lighting, energy and bandwidth reduction over continuous cloud telemetry from
thousands of devices per building, and privacy preservation of occupancy and behavior data that increasingly
carries regulatory weight.
[152]
6.1 Edge-deployed HVAC and energy management
Heating, Ventilation, and Air Conditioning systems are the largest building energy consumer and the primary
optimization target for smart-building AI.
[25,153]
Conventional rule-based control with fixed setpoints results in
substantial energy waste through excessive heating and cooling, poor equipment coordination, and degraded
performance from undetected faults.
Reinforcement-learning HVAC controllers learn policies that balance comfort and energy through direct
interaction with building dynamics. Edge deployment of these controllers eliminates the round-trip-to-cloud
latency that frustrates tight control loops and protects occupancy and behavior data from continuous cloud
transmission.
[38]
Model predictive control extensions execute the optimization on a building-resident edge
server.
[154]
6.2 Lighting and occupancy
Lighting accounts for a substantial share of building electricity consumption, typically higher in commercial
buildings than in residential buildings
[155]
Daylight harvesting adjusts electric lighting based on available natural
light. ML-based lighting control learns optimal strategies from historical occupancy patterns, daylight
availability, and user preferences. Combined occupancy and daylight-harvesting systems have been reported to
substantially reduce lighting energy consumption compared to manual control, with the magnitude of savings
varying by building type, glazing characteristics, and occupancy patterns. Edge-deployed occupancy classifiers,
including privacy-preserving
[149]
vision systems that emit only counts rather than images, are now widespread.
6.3 Federated learning across building portfolios
Federated Learning is increasingly applied across building portfolios held by single owners (commercial REITs,
university campuses, hospital systems) and across portfolios of independent owners participating in voluntary
efficiency programs
[157]
The principal motivations are sensitivity of occupancy and behavior data, contractual
constraints on cross-portfolio data sharing, and the structural advantage of cross-building learning to address
the heterogeneity that limits single-building model transfer. Mitra et al. demonstrated FL benefits for smart-
building applications including HVAC, occupancy, and energy-load forecasting
[52]
; subsequent work has
extended this to fault detection, demand response, and battery management. The principal obstacles to
production FL in buildings remain the non-IID character of inter-building data, the limited willingness of
building owners to participate in cross-organizational training, and the absence of standardized FL
orchestration frameworks suited to building-systems integrators.
[158]
6.4 Renewable integration and on-site storage
Building-integrated renewable energy systems including rooftop solar PV, building-integrated photovoltaics,
and micro-wind turbines reduce grid electricity consumption and associated emissions. Solar generation
forecasting employs sky-camera CNNs for short-horizon prediction and LSTM models trained on numerical
weather prediction outputs for longer-horizon prediction
[159]
Demand-side management adjusts building loads
to better align with renewable generation through pre-cooling, pre-heating, EV charging coordination, and
flexible-load deferral. Battery storage operation determines charging and discharging schedules balancing self-
consumption, arbitrage, grid services, and battery degradation. RL-based control has been reported to increase
solar self-consumption and reduce both grid electricity draw and time-of-use electricity costs, though reported
magnitudes vary substantially by climate, tariff structure, and battery sizing.
[167]
6.5 Case studies in smart buildings and living
Smart building case studies (Table 6) split cleanly along the Edge AI/FL axis: two studies directly instantiate the
transversal lens this review examines, while four apply centralized or cloud-based ML/DL to building-adjacent
health and localization tasks.
Table 6: Smart building case studies with Edge AI and FL annotations
Year / Author
Context
Algorithm /
Method
Dataset
Evaluation Metric
Key Finding
Edge
AI?
FL?
2021 / Mitra,
Ngoko &
Trystram
[52]
Smart-
building
management
(research)
[166]
Federated
personalizatio
n (distributed
learning)
Simulated
/in-situ
building
sensor
streams
Model quality
under
personalization vs.
baseline FL
Federated
personalization
improves in-situ
building-
management learning
while preserving
occupant data locality
Yes
Yes
2023 / Javed et
al.
[161]
Smart-
healthcare
CPS, fog layer
GBG-RPL:
Gini-index +
blockchain
Blackhole/Gre
yhole
detection
[167]
Simulated
CPS
network
traffic
(RPL
routing)
Detection accuracy,
attack mitigation
rate
Blockchain-enabled
Gini-index
framework improves
resilience of health-
monitoring CPS
against routing
attacks
Yes
No
2023 / Abirami
& Karthikeyan
[162]
Remote
elderly/patie
nt monitoring
Optimized
Fuzzy k-NN,
cloud digital
twin
[168]
DS1 &
DS2 voice
datasets
(Kaggle,
UCI ML
Repositor
Prediction
accuracy, F1-score
97.95% accuracy
(DS1) / 91.48%
(DS2); outperforms
NN and kernel-SVM
baselines for early
Parkinson's detection
No
No
Year / Author
Context
Algorithm /
Method
Dataset
Evaluation Metric
Key Finding
Edge
AI?
FL?
y)
[169]
2021 /
Mekruksavanich
& Jitpattanakul
[156]
Wearable-
sensor
activity
recognition
Deep CNN +
RNN hybrid
Wrist-
worn
wearable
sensor
data
Recognition
accuracy
Hybrid CNN-RNN
improves complex
activity-recognition
accuracy over single-
architecture
baselines
Yes
No
2023 / Kamal et
al.
[164]
Indoor
localization
(BLE)
XGBoost +
recursive
continuous
wavelet
transform,
explainable
RSSI
signal
data from
BLE
devices
Localization,
accuracy,
explainability
Explainable XGBoost
model achieves
accurate indoor
localization while
surfacing which RSSI
features drive
predictions
Yes
No
2023 / Fukuda
et al.
[165]
Personal
health
records,
Japan (SHINE
study)
Data-platform
integration for
longitudinal
health
monitoring
Personal
health
record
streams
Platform
adoption/monitori
ng outcomes
Integrated platform
supports continuous
personal-health-
record monitoring
across a national
ecosystem
[170]
Partial
No
Federated personalization across building portfolios
[52]
and fog-layer, blockchain-secured intrusion detection
for health-monitoring cyber-physical systems
[161]
are the two cases where Edge AI and FL jointly shape the
system architecture, consistent with Section 6.3's account of FL's motivations in this domain: cross-building
heterogeneity and the sensitivity of occupancy and health data. The remaining four case studies a cloud-
based digital-twin classifier for early Parkinson's detection,
[162]
a wearable-sensor activity-recognition model
[163]
, an explainable indoor-localization system for BLE devices,
[164]
and a national personal-health-record
monitoring platform
[165]
are centralized or cloud-resident systems that address building-adjacent health and
occupant-monitoring problems without a federated or edge-native architecture. Their inclusion here evidences
the breadth of ML activity in the building-health space rather than the maturity of Edge AI/FL adoption
specifically; as Section 6.3 notes, FL in buildings remains concentrated in research contributions such as
[52]
rather than production-grade, portfolio-wide deployment.
7. Cross-domain synthesis: A transversal taxonomy
7.1 A two-axis taxonomy of urban AI deployment
The taxonomy in Fig. 1 and the density matrix in Fig. 4 are populated from the full corpus of 207 included
studies, classified along two criteria: computation locality was assigned from the deployment location explicitly
described by each study's authors (on-device, edge gateway, fog, or cloud), and learning paradigm was assigned
from the stated training approach (centralized, federated, or hybrid edge-FL). Where a study did not explicitly
state deployment locality or evaluated multiple configurations, it was classified by its primary reported
configuration; studies that could not be unambiguously classified on either axis were excluded from Fig. 4, which
is why cell totals do not sum to exactly 207. Maturity labels in Fig. 3 (Production, Pilot, Research-only) were
assigned based on whether the source study described an operational municipal deployment (Production), a
time-limited or geographically limited trial (Pilot), or a simulation, testbed, or laboratory-only evaluation
(Research-only). This corpus-wide classification is distinct from the curated set of more than thirty case studies
presented in Tables 4 through 6 (Sections 4.5, 5.5, 6.5), which were selected as illustrative examples of the
patterns Figs. 3 and 4 characterize quantitatively across the full corpus, rather than as its complete enumeration.
This review proposes a transversal taxonomy that classifies smart-city AI implementations along two
orthogonal axes. The first axis is computation locality, with values ranging from on-device (TinyML,
microcontroller-class) through edge gateway (single-board computer, edge GPU), fog (district-level server), to
cloud (centralized datacenter) [171] The second axis is learning paradigm, with values ranging from centralized
(single-site training on pooled data), through federated (cross-site model training without data centralization,
including its horizontal, vertical, and hybrid variants), to hybrid edge-plus-FL architectures that combine on-
device inference with federated update flows. The taxonomy admits sixteen cells in principle; in practice, eight
cells dominate the case-study evidence, with the remainder representing immature or theoretical
configurations.
Fig. 3: Two-axis taxonomy of Edge AI and FL deployment in smart cities.
7.2 Cross-domain patterns by application
Cross-tabulating the case-study evidence against taxonomy reveals three consistent patterns. First,
transportation applications cluster heavily on the edge-gateway and on-device columns because of latency
constraints; FL adoption is concentrated in research pilots and cross-fleet manufacturer initiatives rather than
municipal production systems. Second, environmental applications populate the on-device cell densely due to
power constraints on field sensors, but the centralized-cloud column also remains heavily populated for
forecasting and policy use cases that tolerate batch latency. FL adoption is highest in cross-jurisdictional pilots
where centralized pooling is contractually or politically blocked.
[172]
Third, building applications populate the
edge-gateway cell densely because building-resident edge servers are common, and the federated row is the
most actively researched of the three domains for cross-portfolio learning. Fig. 4 visualizes density distribution.
Fig. 4: Method-by-domain density matrix.
7.3 Where Edge AI excels, where FL excels, where they combine
Across the three domains, Edge AI delivers the strongest gains where: (1) latency requirements are tight
(autonomous vehicle perception, signal control, HVAC fast loops); (2) bandwidth is constrained (battery-
operated environmental sensors, distributed leak detection); (3) energy per inference dominates lifecycle cost
(always-on monitoring); or (4) data residency or sovereignty constraints prohibit egress. FL delivers the
strongest gains where: (1) data sovereignty across jurisdictions or operators prevents centralized pooling
(cross-utility, cross-municipality, cross-vendor learning); (2) regulatory constraints (GDPR, sectoral health,
financial privacy) impose costs on centralized pooling; (3) heterogeneity across sources favors cross-site
learning without homogenization; or (4) commercial sensitivities block raw-data sharing. The hybrid pattern,
federated edge learning, is the architectural sweet spot where both pressures apply simultaneously and is the
most active research frontier in 2024-2026.
Fig. 5: Timeline of Edge AI and FL methods in smart cities, 2019-2026.
Fig. 6: Geographic distribution of reviewed case studies.
8. Implementation challenges and strategic solutions
Despite demonstrated benefits across transportation, environmental, and building engineering domains, smart-
city AI implementations face substantial barriers impeding widespread adoption and scaling. These challenges
span technical, organizational, economic, social, and regulatory dimensions, creating complex
interdependencies requiring coordinated solutions.
[173]
8.1 Technical infrastructure and integration
Legacy system integration challenges (proprietary protocols, limited connectivity) are addressed through
gateway devices and middleware platforms, as demonstrated in Singapore's integration of decades-old traffic
controllers with modern AI systems. Connectivity gaps are mitigated through hybrid communication
architectures combining 5G with LPWAN technologies. Stockholm's smart waste system exemplifies this
approach using LoRaWAN for sensor communication where cellular coverage is inadequate. Computational
resource constraints for Edge AI are addressed through model compression (quantization, pruning, knowledge
distillation) and hierarchical edge-cloud architectures, successfully deployed in Amsterdam's building
management systems.
[174]
These infrastructure constraints compound at scale in ways a single-deployment description understates.
Power provisioning is a binding constraint independent of computational capability: a TinyML sensor node's
multi-year battery life depends on inference cadence and radio duty cycle as much as on model size, and energy-
harvesting designs (solar, vibration) introduce their own intermittency that complicates continuous monitoring
commitments. Device lifecycle management firmware updates, model redeployment, and hardware
replacement across a fleet of potentially thousands of heterogeneous edge devices from multiple vendors is
rarely budgeted for in initial deployment costings yet determines whether a five-year-old sensor network still
runs a current, non-vulnerable model or has silently drifted into technical debt.
Network infrastructure heterogeneity is a second compounding factor specific to Edge AI and FL deployments.
5G coverage, LoRaWAN gateway density, and fiber backhaul availability vary sharply within a single city and far
more sharply across the international deployments compared in this review (Section 7), meaning an
architecture validated in one connectivity environment may not transfer to another without redesign. FL is
particularly exposed here: its communication-round overhead (Section 8.2) assumes a baseline of reliable
connectivity that intermittent or bandwidth-constrained municipal networks cannot always guarantee, and
client dropout under poor connectivity directly degrades convergence.
8.2 Edge AI and federated learning-specific limitations
Beyond the general infrastructure and data-management challenges above, Edge AI and FL carry limitations
specific to their distributed architecture that are frequently understated in deployment-oriented literature.
Communication overhead: FL replaces the bandwidth cost of centralizing raw data with the recurring cost of
transmitting model updates across potentially hundreds of communication rounds to reach convergence.
[175]
For
edge and mobile clients on metered or intermittent connections, this recurring cost can exceed the one-time
cost of raw-data transmission it was meant to avoid, particularly for large models; compression techniques such
as gradient quantization and sparsification (Section 3.2) mitigate but do not eliminate this overhead, and the
trade-off between compression aggressiveness and model accuracy is deployment-specific rather than solved
in general.
Privacy risks that persist despite FL. FL is frequently described as privacy-preserving because raw data never
leaves the client, but this characterization is incomplete: gradient updates themselves can leak information
about the underlying training data through model-inversion and membership-inference attacks.
[176]
Differential-
privacy noise injection and secure aggregation (Section 3.3) reduce this risk but introduce their own costs DP
noise degrades model utility, and secure aggregation protocols add computation and communication overhead
of their own
[156]
so privacy hardening in FL is a trade-off to be actively managed, not a property FL provides by
default.
Scalability: FL's convergence behavior degrades as client count grows into the thousands and as data
heterogeneity across clients increases; client dropout, stragglers, and non-IID data distributions the dominant
regime for cross-city, cross-building, and cross-utility urban FL (Section 3.3) slow convergence and can degrade
final model quality relative to centralized training on pooled data
[93]
Aggregation-server throughput and
coordination overhead impose a further, largely unaddressed scalability ceiling for city-scale or cross-
jurisdictional FL deployments.
[177]
Computational constraints: Edge AI's on-device inference budgets (Section 3.2) are demanding enough on their
own; FL clients face a strictly harder requirement, since local training (forward and backward passes over
multiple epochs) is substantially more computing and energy-intensive than inference alone. This asymmetry
means many TinyML-class devices capable of hosting a compressed model for inference cannot feasibly
participate as FL training clients, which in practice restricts urban FL to gateway-class or better hardware and
narrows the population of devices FL can draw on relative to Edge AI's inference-only footprint.
These four limitations are treated as open research problems rather than solved engineering challenges in
Section 9 (non-IID FL, Section 9.2; federated edge learning at scale, Section 9.5; TinyML constraints, Section 9.6);
the discussion here establishes why each currently limits production deployment, which Section 9 then frames
as a research agenda.
8.3 Data management and privacy
Data quality issues (noise, missing values, calibration drift) require automated outlier detection and cross-
sensor validation, implemented effectively in Busan's water monitoring system. Privacy concerns are addressed
through differential privacy, Federated Learning, and edge processing Helsinki's COVID-19 contact tracing
demonstrates privacy-preserving location data analysis. The federated edge learning architectural pattern
described in Section 3.3 directly addresses several of these concerns simultaneously but introduces its own
deployment complexity that municipalities frequently underestimate during procurement
[177]
Cybersecurity
threats require multi-layered security with network segmentation, encryption, and intrusion detection, as
deployed in Singapore's critical infrastructure systems.
[178]
8.4 Financial and organizational barriers
High upfront costs are mitigated through public-private partnerships, phased implementations starting with
high-ROI applications, and grant funding. Technical capacity gaps are addressed through university
partnerships, targeted recruitment, and staff training programs. Organizational resistance requires change
management: leadership commitment, participatory design involving affected staff, and celebrating quick wins
to build momentum.
8.5 Interoperability and standardization
Vendor lock-in risks are mitigated through open standards adoption (FIWARE, OGC SensorThings API) and
open-source platforms, as implemented in multiple European smart-city initiatives. Semantic interoperability
challenges require common data models and ontologies Singapore's data integration framework exemplifies
standardized metadata approaches enabling cross-system integration.
These standardization efforts address data-format interoperability but leave two Edge AI/FL-specific
interoperability problems largely unresolved. First, model-format interoperability across edge accelerators
remains fragmented: a model compressed and validated for one vendor's inference runtime (Section 3.2)
typically requires re-conversion and re-validation for another's, and no standard comparable to FIWARE's role
for data interoperability yet exists for edge-model portability across hardware vendors. Second, FL protocol
interoperability is largely absent: the FL frameworks compared in Table 3 (TensorFlow Federated, PySyft,
Flower, FedML) do not natively interoperate, so a cross-jurisdictional FL deployment spanning agencies that
have standardized on different frameworks requires custom bridging work that, to our knowledge, no included
study in this review's corpus reports having solved at production scale.
[179]
8.6 Stakeholder engagement and governance
Algorithmic Bias and Fairness: AI systems trained on historical data perpetuate and amplify existing societal
biases. Traffic prediction models trained on affluent neighborhood data may underperform in underserved
areas, leading to inequitable infrastructure investment. Facial recognition in surveillance systems exhibits
higher error rates for minority populations. Mitigation requires diverse and representative training datasets,
fairness metrics evaluating performance across demographic groups, bias auditing before deployment, and
ongoing monitoring for disparate impacts. Amsterdam's ethical AI framework and Singapore's bias assessment
protocols exemplify proactive approaches.
[180]
Accountability for Autonomous System Failures: As AI systems make increasingly autonomous decisions in
traffic signal control, emergency response routing, and utility management, accountability frameworks must
address who is responsible when autonomous systems fail, how liability is distributed for accidents involving
autonomous vehicles or failures in critical infrastructure control, and how human oversight is preserved over
algorithmic decision-making. Current legal frameworks inadequately address AI-specific accountability
challenges. Solutions require clear chains of responsibility documented in deployment agreements,
comprehensive logging and audit trails enabling failure analysis, human oversight mechanisms for high-stakes
decisions, and insurance and liability frameworks adapted to AI systems.
Digital Exclusion and Equity: Digital exclusion manifests in multiple dimensions. Technology design assuming
digital literacy and smartphone ownership excludes elderly, low-income, and disabled populations. Data-driven
service optimization may neglect areas generating less data, including underserved neighborhoods with lower
smartphone penetration. Automated decision systems may disadvantage groups underrepresented in training
data. Equity-by-design principles require involving affected communities in system design, multi-channel
service access (digital and non-digital), equity impact assessments evaluating distributional effects, and
proactive outreach to underserved populations.
8.7 Regulatory considerations
Regulatory treatment of Edge AI and FL in urban deployment is fragmented across data-protection, AI-specific,
and sectoral regulatory regimes, and the review's inclusion criteria (Section 2.3) mean this fragmentation is
itself under-documented in the peer-reviewed deployment literature we could draw on most included studies
report technical outcomes without discussing the regulatory approval pathway that preceded deployment.
Where FL is deployed as a data-protection mitigation (Section 3.3), its regulatory standing is not uniformly
settled: GDPR and comparable regimes were drafted around a paradigm of data controllers and processors
handling centralized data, and whether model
regimes remains an active legal question rather than a resolved one, notwithstanding FL's common
characterization as inherently privacy-preserving (Section 8.2).
[50]
Emerging AI-specific regulation illustrates
this uncertainty concretely: the EU AI Act's Annex III explicitly designates AI systems used as safety components
in critical digital infrastructure, road traffic, and water, gas, heating, or electricity supply directly covering
several applications reviewed here (Sections 46) as high-risk, but the compliance deadline for these
and the Commission's interpretive guidance on classification remains in draft form as of this writing.
[50]
A
deployment architecture validated against today's understanding of these requirements may need re-validation
against the finalized guidance before the obligations take effect. Sectoral regulation compounds this further:
health-adjacent building deployments (Section 6, e.g.
[161,162]
) face medical-device or health-data regulation
layered on top of general data protection, while autonomous-vehicle perception (Section 4.2) faces vehicle-
safety certification regimes that vary by jurisdiction and are not harmonized internationally.
Procurement regulation is a further, largely unaddressed constraint: public-sector procurement rules in many
jurisdictions were not designed for AI systems whose behavior evolves post-deployment (through continual
learning, Section 3.2, or through FL's ongoing model updates), creating a mismatch between static procurement
specifications and dynamically updating systems that Section 9.1's open problem on model drift only partially
captures from a technical rather than a contractual angle. We treat this as an open problem alongside those in
Section 9 rather than a solved constraint: no study in our corpus reports a resolved procurement framework for
continually updating municipal AI systems.
[181]
9. Research agenda: Open problems in Edge AI and federated learning for smart cities
The synthesis presented across Sections 4 through 7 identifies eight open problems whose resolution
conditions further maturation of Edge AI and FL in smart-city engineering. The problems are framed as
falsifiable propositions amenable to empirical and methodological investigation, in contrast to the open-ended
research themes typical of review conclusions.
9.1 Model drift under non-stationary urban data
Urban environments exhibit non-stationary distributions over weather, behavior, and infrastructure that
degrade deployed ML models over months and years. Continual-learning and elastic-weight-consolidation
methods mitigate but do not eliminate this. Open question: what monitoring and adaptation cadence preserves
accuracy at acceptable computational cost across the three engineering domains?
9.2 Non-IID Federated Learning in heterogeneous urban deployments
FL convergence and final accuracy degrade under non-IID client data, the dominant regime for cross-city, cross-
building, and cross-utility FL. Personalization, clustering, and meta-learning approaches show promise but lack
standardized evaluation. Open question: what FL personalization strategy and how much local adaptation
maximizes utility under the non-IID profiles characteristic of urban deployments?
9.3 Energy accounting of urban AI
Sustainability claims for smart-city AI rarely account for the embodied and operational energy of the AI itself.
As Edge AI proliferates, the aggregate energy and carbon footprint of millions of always-on inference devices
becomes non-negligible. Open question: under what deployment density and inference cadence does Edge AI
become net energy-positive against the urban systems it manages?
9.4 Standardized fairness and explainability for distributed urban AI
Bias detection and explanation tooling for centralized ML does not transfer cleanly to FL and Edge AI. Fairness
metrics computed on local data may diverge from those computed on the implicit global distribution;
explanations generated at the edge may not match those expected by city operators reviewing centralized
aggregates. Recent work on multi-layer explainable AI frameworks for urban budget allocation suggests one
path, but cross-domain standards remain absent. Open question: what fairness and explainability
instrumentation should be required by procurement standards for municipal AI?
[182]
9.5 Federated edge learning at scale
Federated edge learning combines the latency benefits of Edge AI with the data-sovereignty benefits of FL, but
the architecture is largely empirically validated rather than systematically characterized. Open question: at what
client count and communication-cost regime does federated edge learning outperform centralized cloud
learning on the engineering objectives typical of municipal deployments?
9.6 TinyML for ultra-constrained urban sensing
TinyML enables intelligent sensors operating for years on battery power with kilobytes of RAM, but the model-
class restrictions are severe. The frontier is determining which urban sensing tasks reduce to TinyML-feasible
model classes without unacceptable accuracy loss, and which require gateway-class or fog escalation.
9.7 Cross-domain transfer and shared representation learning
Smart-city domains share underlying spatiotemporal structure that current single-domain models do not
exploit. Foundation models pre-trained on diverse urban data may enable cross-domain transfer, but the corpus
curation, pre-training objectives, and evaluation protocols remain unresolved.
[183]
9.8 Sociotechnical integration
Technical sophistication does not guarantee deployment success. Open problems include participatory design
for affected communities, multi-channel service access combining digital and non-digital pathways, equity
impact assessments, and proactive outreach to underserved populations. Section 8.5 introduced these themes;
advancing them requires interdisciplinary collaboration between AI researchers, urban planners, and policy
scholars that current institutional structures do not adequately support.
10. Limitations of this review
This review is subject to several limitations that should inform how its findings are used. First, as noted in
Section 1.3, this work adopts a narrative engineering-review approach rather than a systematic review or meta-
analysis: reported outcomes are synthesized qualitatively rather than pooled through formal statistical
aggregation, and readers should not interpret the ranges reported in Sections 4 through 6 and in the Conclusion
as pooled effect sizes.
Second, screening and full-text eligibility assessment (Section 2.4) were conducted by a single reviewer. No
second independent screener was used, and no inter-rater agreement statistic (e.g., Cohen's kappa) was
computed. This departs from PRISMA good practice for systematic reviews and increases the risk of inclusion
and classification error relative to dual-reviewer protocols.
Third, the eligibility criteria (Section 2.3) restricted the corpus to peer-reviewed, English-language, full-text-
retrievable studies, and grey literature was excluded except for a small set of contextual UN, World Bank, and
municipal reports. These restrictions plausibly under-represent deployments documented only in non-English
venues, in vendor or municipal grey literature, or in outlets outside the five searched databases, and may bias
the corpus toward jurisdictions and institutions with stronger English-language peer-reviewed publication
practices.
Fourth, the search window (January 2021 to February 2026) was chosen to capture the post-2020 transversal
turn toward Edge AI and FL (Section 1.4), but Edge AI and FL methods continue to evolve rapidly; findings
synthesized here should be read as a snapshot of the literature as of the search date rather than a permanently
current account.
Fifth, the case-study tables in Sections 4.5, 5.5, and 6.5 present a curated subset of "more than thirty" case
studies drawn from the 207-study corpus, selected to illustrate transversal Edge AI and FL deployment patterns
rather than to exhaustively or representatively summarize the full corpus. The quantitative outcomes reported
within these case studies and in the Conclusion (e.g., energy-saving percentages) are as reported by the original
authors and were not independently re-verified or re-analyzed by this review; they may be subject to
publication bias toward favorable results.
Sixth, the six comparator surveys used to position this review's contribution (Section 2.5) were selected
purposively to span the range from broad AI-in-smart-cities reviews to paradigm-focused FL and Edge AI
surveys, rather than identified through a systematic search of comparator reviews. A differently constructed
comparator set could yield a different positioning.
Finally, classification of case studies into the cells of the two-axis taxonomy (Section 7.1, Figs. 3 and 4) required
subjective judgment regarding a study's dominant computation locality and learning paradigm where a
deployment spanned multiple tiers or paradigms. This classification was not independently validated by a
second reviewer, and borderline cases may be classified differently on reanalysis.
11. Conclusion
This review synthesized over two hundred peer-reviewed studies on Edge AI and Federated Learning in smart-
city engineering across transportation, environmental, and building domains, applying a PRISMA 2020-
compliant search and selection protocol with documented inclusion criteria. Four contributions were delivered:
a transversal two-axis taxonomy classifying urban AI by computation locality and learning paradigm; a five-
layer reference architecture extended with explicit Edge AI and FL placement annotations; a structured cross-
domain synthesis of more than thirty international case studies; and an evidence-based research agenda
framing eight open problems as falsifiable propositions. The synthesis supports three quantified findings. First,
Edge AI delivers consistent latency, energy, and bandwidth benefits across transportation and building
applications, with consistent reports of measurable energy reduction in lighting and HVAC control, and sub-
second inference latencies for vehicular perception and signal control, but with deployment maturity differing
across domains by an order of magnitude. Second, Federated Learning is most impactful where data sovereignty,
regulatory constraint, or cross-jurisdictional coordination dominates the problem geometry particularly in
cross-utility environmental monitoring and cross-portfolio building energy modeling, but production
deployments meaningfully lag research pilots, consistent with the Pilot- and Research-only-dominated cells of
Fig. 3. Third, the federated edge learning hybrid is the architectural pattern most likely to dominate the next
deployment generation where both latency and sovereignty pressures apply, and the eight open problems in
Section 9 condition its maturation. For city governments and urban planners, the practical implication is that
procurement and architectural decisions should explicitly position each deployment in the proposed taxonomy,
treating computation locality and learning paradigm as design parameters rather than vendor defaults. For
technology vendors and system integrators, the implication is that Edge AI and FL frameworks must be
supported as first-class deployment options with documented privacy, energy, and adaptation guarantees. For
researchers, the agenda in Section 9 identifies eight problems where field studies and methodological work
would directly affect deployment readiness. The smart-city domain is iterative and contextual; the taxonomy
and reference architecture proposed here are intended as scaffolding for that iteration, not as a fixed
prescription.
Declaration
[1]
International Journal of Information Management Data Insights
[2]
N. Kapoor, N. Ahmad, S. K. Nayak, S. P. Singh, P. V. Ilavarasan, P. Ramamoorthy, Identifying
infrastructural gap areas for smart and sustainable tribal village development: A data science
approach from India, International Journal of Information Management Data Insights, 2021, 1,
100041, doi: 10.1016/j.jjimei.2021.100041.
[3]
E. Manibardo, I. Laña, J. Del Ser, Deep learning for road traffic forecasting: Does it make a
difference?, IEEE Transactions on Intelligent Transportation Systems, 2021, 7, 6164-6188, doi:
10.1109/TITS.2021.3083957.
[4]
Z. S. Klos, J. Kalkowska, J. Kasprzak, Towards a Sustainable Future - Life Cycle Management, Springer
International Publishing, 2022, doi: 10.1007/978-3-030-77127-0.
[5]
M. Elsisi, M.-Q. Tran, K. Mahmoud, M. Lehtonen, M. M. F. Darwish, Deep Learning-Based Industry 4.0
and Internet of Things towards Effective Energy Management for Smart Buildings, Sensors, 2021, 21,
1038, doi: 10.3390/s21041038.
[6]
P. Mukuyu, S. Warner, D. V. Chapman, N. Jayathilake, C. Dickens, J. Mateo-Sagasta, Innovations in water
quality monitoring and management in Africa: Towards developing an African Water Quality Program
(AWaQ), Colombo, Sri Lanka, 2024, doi: 10.5337/2023.217.
[7]
L. Z. Ghomari, Artificial intelligence in intelligent transportation systems, Journal of Intelligent
Management Systems and Engineering, 2025, 6, 2642, doi: 10.1108/JIMSE-11-2024-0035.
[8]
O. Farooq, A. Khan, The Impact of Machine Learning on Climate Change Modeling and Environmental
Sustainability, Artificial Intelligence and Machine Learning Research, 2025, 6.
[9]
P. Antwi-Afari, D.-G. Owusu-Manu, S. T. Ng, G. Asumadu, Modeling the smartness or smart
Journal of Urban Management, 2021, 10, 369381,
doi: 10.1016/j.jum.2021.06.005.
[10]
P. Bellini, P. Nesi, G. Pantaleo, IoT-enabled smart cities: A review of concepts, frameworks and key
technologies, Applied Sciences, 2022, 12, 1607, doi: 10.3390/app12031607.
[11]
M. Bauer, L. Sanchez, J. Song, IoT-enabled smart cities: Evolution and outlook, Sensors, 2021, 21, 4511,
doi: 10.3390/s21134511.
[12]
I. H. Sarker, Machine learning: Algorithms, real-world applications and research directions, SN
Computer Science, 2021, 2, 160, doi: 10.1007/s42979-021-00592-x.
[13]
I. H. Sarker, Deep learning: A comprehensive overview on techniques, taxonomy, applications and
research directions, SN Computer Science, 2021, 2, 420, doi: 10.1007/s42979-021-00815-1.
[14]
P. Li, H. Zhang, Y. Wu, L. Qian, R. Yu, D. Niyato, X. Shen, Filling the missing: exploring generative ai for
enhanced federated learning over heterogeneous mobile edge devices, IEEE Transactions on Mobile
Computing, 2024, 23, 1000110015, doi: 10.1109/TMC.2024.3371772.
[15]
H. Nguyen, D. Nawara, R. Kashef, Connecting the indispensable roles of IoT and artificial intelligence
in smart cities: A survey, Journal of Information and Intelligence, 2024, 2, 261285, doi:
10.1016/j.jiixd.2024.01.003.
[16]
I. H. Sarker, M. M. Hoque, Md. K. Uddin, T. Alsanoosy, Mobile data science and intelligent apps:
concepts, ai-based modeling and research directions, Mobile Networks and Applications, 2021, 26,
285303, doi: 10.1007/s11036-020-01650-z.
[17]
D. Luckey, H. Fritz, D. Legatiuk, K. Dragos, K. Smarsly, Artificial Intelligence Techniques for Smart City
Applications, Proceedings of the 18th International Conference on Computing in Civil and Building
Engineering, Springer, 2021, 315, doi: 10.1007/978-3-030-51295-8_1.
[18]
D. Correia, J. L. Marques, L. Teixeira, The state-of-the-art of smart cities in the European union, Smart
Cities, 2022, 5, 17761810, doi: 10.3390/smartcities5040089.
[19]
T. Robinson, M. Ji, Smart City Seoul: Solving the Urban Puzzle. In: Sustainable, Smart and Solidary
Seoul. Sustainable Development Goals Series. Springer, Cham. 2022, doi: 10.1007/978-3-031-13595-
8_5.
[20]
I. Dawar, N. Kumar, Making smart cities smarter, Smart Cities and Artificial Intelligence, 2023, 240
262, doi: 10.4018/979-8-3693-0892-9.ch012.
[21]
Sensors
[22]
[23]
U. Rajasekaran, A. Malini, M. Murugan, Artificial intelligence in autonomous vehicles—a survey of
trends and challenges, in: Artificial Intelligence for Autonomous Vehicles, (ed) S. Rajendran, M.
Sabharwal, Y-C. Hu, R. K.Dhanaraj, B. Balusamy, Wiley, 2024, 124, doi:
10.1002/9781119847656.ch1.
[24]
Computer Systems Science and Engineering
[25]
Energies
[26]
Connection Science
[27]
H. M. K. K. M. B. Herath, G. M. K. B. Karunasena, H. M. W. T. Herath, Development of an IoT Based
systems to mitigate the impact of COVID-19 pandemic in smart cities, IoT and Smart City Applications,
2021, 287309, doi: 10.1007/978-3-030-72065-0_16.
[28]
I. A. T. Hashem, R. S. A. Usmani, M. S. Almutairi, A. O. Ibrahim, A. Zakari, F. Alotaibi, S. M. Alhashmi, H.
Chiroma, Urban computing for sustainable smart cities: Recent advances, taxonomy, and open
research challenges, Sustainability, 2023, 15, 3916, doi: 10.3390/su15053916.
[29]
T. V. Hoang, Impact of integrated artificial intelligence and internet of things technologies on smart
city transformation, Journal of Technical Education Science, 2024, 19, 6473, doi:
10.54644/jte.2024.1532.
[30]
O. El Ghati, O. Alaoui-Fdili, O. Chahbouni, N. Alioua, W. Bouarifi, Artificial intelligence-powered visual
internet of things in smart cities: A comprehensive review, Sustainable Computing: Informatics and
Systems, 2024, 43, 101004, doi: 10.1016/j.suscom.2024.101004.
[31]
H. B. McMahan, E. Moore, D. Ramage, S. Hampson, B. A. y Arcas, Communication-efficient learning of
deep networks from decentralized data, Proceedings of the 20th International Conference on
Artificial Intelligence and Statistics, 2017, 12731282.
[32]
Z. Geng, W. Zhang, A Meta-Learning Framework Integrating Multi-Source Spatio-Temporal Data for
Adaptive and Accurate Cross-City Traffic Flow Prediction, Journal of Cities & Infrastructure, 2025, doi:
10.64229/n7hrn835.
[33]
S. S. Gill, M. Golec, J. Hu, M. Xu, J. Du, H. Wu, G. K. Walia, S. S. Murugesan, B. Ali, M. Kumar, K. Ye, P. Verma,
S. Kumar, F. Cuadrado, S. Uhlig, Edge AI: A taxonomy, systematic review and future directions, Cluster
Computing, 2025, 28, 18, doi: 10.1007/s10586-024-04686-y.
[34]
A. Alvarez, C. Mora, M. Mutzu Martis, C. Garau, The downgrade of leading smart city,
Gervasi, B. Murgante, C. Garau, D. Taniar,
A.M.A. Rocha, M. N. Faginas Lago, (eds) Computational Science and Its Applications ICCSA 2024
Workshops. ICCSA 2024. Lecture Notes in Computer Science, Springer, Cham., 2024, 14823,
10.1007/978-3-031-65329-2_6.
[35]
Proceedings
of the IEEE
[36]
A. S. Syed, D. Sierra-Sosa, A. Kumar, A. Elmaghraby, IoT in smart cities: A survey of technologies,
practices and challenges, Smart Cities, 2021, 4, 429475, doi: 10.3390/smartcities4020024.
[37]
C. Englund, E. E. Aksoy, F. Alonso-Fernandez, M. D. Cooney, S. Pashami, B. Åstrand, AI perspectives in
smart cities and communities to enable road vehicle automation and smart traffic control, Smart
Cities, 2021, 4, 783802, doi: 10.3390/smartcities4020040.
[38]
M. A. Ramírez-Moreno, S. Keshtkar, D. A. Padilla-Reyes, E. Ramos-López, M. García-Martínez, M. C.
Hernández-Luna, A. E. Mogro, J. Mahlknecht, J. I. Huertas, R. E. Peimbert-García, R. A. Ramírez-
Mendoza, A. M. Mangini, M. Roccotelli, B. L. Pérez-Henríquez, S. C. Mukhopadhyay, J. de Jesús Lozoya-
Santos, Sensors for sustainable smart cities: A review, Applied Sciences, 2021, 11, 8198, doi:
10.3390/app11178198.
[39]
S. Cui, Y. Fu, H. Fu, W. Shen, Edge-to-cloud computing and intelligence for IoT-based Structural Health
Monitoring: A comprehensive review, Advanced Engineering Informatics, 2026, 71, 104300, doi:
10.1016/j.aei.2025.104300.
[40]
L. Liu, X. Guo, C. Lee, Promoting smart cities into the 5G era with multi-field Internet of Things (IoT)
applications powered with advanced mechanical energy harvesters, Nano Energy, 2021, 88, 106304,
doi: 10.1016/j.nanoen.2021.106304.
[41]
I. Shahrour, X. Xie, Role of internet of things (IoT) and crowdsourcing in smart city projects, Smart
Cities, 2021, 4, 12761292, doi: 10.3390/smartcities4040068.
[42]
L. Pioli, D. D. J. de Macedo, D. G. Costa, M. A. R. Dantas, Towards an AI-Driven Data Reduction
Framework for Smart City Applications, Sensors, 2024, 24, 358, doi: 10.3390/s24020358.
[43]
W. Cao, J. Zhang, C. Cai, Q. Chen, Y. Zhao, Y. Lou, W. Jiang, G. Gui, CNN-based intelligent safety
surveillance in green IoT applications, China Communications, 2021, 18, 108119, doi:
10.23919/JCC.2021.01.010.
[44]
R. Wang, C. Xu, R. Dong, Z. Luo, R. Zheng, X. Zhang, A secured big-data sharing platform for materials
genome engineering: State-of-the-art, challenges and architecture, Future Generation Computer
Systems, 2023, 142, 5974, doi: 10.1016/j.future.2022.12.026.
[45]
I. H. Sarker, AI-based modeling: Techniques, applications and research issues towards automation,
intelligent and smart systems, SN Computer Science, 2022, 3, 158, doi: 10.1007/s42979-022-01043-
x.
[46]
X. He, Q. Ai, J. Wang, F. Tao, B. Pan, R. C. Qiu, B. Yang, Situation awareness of energy internet of things
in smart city based on digital twin: from digitization to informatization, IEEE Internet of Things
Journal, 2023, 10, 74397458, doi: 10.1109/JIOT.2022.3203823
[47]
M. Paricherla, S. Babu, K. Phasinam, H. Pallathadka, A. S. Zamani, V. Narayan, S. K. Shukla, H. S.
Mohammed, Towards development of machine learning framework for enhancing security in internet
of things, Security and Communication Networks, 2022, 4477507, doi: 10.1155/2022/4477507.
[48]
H. A Tarish, Anomaly detection improvement in computer communication networks using machine
learning techniques, Journal of Cybersecurity and Information Management, 2025, 15, 1, doi:
10.54216/JCIM.150120.
[49]
Z.-G. Chen, Z.-H. Zhan, S. Kwong, J. Zhang, Evolutionary computation for intelligent transportation in
smart cities: A survey, IEEE Computational Intelligence Magazine, 2022, 17, 83102, doi:
10.1109/MCI.2022.3155330.
[50]
B. Fabregue, Artificial intelligence governance in smart cities: A European regulatory perspective,
Journal of Autonomous Intelligence, 2024, 7, 672, doi: 10.32629/jai.v7i2.672.
[51]
Z. Ullah, M. Naeem, A. Coronato, P. Ribino, G. De Pietro, Blockchain applications in sustainable smart
cities, Sustainable Cities and Society, 2023, 97, 104697, doi: 10.1016/j.scs.2023.104697.
[52]
A. Mitra, Y. Ngoko, D. Trystram, Impact of federated learning on smart buildings, 2021 International
Conference on Artificial Intelligence and Smart Systems (ICAIS), IEEE, 2021, 9399, doi:
10.1109/ICAIS50930.2021.9395938.
[53]
A. R. Jafari, V. González, L. Martín, L. Sánchez, J. Lanza, S. M. Raza, M. Alvi, K. Kaewnoparat, R. Minerva,
N. Crespi, Data enrichment toolchain: A use-case for correlation analysis of air quality, traffic, and
y, Internet of Things, 2024, 26, 101232, doi:
10.1016/j.iot.2024.101232.
[54]
M. Saleem, S. Abbas, T. M. Ghazal, M. A. Khan, N. Sahawneh, M. Ahmad, Smart cities: Fusion-based
intelligent traffic congestion control system for vehicular networks using machine learning
techniques, Egyptian Informatics Journal, 2022, 23, 417426, doi: 10.1016/j.eij.2022.03.003.
[55]
Y. Sun, X. Zhao, R. Lovreglio, E. Kuligowski, AI for large-scale evacuation modeling: Promises and
challenges, in: Interpretable Machine Learning for the Analysis, Design, Assessment, and Informed
Decision Making for Civil Infrastructure, Woodhead Publishing Series in Civil and Structural
Engineering, 2024, 185204, doi: 10.1016/B978-0-12-824073-1.00014-9.
[56]
A. J. Zapata-Sierra, E. Salmerón-Manzano, A. Alcayde, M. L. Zapata-Castillo, F. Manzano-Agugliaro, The
scientific landscape of smart water meters: A comprehensive review, Water, 2024, 16, 113, doi:
10.3390/w16010113.
[57]
S. A. A. Bokhari, S. Myeong, Artificial intelligence-based technological-oriented knowledge
management, innovation, and e-service delivery in smart cities: Moderating role of e-governance,
Applied Sciences, 2022, 12, 8732, doi: 10.3390/app12178732.
[58]
O. Aouedi, T.-H. Vu, A. Sacco, D. C. Nguyen, K. Piamrat, G. Marchetto, Q.-V. Pham, A survey on intelligent
internet of things: Applications, security, privacy, and future directions, IEEE Communications Surveys
& Tutorials, 2025, 27, 12381292, doi: 10.1109/COMST.2024.3430368.
[59]
J. Ansel, E. Yang, H. He, N. Gimelshein, A. Jain, M. Voznesensky, B. Bao, P. Bell, D. Berard, E. Burovski, G.
Chauhan, A. Chourdia, W. Constable, A. Desmaison, Z. DeVito, E. Ellison, W. Feng, J. Gong, M. Gschwind,
B. Hirsh, S. Huang, K. Kalambarkar, L. Kirsch, M. Lazos, M. Lezcano, Y. Liang, J. Liang, Y. Lu, C. K. Luk, B.
Maher, Y. Pan, C. Puhrsch, M. Reso, M. Saroufim, M. Y. Siraichi, H. Suk, S. Zhang, M. Suo, P. Tillet, X. Zhao,
E. Wang, K. Zhou, R. Zou, X. Wang, A. Mathews, W. Wen, G. Chanan, P. Wu, S. Chintala, PyTorch 2: Faster
Machine Learning Through Dynamic Python Bytecode Transformation and Graph Compilation,
Proceedings of the 29th ACM International Conference on Architectural Support for Programming
Languages and Operating Systems, ACM, 2024, 2, 929947, doi: 10.1145/3620665.3640366.
[60]
Adoption: A Review of the Literature, Journal of Open Innovation: Technology, Market, and Complexity,
2021, 7, 146, doi: 10.3390/joitmc7020146.
[61]
city and sustainable energy—evidence from the European union
capital cities, Energies, 2024, 17, 4678, doi: 10.3390/en17184678.
[62]
prospects of energy in smart grids and smart cities systems, Energy Conversion and Management: X,
2025, 27, 101162, doi: 10.1016/j.ecmx.2025.101162.
[63]
C. R. Banbury, V. Janapa Reddi, M. Lam, W. Fu, A. Fazel, J. Holleman, X. Huang, R. Hurtado, D. Kanter, A.
Lokhmotov, D. Patterson, D. Pau, J.-S. Seo, J. Sieracki, U. Thakker, M. Verhelst, P. Yadav, Benchmarking
TinyML Systems: Challenges and Direction, arXiv:2003.04821, doi: 10.48550/arXiv.2003.04821.
[64]
F. Trindade Neves, M. Aparicio, M. de Castro Neto, The impacts of open data and explainable AI on
real estate price predictions in smart cities, Applied Sciences, 2024, 14, 2209, doi:
10.3390/app14052209.
[65]
R. David, J. Duke, A. Jain, V. Janapa Reddi, N. Jeffries, J. Li, N. Kreeger, I. Nappier, M. Natraj, T. Wang, P.
Warden, R. Rhodes, TensorFlow Lite Micro: Embedded machine learning for TinyML systems,
Proceedings of Machine Learning and Systems, 2021, 3, 800811.
[66]
H. Peng, C. Ding, T. Geng, S. Choudhury, K. Barker, A. Li, Evaluating emerging AI/ML accelerators: IPU,
RDU, and NVIDIA/AMD GPUs, Companion of the 15th ACM/SPEC International Conference on
Performance Engineering, ACM, 2024, 1420, doi: 10.1145/3629527.3651428.
[67]
P. Kumar, Time series traffic flow prediction with hyper-parameter optimized ARIMA models for
intelligent transportation system, Journal of Scientific & Industrial Research, 2022, 81, 408-415,
Accessed 13 January 2026.
[68]
Q. Yang, Y. Liu, T. Chen, Y. Tong, Federated machine learning: Concept and applications, ACM
Transactions on Intelligent Systems and Technology, 2019, 10, 12:112:19, doi: 10.1145/3298981.
[69]
A. B. M. S. Rahman, M. Lee, S. Venkatesan, J. Lim, C. Shin, A Comparative Study Between Linear
Regression and Support Vector Regression Model Based on Environmental Factors of a Smart Bee
Farm, Smart Media Journal, 2022, 11, 3847, doi: 10.30693/SMJ.2022.11.5.38.
[70]
Z. Yang, S. Zhang, Resource-efficient heterogenous federated continual learning on edge, 2024 Design,
Automation & Test in Europe Conference & Exhibition (DATE), IEEE, 2024, 16, doi:
10.23919/DATE58400.2024.10546889.
[71]
H. C. Nguyen, M. Liu, Short-term solar irradiance forecasting from future sky images generation, AI
2023: Advances in Artificial Intelligence, Springer, 2023, 1527, doi: 10.1007/978-981-99-8388-9_2.
[72]
M. W.-E., Geospatial environmental data for planetary health applications, Geospatial Technologies for
Environmental Health, Springer, 2023, 123141, doi: 10.1007/978-981-19-8765-6_7.
[73]
R. M. Alsina-Pagès, G. J. Ginovart-Panisello, M. Freixes, A. Radicchi, A Soundwalk in the heart of
Poblenou superblock in Barcelona: Preliminary study of the acoustic events, Noise Mapping, 2021, 8,
207216, doi: 10.1515/noise-2021-0016.
[74]
S. I. Imtiaz, S. ur Rehman, A. R. Javed, Z. Jalil, X. Liu, W. S. Alnumay, DeepAMD: Detection and
identification of Android malware using high-efficient deep artificial neural network, Future
Generation Computer Systems, 2021, 115, 844856, doi: 10.1016/j.future.2020.10.008.
[75]
M. A. Ezzat, M. A. Abd El Ghany, S. Almotairi, M. A.-M. Salem, Horizontal review on video surveillance
for smart cities: edge devices, applications, datasets, and future trends, Sensors, 2021, 21, 3222, doi:
10.3390/s21093222.
[76]
A. Al-Kaff, Navigating the Future: AI innovations for intelligent mobility in smart cities, SAE
International, 2023, doi: 10.4271/2023-01-1901.
[77]
T. K. Rusch, M. M. Bronstein, S. Mishra, A Survey on Oversmoothing in Graph Neural Networks, arXiv,
2023, doi: 10.48550/arXiv.2303.10993.
[78]
A. Belhadi, Y. Djenouri, G. Srivastava, D. Djenouri, J. C.-W. Lin, G. Fortino, Deep learning for pedestrian
collective behavior analysis in smart cities: A model of group trajectory outlier detection, Information
Fusion, 2021, 65, 1320, doi: 10.1016/j.inffus.2020.08.003.
[79]
S. Y. He, S. Tao, Z. Wang, S. Luo, Data for travel behaviour research: Recent advances, challenges and
opportunities in the era of smart cities, Handbook of Travel Behaviour, Edward Elgar Publishing,
2024, 197218, doi: 10.4337/9781839105746.00018.
[80]
C. Xu, H. Wu, Y. Zhang, S. Dai, H. Liu, J. Tian, A Real-Time Complex Road AI Perception Based on 5G-
V2X for Smart City Security, Wireless Communications and Mobile Computing, 2022, 2022, 4405242,
doi: 10.1155/2022/4405242
[81]
R. Panchalingam, K. C. Chan, A state-of-the-art review on artificial intelligence for smart buildings,
Intelligent Buildings International, 2021, 13, 203226, doi: 10.1080/17508975.2019.1613219.
[82]
J. S. Saini, S. Arora, S. Kamboj, Prediction of Smart Building and Smart City Resources using AI-
techniques, 2023 2nd International Conference for Innovation in Technology (INOCON), IEEE, 2023,
15, doi: 10.1109/INOCON57975.2023.10101326.
[83]
U. Bissarinova, A. Tleuken, S. Alimukhambetova, H. A. Varol, F. Karaca, DL-SLICER: Deep learning for
satellite-based identification of cities with enhanced resemblance, Buildings, 2024, 14, 551, doi:
10.3390/buildings14020551.
[84]
L. Mabon, W.-Y. Shih, Urban greenspace as a climate change adaptation strategy for subtropical Asian
cities: A comparative study across cities in three countries, Global Environmental Change, 2021, 68,
102248, doi: 10.1016/j.gloenvcha.2021.102248.
[85]
R. Jayavadivel, M. Arunachalam, G. Nagarajan, B. Prabhu Shankar, C. Viji, N. Rajkumar, K. R.
Senthilkumar, Historical overview of AI adoption in libraries, AI-assisted library reconstruction, IGI
Global Scientific Publishing, 2024, 267289, doi: 10.4018/979-8-3693-2782-1.ch015.
[86]
M. Lymperaiou, G. Stamou, A survey on knowledge-enhanced multimodal learning, Artificial
Intelligence Review, 2024, 57, 284, doi: 10.1007/s10462-024-10825-z.
[87]
N. C. A. Sallang, M. T. Islam, M. S. Islam, H. Arshad, A CNN-Based Smart Waste Management System
Using TensorFlow Lite and LoRa-GPS Shield in Internet of Things Environment, IEEE Access, 2021, 9,
153560153574, doi: 10.1109/ACCESS.2021.3128314.
[88]
T. -H. Chen, Y.-T. Wang, C.-H. Wu, C.-F. Kuo, H.-T. Cheng, S.-W. Huang, C. Lee, A colonial serrated polyp
classification model using white-light ordinary endoscopy images with an artificial intelligence
model and TensorFlow chart, BMC Gastroenterology, 2024, 24, 99, doi: 10.1186/s12876-024-03181-
3.
[89]
2024 17th
International Conference of Education, Research and Innovation (ICERI), 2024, 1068910696, doi:
10.21125/iceri.2024.2775.
[90]
L. Pilgram, F. K. Dankar, J. Drechsler, M. Elliot, J. Domingo-Ferrer, P. Francis, M. Kantarcioglu, L. Kong,
B. Malin, K. Muralidhar, P. Myles, F. Prasser, J. L. Raisaro, C. Yan, K. El Emam, A consensus privacy
metrics framework for synthetic data, Patterns, 2025, 6, 101320, doi: 10.1016/j.patter.2025.101320.
[91]
M. Ciotola, G. Guarino, A. Mazza, G. Poggi, G. Scarpa, A comprehensive benchmarking framework for
sentinel-2 sharpening: methods, dataset, and evaluation metrics, Remote Sensing, 2025, 17, 1983,
doi: 10.3390/rs17121983.
[92]
B. H. Reddy, A. S. Vickram, P. R. Karthikeyan, Classification of fire and smoke images using random
forest algorithm in comparison with decision tree to measure accuracy, precision, recall, F-score, AIP
Conference Proceedings, 2024, 020077, doi: 10.1063/5.0198489.
[93]
P. Kairouz, H. B. McMahan, B. Avent, A. Bellet, M. Bennis, A. N. Bhagoji, K. A. Bonawitz, Z. Charles, G.
Cormode, R. Cummings, R. G. L. D'Oliveira, H. Eichner, S. El Rouayheb, D. Evans, J. Gardner, Z. Garrett,
A. Gascón, B. Ghazi, P. B. Gibbons, M. Gruteser, Z. Harchaoui, C. He, L. He, Z. Huo, B. Hutchinson, J. Hsu,
Liu, P. Mittal, M. Mohri, R. Nock, A. Özgür, R. Pagh, M. Raykova, H. Qi, D. Ramage, R. Raskar, D. Song, W.
Song, S. U. Stich, Z. Sun, A. T. Suresh, F. Tramèr, P. Vepakomma, J. Wang, L. Xiong, Z. Xu, Q. Yang, F. X. Yu,
H. Yu, S. Zhao, Advances and open problems in federated learning, Foundations and Trends in Machine
Learning, 2021, 14, 1210, doi: 10.1561/2200000083.
[94]
J. Zhou, Y. Zhang, Y. Qiu, State-of-the-art review of machine learning and optimization algorithms
applications in environmental effects of blasting, Artificial Intelligence Review, 2024, 57, doi:
10.1007/S10462-023-10636-8.
[95]
Scientific Reports,
2024, 14, 13829, doi: 10.1038/s41598-024-64483-w.
[96]
T. Otto, M. Klöppel-Gersdorf, I. Partzsch, A framework for urban C-ITS GLOSA evaluation, Smart
Energy for Smart Transport, Springer, 2023, 382392, doi: 10.1007/978-3-031-23721-8_33.
[97]
A. Arora, A. Jain, D. Yadav, V. Hassija, V. Chamola, B. Sikdar, Next generation of multi-agent driven smart
city applications and research paradigms, IEEE Open Journal of the Communications Society, 2023, 4,
21042121, doi: 10.1109/OJCOMS.2023.3310528.
[98]
J. Ye, J. Zhao, K. Ye, C. Xu, How to build a graph-based deep learning architecture in traffic domain: A
survey, IEEE Transactions on Intelligent Transportation Systems, 2022, 23, 39043924, doi:
10.1109/TITS.2020.3043250.
[99]
D. Li, T. Nanseki, Y. Chomei, J. Kuang, A review of smart agriculture and production practices in
Japanese large-scale rice farming, Journal of the Science of Food and Agriculture, 2023, 103, 1609
1620, doi: 10.1002/jsfa.12204.
[100]
D. Juanatey, M. Naya, T. Baamonde, F. Bellas, Developing a Simulation Model for Autonomous Driving
Education in the Robobo SmartCity Framework, The 4th XoveTIC Conference, MDPI, 2021, 49, doi:
10.3390/engproc2021007049.
[101]
Ishwarappa, J. Anuradha, Big data based stock trend prediction using deep CNN with reinforcement-
LSTM model, International Journal of System Assurance Engineering and Management, 2021, doi:
10.1007/s13198-021-01074-2.
[102]
A. Khalil, H. Farman, M. M. Nasralla, B. Jan, J. Ahmad, Artificial Intelligence-based intrusion detection
system for V2V communication in vehicular adhoc networks, Ain Shams Engineering Journal, 2024,
15, 102616, doi: 10.1016/j.asej.2023.102616
[103]
G. Cepeliauskaite, B. Keppner, Z. Simkute, Z. Stasiskiene, L. Leuser, I. Kalnina, N. Kotovica, J. Andin, M.
Muiste, Smart-mobility services for climate mitigation in urban areas: case studies of Baltic countries
and germany, Sustainability, 2021, 13, 4127, doi: 10.3390/su13084127.
[104]
A. Patronov, K. Papadopoulos, O. Engkvist, Has artificial intelligence impacted drug discovery?,
Artificial Intelligence in Drug Discovery, Springer, 2022, 153176, doi: 10.1007/978-1-0716-1787-
8_6.
[105]
M. Alymani, L. A. Almoqhem, D. A. Alabdulwahab, A. A. Alghamdi, H. Alshahrani, K. Raza, Enabling
smart parking for smart cities using Internet of Things (IoT) and machine learning, PeerJ Computer
Science, 2025, 11, e2544, doi: 10.7717/peerj-cs.2544.
[106]
N. El Karkouri, L. Hassine, Y. Ledmaoui, H. Chaibi, R. Saadane, N. Enneya, M. El Aroussi, Enhancing
route optimization in road transport systems through machine learning: a case study of the Dakhla-
Paris corridor, Future Transportation, 2025, 5, 60, doi: 10.3390/futuretransp5020060.
[107]
X. Xiao, Z. Peng, Y. Lin, Z. Jin, W. Shao, R. Chen, N. Cheng, G. Mao, Parking Prediction in Smart Cities: A
Survey, IEEE Transactions on Intelligent Transportation Systems, 2023, 24, 1030210326, doi:
10.1109/TITS.2023.3279024.
[108]
V. Adewopo, N. Elsayed, Z. Elsayed, M. Ozer, C. Zekios, A. Abdelgawad, M. Bayoumi, big data and deep
learning in smart cities: a comprehensive dataset for ai-driven traffic accident detection and
computer vision systems, 2024 IEEE SoutheastCon, IEEE, 2024, 675680, doi:
10.1109/SoutheastCon52093.2024.10500288
[109]
M. L. Francies, A. T. Khalil, H. M. Amer, M. M. Ata, A framework for continual learning in real-time
traffic forecasting utilizing spatialtemporal graph convolutional recurrent networks, Complex &
Intelligent Systems, 2025, 11, doi: 10.1007/s40747-025-02049-7.
[110]
X. Chen, J. Wang, K. Xie, TrafficStream: A Streaming Traffic Flow Forecasting Framework Based on
Graph Neural Networks and Continual Learning, Proceedings of the 30th International Joint
Conference on Artificial Intelligence (IJCAI), 2021, 36203626, doi: 10.24963/ijcai.2021/498.
[111]
A. Chahal, P. Gulia, N. S. Gill, N. Sultana, Aquila Optimizer-Based Hybrid Predictive Model for Traffic
Congestion in an IoT-Enabled Smart City, International Journal of Intelligent Systems, 2024, 2024,
5577278, doi: 10.1155/2024/5577278.
[112]
G. Kaur, S. K. Grewal, A. Jain, Federated Learning based Spatio-Temporal framework for real-time
traffic prediction, Wireless Personal Communications, 2024, 136, 849865, doi: 10.1007/S11277-
024-11292-Z
[113]
M. Du, P. Yang, Y. Liu, RDP-FedGAN: A Rényi-differential privacy empowered federated learning GAN
in smart parking, IEEE Transactions on Vehicular Technology, 2024, 74, 100-109, doi: doi:
10.1109/TVT.2024.3462108.
[114]
S. Z. Marshoodulla, G. Saha, SDIoTPark: A data analytics framework for smart parking using SDN-
BASED IoT, IEEE Internet of Things Journal, 2024, 11, 2003020039, doi:
10.1109/JIOT.2024.3373133.
[115]
E. Collini, P. Nesi, G. Pantaleo, Deep learning for short-term prediction of available bikes on bike-
sharing stations, IEEE Access, 2021, 9, 124337124347, doi: 10.1109/ACCESS.2021.3110794
[116]
M. J. Harrison, S. Liu, Climate change adaptation strategies in coastal cities of Australia: A comparative
study of Sydney, Melbourne, and Brisbane, Studies in Social Science & Humanities, 2024, 3, 6471.
[117]
Alwiyah, D. P. Lazirkha, The impact of artificial intelligence in smart city air purifier systems, Aptisi
Transactions on Technopreneurship, 2022, 4, 205214, doi: 10.34306/att.v4i2.180.
[118]
M. Melchiorri, S. Freire, M. Schiavina, A. Florczyk, C. Corbane, L. Maffenini, M. Pesaresi, P. Politis, F.
Szabo, D. Ehrlich, P. Tommasi, D. Airaghi, L. Zanchetta, T. Kemper, The multi-temporal and multi-
dimensional global urban centre database to delineate and analyse world cities, Scientific Data, 2024,
11, 82, doi: 10.1038/s41597-023-02691-1.
[119]
K. Sako, B. N. Mpinda, P. C. Rodrigues, Neural networks for financial time series forecasting, Entropy,
2022, 24, 657, doi: 10.3390/e24050657.
[120]
R. Keisler, Forecasting Global Weather with Graph Neural Networks, arXiv, 2022, doi:
10.48550/arXiv.2202.07575. Accessed 2 August 2025.
[121]
R. Mathumitha, P. Rathika, K. Manimala, Intelligent deep learning techniques for energy consumption
forecasting in smart buildings: a review, Artificial Intelligence Review, 2024, 57, doi:
10.1007/S10462-023-10660-8.
[122]
H. C. Bhandari, Y. R. Pandeya, K. Jha, S. Jha, Recent advances in electrical engineering: exploring graph
neural networks for weather prediction in data-scarce environments, Environmental Research
Communications, 2024, 6, 105010, doi: 10.1088/2515-7620/ad8108.
[123]
B. Wang, S. Zhu, M. Cane, Y. Ma, Global weather forecasting through deep learning: A comprehensive
survey, TRCN, 2024. Available at: TRCN. Accessed 28 July 2025.
[124]
P. Ghosh, A. Neufeld, J. K. Sahoo, Forecasting directional movements of stock prices for intraday
trading using LSTM and random forests, Finance Research Letters, 2022, 46, 102280, doi:
10.1016/j.frl.2021.102280.
[125]
P. Hewage, M. Trovati, E. Pereira, A. Behera, Deep learning-based effective fine-grained weather
forecasting model, Multimedia Tools and Applications, 2021, 80, 343366, doi: 10.1007/S10044-020-
00898-1.
[126]
P. A. Khajgiwale, A. Naveen, N. Jethani, A. Shitole, Comparative analysis of machine learning methods
in climate data extraction and forecasting, 2025 Emerging Technologies for Intelligent Systems
(ETIS), Trivandrum, India, 2025, 1-7, doi: 10.1109/ETIS64005.2025.10961537.
[127]
S. Y. Jhin, S. Kim, N. Park, Addressing prediction delays in time series forecasting: A continuous GRU
approach with derivative regularization, Proceedings of the ACM on Management of Data, 2024, 2,
12341245, doi: 10.1145/3637528.3671969.
[128]
Z. Xu et al., DGFormer: A physics-guided station level weather forecasting model with dynamic
spatial-temporal graph neural network, Intelligent Data Analysis, 2024, 28, 499533, doi:
10.1007/S10707-024-00511-1.
[129]
Y. He, C. L. Tan, Y. Lu, Y. Qu, J. Liu, Solar radiation and daily light integral characteristics of photovoltaic-
green roof system in tropical area: A case study in Singapore, Solar Energy, 2024, 279, 112807, doi:
10.1016/j.solener.2024.112807.
[130]
G. U. Fayomi, S. E. Mini, C. M. Chisom, O. S. I. Fayomi, N. E. Udoye, O. Agboola, D. Oomole, Smart waste
management for smart city: Impact on industrialization, IOP Conference Series: Earth and
Environmental Science, 2021, 655, 012040, doi: 10.1088/1755-1315/655/1/012040.
[131]
S. Shahab, M. Anjum, Solid Waste Management scenario in India and illegal dump detection using
deep learning: an ai approach towards the sustainable waste management, Sustainability, 2022, 14,
15896, doi: 10.3390/su142315896.
[132]
B. Ugwoke, S. P. Corgnati, P. Leone, R. Borchiellini, J. M. Pearce, Low emissions analysis platform model
for renewable energy: Community-scale case studies in Nigeria, Sustainable Cities and Society, 2021,
67, 102750, doi: 10.1016/j.scs.2021.102750.
[133]
E. Velasco, Circular economy in Singapore: waste management, food and agriculture, energy, and
transportation, Urban Resilience and Sustainability, 2024, 2, 110150, doi:
https://doi.org/10.3934/urs.2024007.
[134]
CNN-LSTM with transfer learning for sustainable environment, Multimedia Tools and Applications,
2023, 83, 2950529529, doi: https://doi.org/10.1007/s11042-023-16677-z.
[135]
T. Wang, F. Zhang, H. Gu, H. Hu, M. Kaur, A research study on new energy brand users based on
principal component analysis (PCA) and fusion target planning model for sustainable environment
of smart cities, Sustainable Energy Technologies and Assessments, 2023, 57, 103262, doi:
https://doi.org/10.1016/j.seta.2023.103262.
[136]
S. Alshmrany, Adaptive learning style prediction in e-learning environment using levy flight
distribution based CNN model, Cluster Computing, 2022, 25, 523536, doi: 10.1007/s10586-021-
03403-3.
[137]
S. M. M. Islam, K. H. Talukder, Smart environment index prediction of smart city using polynomial
regression, Khulna University Studies, 2022, 676689, doi: 10.53808/KUS.2022.ICSTEM4IR.0130-se.
[138]
Z. Ben-Bouallegue, M. C. A. Clare, L. Magnusson, E. Gascon, M. Maier-Gerber, M. Janousek, M. Rodwell,
F. Pinault, J. S. Dramsch, S. T. K. Lang, B. Raoult, F. Rabier, M. Chevallier, I. Sandu, P. Dueben, M. Chantry,
F. Pappenberger, The rise of data-driven weather forecasting, Bulletin of the American Meteorological
Society, 2024, 105, doi: https://doi.org/10.1175/BAMS-D-23-0162.1.
[139]
M. V. du Bray, R. Stotts, R. Southee, A. Wutich, Beyond extreme: Heat emergency and water insecurity
for people experiencing houselessness in Phoenix, Arizona, USA during and after the heatwave of
2023, Human Organization, 2023, 51, 799808, doi: 10.1007/S10745-023-00447-4.
[140]
B. Bochenek, Z. Ustrnul, Machine learning in weather prediction and climate analyses—applications
and perspectives, Atmosphere, 2022, 13, 180, doi: 10.3390/atmos13020180.
[141]
L. Chen, B. Han, X. Wang, J. Zhao, W. Yang, Z. Yang, Machine learning methods in weather and climate
applications: A survey, Applied Sciences, 2023, 13, 12019, doi:
https://doi.org/10.3390/app132112019.
[142]
T. A. Amnuaylojaroen, Advancements and challenges of artificial intelligence in climate modeling for
sustainable urban planning, Frontiers in Artificial Intelligence, 2025, 8, 1517986, doi:
10.3389/frai.2025.1517986.
[143]
R. Snaiki, S. S. Parida, A data-driven physics-informed stochastic framework for hurricane-induced
risk estimation of transmission tower-line systems under a changing climate, International Journal of
Electrical Power & Energy Systems, 2023, 280, 115673, doi: 10.1016/j.engstruct.2023.115673.
[144]
B. N. Rao, M. Praveen, D. R. Babu, A review on the role of AI in optimizing renewable energy grid
management, International Journal of Scientific Research in Engineering and Management, 2024, 20,
7697.
[145]
P. Xue, C. Huang, Y. Zhong, M. Notaro, M. B. Kayastha, X. Zhou, C. Zhao, C. P.-Lidard, C. Cruz, E. Kemp
Enhancing winter climate simulations of the Great Lakes: insights from a new coupled lakeice
atmosphere (CLIAv1) system on the importance of integrating 3D hydrodynamics with a regional
climate model, Geoscientific Model Development, 2025.18, doi: 10.5194/gmd-18-4293-2025.
[146]
B. Bonev, T. Kurth, A. Mahesh, FourCastNet 3: A geometric approach to probabilistic machine-learning
weather forecasting at scale, arXiv, arXiv:2507.12144, 2025, doi: 10.48550/arXiv.2507.12144.
[147]
A. Khandekar, J. Cross, A. Maringanti, Scale and modularity in thermal governance: The replication of
Urban Studies, 2024, 61, 28682886, doi: 10.1177/00420980231195193.
[148]
M. Ma, Y. Huang, Y. Zhang, Y. Li, J. Zhang, HiSTGNN: Hierarchical spatio-temporal graph neural network
for weather forecasting, Information Sciences, 2023, 646, 119406, doi: 10.1016/j.ins.2023.119406.
[149]
M. Méndez, M. G. Merayo, M. Núñez, Machine learning algorithms to forecast air quality: a survey,
Artificial Intelligence Review, 2023, 56, 1003110066, doi: 10.1007/S10462-023-10424-4.
[150]
A. Siddique, N. Alasbali, M. Driss, W. Boulila, M. S. Alshehri, Jawad Ahmad Sustainable collaboration:
Federated learning for environmentally conscious forest fire classification in Green Internet of Things
(IoT), Internet of Things, 2024, 25, 101013, doi: 10.1016/j.iot.2023.101013.
[151]
M. Addad, S. Al-Taani, Leveraging AI for energy-efficient smart cities: architectural and urban planning
solutions for sustainable growth - a comparative case study of Amman city and international
examples, Journal of Civil Engineering and Architecture, 2024, 18, doi: 10.17265/1934-
7359/2024.12.006.
[152]
R. M. Moalem, M. A. Mosgaard, A critical review of the role of repair Cafés in a sustainable circular
transition, Sustainability, 2021, 13, 12351, doi: 10.3390/su132212351.
[153]
K. Alanne, S. Sierla, An overview of machine learning applications for smart buildings, Sustainable
Cities and Society, 2022, 76, 103445, doi: 10.1016/j.scs.2021.103445.
[154]
Z. Nie, H. Farzaneh, Role of model predictive control for enhancing eco-driving of electric vehicles in
urban transport system of Japan, Sustainability, 2021, 13, 9173, doi: 10.3390/su13169173.
[155]
S. Müller, G. Rossi, V. Ivanova, Smart city development: challenges and opportunities in Europe,
International Journal of Engineering Sciences and Technologies, 2024, 2, doi:
https://doi.org/10.58531/ijest/2/1/4.
[156]
K. A. Bonawitz, V. Ivanov, B. Kreuter, A. Marcedone, H. B. McMahan, S. Patel, D. Ramage, A. Segal, K.
Seth, Practical secure aggregation for privacy-preserving machine learning, Proceedings of the 2017
ACM SIGSAC Conference on Computer and Communications Security, ACM, 2017, 11751191, doi:
10.1145/3133956.3133982.
[157]
D.-K. Lee, Analysis of the potential value of cultural ecosystem services: A case study of Busan City,
Republic of Korea, Ecosystem Services, 2024, 65, 101596, doi: 10.1016/j.ecoser.2024.101596.
[158]
P. Kumar, K. Silambarasan, Enhancing the performance of healthcare service in IoT and cloud using
optimized techniques, IETE Journal of Research, 2022, 68, 14751484, doi:
10.1080/03772063.2019.1654934.
[159]
P. Gonzalez-Pizarro, L. Brazzi, S. Koch, A. Trinks, J. Muret, N. Sperna Weiland, G. Jovanovic, A.
Cortegiani, T. D. Fernandes, P. Kranke, A. Malisiova, P. McConnell, L. Misquita, C. S. Romero, F. Bilotta,
E. De Robertis, W. Buhre, European Society of Anaesthesiology and Intensive Care consensus
document on sustainability: 4 scopes to achieve a more sustainable practice, European Journal of
Anaesthesiology, 2024, 41, 260277, doi: 10.1097/EJA.0000000000001942.
[160]
I.-S. Sorlei, N. Bizon, P. Thounthong, M. Varlam, E. Carcadea, M. Culcer, M. Iliescu, M. Raceanu, Fuel cell
electric vehicles—a brief review of current topologies and energy management strategies, Energies,
2021, 14, 252, doi: 10.3390/en14010252.
[161]
M. Javed, N. Tariq, M. Ashraf, F. A. Khan, M. Asim, M. Imran, Securing smart healthcare cyber-physical
systems against blackhole and greyhole attacks using a blockchain-enabled Gini index framework,
Sensors, 2023, 23, 9372, doi: 10.3390/s23239372.
[162]
L. Abirami, J. Karthikeyan, Digital Twin-Based Healthcare System (DTHS) for earlier parkinson
disease identification and diagnosis using optimized fuzzy based k-nearest neighbor classifier model,
IEEE Access, 2023, 11, 9666196672, doi: 10.1109/ACCESS.2023.3312278.
[163]
S. Mekruksavanich, A. Jitpattanakul, Deep Convolutional Neural Network with RNNs for Complex
Activity Recognition Using Wrist-Worn Wearable Sensor Data, Electronics, 2021, 10, 1685, doi:
10.3390/electronics10141685.
[164]
A. H. M. Kamal, Md. G. R. Alam, M. R. Hassan, T. S. Apon, M. M. Hassan, Explainable indoor localization
of BLE devices through RSSI using recursive continuous wavelet transformation and XGBoost
classifier, Future Generation Computer Systems, 2023, 141, 230242, doi:
10.1016/j.future.2022.11.001.
[165]
H. Fukuda, F. Murata, S. Azuma, M. Fujimoto, S. Kudo, Y. Kobayashi, K. Saho, K. Nakahara, R. Ono,
Development of a data platform for monitoring personal health records in Japan: The Sustaining
Health by Integrating Next-generation Ecosystems (SHINE) Study, PLoS ONE, 2023, 18, e0281512,
doi: 10.1371/journal.pone.0281512.
[166]
A. Baroncelli, M. Ruberti, Smart sport arenas make cities smarter, Managing Smart Cities, Springer,
2022, 89104, doi: 10.1007/978-3-030-93585-6_6.
[167]
L. Zhong, X. Deng, A cloud and IoT-enabled workload-aware healthcare framework using ant colony
optimization algorithm, International Journal of Advanced Computer Science and Applications, 2023,
14, doi: 10.14569/IJACSA.2023.0140394.
[168]
Y. Cao, M. Erdt, C. Robert, N. B. Naharudin, S. Q. Lee, Y.-L. Theng, Decision-making factors toward the
adoption of smart home sensors by older adults in singapore: mixed methods study, JMIR Aging, 2022,
5, e34239, doi: 10.2196/34239.
[169]
H. M. K. K. M. B. Herath, Internet of Things (IoT) enable designs for identify and control the COVID-19
Pandemic, Internet of Things and COVID-19, Springer, 2021, 423436, doi: 10.1007/978-3-030-
69744-0_24.
[170]
A. H. Hassan, R. bin Sulaiman, M. A. Abdulgabber, H. Kahtan, Balancing technological advances with
user needs: user-centered principles for ai-driven smart city healthcare monitoring, International
Journal of Advanced Computer Science and Applications, 2023, 14, doi:
10.14569/IJACSA.2023.0140341.
[171]
T. Yang, Y. Mei, L. Xu, H. Yu, Y. Chen, Application of question answering systems for intelligent
agriculture production and sustainable management: A review, Resources, Conservation and
Recycling, 2024, 204, 107497, doi: 10.1016/j.resconrec.2024.107497.
[172]
A. Nikolaeva, Smart Cities and (Smart) Cycling: Exploring the synergies in Copenhagen and
Amsterdam, Journal of Urban Technology, 2024, 31, 2949, doi: 10.1080/10630732.2024.2322007.
[173]
T. Natalia, S. Kumar Joshi, S. Dixit, H. Kanakadurga Bella, P. Chandra Jena, A. Vyas, Enhancing smart
city services with AI: A field experiment in the context of industry 5.0, BIO Web of Conferences, 2024,
86, 01063, doi: 10.1051/bioconf/20248601063.
[174]
G. F. Huseien, K. W. Shah, Potential applications of 5g network technology for climate change control:
a scoping review of singapore, Sustainability, 2021, 13, 9720, doi: 10.3390/su13179720.
[175]
for improving communication efficiency, arXiv:1610.05492, doi: 10.48550/arXiv.1610.05492.
[176]
Z. Li, A. Lowy, J. Liu, T. Koike-Akino, K. Parsons, B. Malin, Y. Wang, Analyzing inference privacy risks
through gradients in machine learning, Proceedings of the ACM on Management of Data, 2024, 2,
34663480, doi: 10.1145/3658644.3690304.
[177]
M. Sheikh, J. Aghaei, H. Chabok, M. Roustaei, T. Niknam, A. Kavousi-Fard, João P. S. Catalão, Synergies
between transportation systems, energy hub and the grid in smart cities, IEEE Transactions on
Intelligent Transportation Systems, 2021, 23, 7371-7385, doi: 10.1109/TITS.2021.3069354.
[178]
J. Alshehri, A. Alhamed, M. M. H. Rahman, A systematic literature review on cybersecurity risk
management in smart cities, 2024 International Conference on Artificial Intelligence in Information
and Communication (ICAIIC), IEEE, 2024, 407412, doi: 10.1109/ICAIIC60209.2024.10463312.
[179]
G. Lastro, J. -A. Gross, C. Neureiter, A Standards-based approach for cross-domain modelling of smart
city system architectures, 2023 8th International Conference on Smart and Sustainable Technologies
(SpliTech), Split/Bol, Croatia, 2023, 1-6, doi: 10.23919/SpliTech58164.2023.10193583.
[180]
M. S. Gonzalez, M. E. Santos, Sustainable cities, smart investments: A characterization of a thousand
days-san miguel, a program for vulnerable early childhood in Argentina, Sustainability, 2023, 15,
12205, doi: 10.3390/su151612205.
[181]
A. Hagiu, J. Wright, Artificial intelligence and competition policy, International Journal of Industrial
Organization, 2025, 103134, doi: 10.1016/j.ijindorg.2025.103134.
[182]
M. A. Khan, H. Park, Exploring explainable artificial intelligence techniques for interpretable neural
networks in traffic sign recognition systems, Electronics, 2024, 13, 306, doi:
10.3390/electronics13020306.
[183]
J. Schneider, C. Meske, P. Kuss, Foundation Models, Business & Information Systems Engineering, 2024,
66, 221231, doi: 10.1007/s12599-024-00851-0.
Publisher Note: The views, statements, and data in all publications solely belong to the authors and
contributors. GR Scholastic is not responsible for any injury resulting from the ideas, methods, or products
mentioned. GR Scholastic remains neutral regarding jurisdictional claims in published maps and institutional
affiliations.
Open Access
This article is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License, which
permits the non-commercial use, sharing, adaptation, distribution and reproduction in any medium or format,
as long as appropriate credit to the original author(s) and the source is given by providing a link to the Creative
Commons License and changes need to be indicated if there are any. The images or other third-party material
in this article are included in the article's Creative Commons License, unless indicated otherwise in a credit line
to the material. If material is not included in the article's Creative Commons License and your intended use is
not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly
from the copyright holder. To view a copy of this License, visit: https://creativecommons.org/licenses/by-
nc/4.0/
© The Author(s) 2026
