Supervised Machine Learning Approaches for Predicting Key Pollutants and for the Sustainable Enhancement of Urban Air Quality: A Systematic Review

Abstract

Urban air pollution is a pressing global issue driven by factors such as swift urbanization, population expansion, and heightened industrial activities. To address this challenge, the integration of Machine Learning (ML) into smart cities presents a promising avenue. Our article offers comprehensive insights into recent advancements in air quality research, employing the PRISMA method as a cornerstone for the reviewing process, while simultaneously exploring the application of frequently employed ML methodologies. Focusing on supervised learning algorithms, the study meticulously analyzes air quality data, elucidating their unique benefits and challenges. These frequently employed ML techniques, including LSTM (Long Short-Term Memory), RF (Random Forest), ANN (Artificial Neural Networks), and SVR (Support Vector Regression), are instrumental in our quest for cleaner, healthier urban environments. By accurately predicting key pollutants such as particulate matter (PM), nitrogen oxides (NO_x), carbon monoxide (CO), and ozone (O₃), these methods offer tangible solutions for society. They enable informed decision-making for urban planners and policymakers, leading to proactive, sustainable strategies to combat urban air pollution. As a result, the well-being and health of urban populations are significantly improved. In this revised abstract, the importance of frequently employed ML methods in the context of air quality is explicitly emphasized, underlining their role in improving urban environments and enhancing the well-being of urban populations.

1. Introduction

In recent years, there has been a significant increase in public awareness of the effects of population growth and activity on the environment. The United Nations estimates that there are currently 7.6 billion people living in the world, with 4.2 billion (or around 55%) of them living in cities [1]. By 2050, it is estimated that the urban population will double compared to its current size. This rapid growth in urban areas has led to a significant increase in industrialization to meet the demands of the expanding population.

However, this rapid pace of industrial and economic development has led to harmful repercussions for the environment, particularly in terms of air quality, posing a significant challenge to both human health and the environment. The influx of people into cities and the subsequent increase in industrial activity have resulted in elevated levels of air pollutants. These pollutants are often released from vehicular emissions, power generation, and manufacturing processes, accumulating in the atmosphere and contributing to the formation of harmful smog and haze. As a consequence, urban residents are increasingly exposed to poor air quality, which can lead to a range of respiratory and cardiovascular ailments. Additionally, the long-range transport of these pollutants can adversely affect not only local communities but also distant regions, amplifying the global scale of the issue. Urgent measures are required to mitigate air quality degradation, including the implementation of cleaner technologies, enhanced urban planning, and the promotion of sustainable transportation systems to ensure a healthier future for both the planet and its inhabitants.

The concept of a smart city offers an opportunity to effectively and efficiently address these challenges. With air pollution posing a threat to urban living, it is crucial to address this issue promptly and wisely. A smart city refers to an urban locality that leverages information and communication technologies (ICT) to enhance the quality of life for its residents. It achieves this by offering improved healthcare, transportation, and energy services while enabling the government to optimize resource utilization for the well-being of its citizens.

In the context of air quality research, it is essential to recognize the value of many qualitative tools. These tools include chemical analysis: Ambient Sampling, Source Profiling, and Receptor Modeling, which hold profound significance in the field of air quality management. These methodologies offer a systematic approach to comprehending the multifaceted nature of air pollution, facilitating the precise identification and quantification of pollution sources. By combining the insights derived from chemical analysis with rigorous statistical techniques, these models equip environmental scientists and policymakers with a robust foundation for devising targeted pollution control strategies, establishing regulatory benchmarks, and executing efficacious measures to ameliorate air quality. Their pivotal role in safeguarding public health, preserving ecological integrity, and promoting sustainable development through the foundational mitigation of air pollution renders them indispensable tools in contemporary environmental science and policy paradigms. As such, these qualitative models occupy a central place in the arsenal of methodologies available for addressing the complex challenge of air quality management [2].

Moreover, the convergence of ICT with ML presents substantial potential for fostering groundbreaking solutions. An area that particularly stands out in smart city applications is the enhancement of air quality monitoring. ML algorithms demonstrate the capability to effectively interpret and analyze extensive datasets concerning air quality, sourced from diverse channels such as sensors and weather stations [3]. In the field of ML, we encounter a trio of techniques: supervised, unsupervised, and semi-supervised methods. Supervised ML entails instructing a model using meticulously labeled data, where the expected outcome is already discerned. On the other hand, unsupervised ML embarks on the uncharted terrain of unlabeled data, excelling in tasks such as categorizing data groups into clusters founded on shared characteristics. The hybrid amalgam known as semi-supervised ML presents an appealing avenue, especially in situations where labeled data are sparse. As a result of this analysis, these algorithms can discern patterns, predict pollution levels, and offer real-time monitoring of the city’s air quality status.

The motivation behind this study is multifaceted and driven by a confluence of factors. First and foremost, there is a growing global concern for environmental well-being, with a pressing need to address the adverse impacts of air pollution on ecosystems and wildlife. Additionally, public health is a paramount consideration, as air pollution poses significant health risks to communities, making it imperative to gain a deeper understanding of the issue. Policy and regulatory frameworks related to air quality are of great importance, and this study seeks to provide evidence-based insights to inform and shape effective policies. Furthermore, the study is motivated by the remarkable technological advancements in supervised learning and data analytics, which offer new avenues for air quality research. It also aligns with urban planning efforts, aiming to create cities that minimize air pollution and enhance residents’ quality of life. In light of global sustainability goals and commitments, this research contributes to broader objectives. Moreover, there is a genuine aspiration for scientific advancement and the pursuit of interdisciplinary collaboration, uniting fields such as environmental science, data science, and public health to tackle the complex challenge of air quality comprehensively.

This study aims to conduct a thorough review of recent literature on air pollution, with a specific focus on the use of supervised learning methods in the field and the identification of key pollutants. We aspire to identify and highlight the most prevalent pollutants and the most employed techniques. We commence with an examination of related work (Section 2) and describe our study methodology within the context of prior research (Section 3). Section 4 delves into the application of supervised ML techniques in air quality monitoring, while Section 5 discusses our analysis results, limitations, and potential future research directions. Finally, Section 6 summarizes key findings and conclusions.

2. Related Works

2.1. Qualitative Approaches in the Air Quality Field

A wide range of qualitative approaches work together in the complex field of environmental science and air quality management to provide a deep comprehension and efficient reduction of air pollution. These essential instruments include the critical procedures of Chemical Analysis, Source Profiling, Receptor Modeling, and Ambient Sampling. Ambient Sampling is the first step in the process, which is a fundamental step that involves the systematic collection of air samples from various geographic locations. With the use of this methodical technique, it is possible to obtain discrete air quality snapshots and perform detailed quantification of a wide range of contaminants, including heavy metals, volatile organic compounds, and particle matter. The next step involves subjecting the gathered samples to a thorough chemical analysis, which reveals a wealth of information about the makeup of the air environment and the existence of various pollutants. After that, Source Profiling takes the lead in locating and describing the potential sources of these contaminants. A variety of methodological techniques are used by environmental scientists to track the sources of pollutants identified in the samples they have gathered. This involves a multifaceted approach that amalgamates data from chemical analysis, isotopic inquiry, and meteorological insights, enabling the discrimination between natural contributors, such as dust and sea salt, and anthropogenic sources, encompassing industrial emissions, vehicular traffic, and residential heating. Source Profiling stands as an instrumental undertaking in ascertaining the principal contributors to air pollution. Finally, Receptor Modeling assumes a pivotal role, capitalizing on the outcomes of chemical analysis to meticulously allot responsibility among the varied pollution sources. By employing a blend of statistical and mathematical techniques that meticulously consider the chemical composition of pollutants, Receptor Modeling meticulously unravels the intricate amalgam found within the samples. This necessitates the application of methodologies such as chemical mass balance, positive matrix factorization (PMF), and factor analysis, all directed toward delivering a comprehensive dissection of the contributions from disparate sources. These receptor models not only foster a more profound understanding but also serve as invaluable resources for both environmental agencies and policymakers, endowing them with the requisite insights to develop meticulously precise pollution control strategies and regulatory frameworks, all directed toward the augmentation of air quality [2].

2.2. Air Quality Analysis and Forecasting

Multiple reviews have illuminated various facets of air quality analysis, forecasting, and related domains. Notably, the role of data quality in constructing effective Air Quality Models (AQMs), highlights challenges stemming from limited data availability [4]. Some studies provide a contemporary overview of air pollution forecasting, spotlighting models like the Air Quality Index (AQI) and discussing predictor variables and emerging trends [5]. In the domain of air quality prediction, there has been a substantial body of research focused on ML algorithms, with particular attention given to Artificial Neural Networks (ANN). The primary objective of this research was to elucidate the connection between the Air Quality Index (AQI) and various ML approaches, including ANN and Multiple Linear Regression (MLR) [6]. Furthermore, the investigation extends to encompass ANN, Deep Neural Networks (DNN), Support Vector Machines (SVM), and Fuzzy Logic, highlighting the need for further inquiry, particularly within the realm of DNN [7]. This research journey culminates in an assessment of data decomposition techniques in conjunction with ML models such as ANN, SVM, and the Emerging Learning Machine (ELM) for air pollution forecasting. This underscores the sustained endeavor to craft more precise and efficacious strategies to combat the pressing issue of air pollution [8].

Deep learning has gained significant prominence in the field of air pollution epidemiology, a trend that becomes readily apparent in the reviewed literature. One study [9] delves into the potential of deep learning for source apportionment and forecasting, with an emphasis on its spatial and temporal analysis capabilities. Additionally, another study [10] provides a comparative analysis of non-deep and deep learning approaches in modeling air pollutant correlations. Furthermore, the application of deep learning takes center stage in [11], focusing specifically on its utilization for time series air quality forecasting.

Shifting the focus to indoor environments, the versatility of ML comes to the forefront. In [12], an extensive review examines studies employing ML to predict occupancy behavior and patterns with a focus on its related applications to indoor air quality and thermal comfort, favoring algorithms grounded in neural networks. Furthermore, Ref. [13] establishes a connection between ML and household cooking practices, exploring its implications for carbon neutrality. In parallel, Ref. [14] conducts a comprehensive survey of ML applications in heating, ventilation, and air conditioning (HVAC) systems and their impact on building performance. Lastly, Ref. [15] employs ANN and Reinforcement Learning (RL) models to investigate indoor air quality and thermal comfort.

2.3. Machine Learning in Urban and Industrial Planning

Other reviews have highlighted the substantial role of Machine Learning (ML) models in analyzing air quality data in city planning and urban sustainability. By identifying areas with poor air quality and comprehending the contributing factors, city planners can make informed decisions to optimize land use, transportation infrastructure, and green spaces, ultimately enhancing the quality of life for urban residents. These reviews emphasize the importance of ML in advancing urban sustainability and smart city development [16,17].

At the dynamic intersection of industrial and urban planning, Machine Learning (ML) has emerged as a transformative and versatile tool, as affirmed by numerous comprehensive reviews. These comprehensive analyses accentuate the pivotal role that Machine Learning plays in the context of Industry 4.0, where data-driven decision-making takes center stage [18]. Notably, some reviews shed light on Machine Learning’s influential applications in optimizing energy efficiency within the industrial sector [19]. They also provide valuable insights into the practical challenges and opportunities encountered when implementing large-scale Machine Learning systems in real-world industrial settings [20]. Collectively, these reviews underscore the profound impact of Machine Learning on industrial planning, heralding greater efficiency and sustainability within the dynamic landscape of modern manufacturing.

2.4. Machine Learning in Climate Change Context

This interconnected web of research showcases the diverse applications of ML within urban and industrial planning. However, it is essential to recognize that the influence of ML extends beyond these domains, finding relevance within the pressing global challenge of climate change. The transition to this broader context is imperative. As articulated in [21], ML lends its capabilities to the exploration of spatial techniques in climate change research, offering new avenues for understanding and mitigating climate-related challenges. Additionally, Ref. [22] demonstrates how ML can be harnessed to discern the multifaceted impacts of climate change on public health, particularly emphasizing disparities in evidence distribution. This seamless bridge from urban and industrial planning to the wider climate change landscape underscores the indispensable and interconnected role of Machine Learning in addressing contemporary global challenges.

These methods, including qualitative ones, play a crucial role in addressing climate change and air pollution mitigation while aligning with the Sustainable Development Goals (SDGs). They offer precision in identifying and characterizing pollution sources, enabling targeted mitigation efforts and emission reductions, supporting public health (SDG 3), fostering sustainable urban planning (SDG 11), and contributing to climate action by reducing greenhouse gas emissions (SDG 13). Moreover, they have broader positive impacts on terrestrial biodiversity (SDG 15) and exemplify the importance of partnerships (SDG 17) among governments, industries, and the scientific community for effective implementation and achieving the SDGs [23].

These reviews collectively deepen our understanding of various aspects encompassing air quality analysis, forecasting techniques, and the associated challenges, underscoring the pivotal role prediction plays in safeguarding human health, the environment, and energy efficiency. However, our approach diverges as we concentrate on investigating the implementation of Supervised ML methods in the context of air quality (Figure 1). This distinct perspective redirects our attention toward the application of Supervised ML techniques, yielding novel insights into comprehending air quality analysis and forecasting, accompanied by unique challenges that demand attention.

3. Method

The study followed the PRISMA guidelines [24], and a detailed checklist is provided in Table S1 in the Supplementary Material. PRISMA ensures a structured and transparent approach to comprehensive literature reviews, enhancing rigor and reproducibility. It involves defining research questions, selecting studies, assessing quality, and synthesizing findings. By applying PRISMA, we collected recent articles on ML and urban air quality, focusing on environmental challenges in urban areas. We analyzed articles using supervised ML evaluating algorithm choice, preprocessing, and more. This yielded 59 relevant articles for our comprehensive analysis (Figure 2).

3.1. Database Collection

The review process commenced with a comprehensive collection of academic articles related to ML and urban quality. Reputable databases such as Google Scholar, MDPI, IEEE Xplore, Springer, and Elsevier were utilized to access a diverse array of journal articles. To maintain the review’s relevance and currency, the focus was on recent publications.

3.2. Initial Selection

The initial selection phase involved a meticulous process of narrowing down a pool of approximately 507 academic articles obtained from reputable databases such as Google Scholar, MDPI, IEEE Xplore, Springer, and Elsevier. This aimed to identify articles aligned with the research objective, focusing on titles and abstracts that explored the application of ML methods (supervised, unsupervised, reinforcement, and semi-supervised) in addressing environmental challenges in the context of urban air quality. To maintain relevance and currency, emphasis was placed on recent publications, and this phase served the purpose of refining the initial dataset to a more manageable set of articles by the research criteria.

3.3. Preliminary Screening

Within the selected articles, special emphasis was placed on the concept of urban air quality. Articles on urban air quality typically cover sources of pollution, health impacts, environmental sustainability, monitoring methods, regulatory aspects, mitigation strategies, community engagement, technological solutions, case studies, and policy recommendations. The review delved deeper into the selected articles, conducting a comprehensive evaluation of the employed supervised ML methodologies. Key factors such as algorithm selection, data preprocessing techniques, feature engineering, and model evaluation, were closely examined to understand the efficacy and robustness of the approaches.

3.4. Assessment and Retrieval

In the selection of final articles for this review, a meticulous approach was adopted to ensure the inclusion of high-quality, indexed research with a diverse representation of supervised learning methods. The chosen articles spanned a wide range of methodologies, including classification, regression, deep learning, and hybrid models, providing a comprehensive examination of their applications in the context of urban air quality research. Additionally, special consideration was given to incorporating various pollutants, such as PM₁₀, PM_2.5, SO₂, CO₂, and AQI, thus offering insights into the multifaceted challenges of air pollution management. This stringent selection process resulted in the compilation of 62 articles that collectively form the foundation for a comprehensive and insightful analysis of the intersection between supervised learning and diverse pollutants in urban air quality research.

3.5. Synthesis and Presentation

The reviewed articles were synthesized to create a coherent and well-structured narrative. The review commenced with an introduction that outlined the importance of applying ML in the context of urban air quality research. It then proceeded to explore the various supervised ML approaches used to tackle urban challenges and their implications on urban air quality outcomes. Special attention was given to highlighting the most noteworthy studies and their potential impact on future research and urban planning strategies.

4. Air Quality Monitoring with Supervised Learning

4.1. Air Quality Field

4.1.1. Air Quality Landscape

Economic expansion leads to urbanization and industrialization, contributing to environmental decline. The United Nations notes that 55% of the global population (4.2 billion) resides in cities, with an anticipated doubling by 2050 [1]. Such rapid urban growth heightens energy demands, driving resource consumption and carbon emissions [25,26]. Urbanization is consistently linked to environmental deterioration [27].

Industrialization, a pivotal contributor to pollution, exhibits a direct impact on emissions. A 1% industrialization rise is associated with an 11% per capita emissions increase [28]. This has immediate and long-term environmental effects [29]. These links intertwine urbanization, industrialization, climate, and urban residents’ well-being [30]. Climate change, an outcome of this intricate interplay, profoundly affects urban environments, notably air quality. Addressing local air pollution and climate change necessitates joint exploration [31]. It is predicted that there will be an increase in ozone-related fatalities due to climate change induced by greenhouse gases [32]. they are showcasing the intimate connection between air pollution and climate change.

According to [33], the average person inhales approximately 11,000 L of air per day. This air is composed of 78.19% nitrogen, 21.94% oxygen, and 0.032% carbon dioxide, with trace amounts of other gasses [34]. However, when additional elements, known as pollutants, are introduced into the air, the composition can change, leading to air pollution [35]. It is concerning that 95% of the global population breathes this polluted air, as reported by the World Health Organization (WHO). The 2014 data report identifies air pollution as one of the top eight causes of death worldwide. Specifically, in the European Union, it is estimated to be responsible for 400,000 premature deaths [36]. The WHO reports a staggering 6.5 million annual deaths attributed to air pollution [37]. Regrettably, the COVID-19 pandemic has exacerbated this already critical situation. Studies addressing the connection between air pollution and COVID-19 mortality have emerged, with a particular focus on the valuable contribution of AI-enabled imaging techniques [38]. In addition, investigations were carried out to analyze death data through ML experiments, as well as to assess the impact of the Wuhan lockdown on air quality and health. These assessments employed RF-based weather normalization methods for four pollutants in 30 Chinese cities [39]. Furthermore, research findings have indicated that a significant proportion, over 70%, of SARS-CoV-2 deaths in Italy may be attributed to air pollution and PM_2.5 [40].

The intricate interplay between air quality and health is just one facet of a much broader nexus that encompasses several critical disciplines, as depicted in Figure 3. This complex interconnection serves as a pivotal focal point for addressing contemporary urban challenges. The amalgamation of air quality considerations with urban planning, industrial planning, transportation, green infrastructure, public health, and climate change has garnered significant attention in recent research and policy discussions.

4.1.2. Pollutants and Air Quality Indices

Air pollution can be categorized into indoor and outdoor pollution [41]. Outdoor pollution can stem from natural events like fires and volcanic eruptions, as well as human activities such as manufacturing, burning, and transportation. On the other hand, indoor pollution can be influenced by activities within the home, such as cooking, smoking, and burning fuels, as well as proximity to outdoor pollution sources like highways and industrial areas [42]. It is estimated that indoor air pollution causes 4.3 million deaths, while outdoor air pollution causes 3 million deaths [43]. Studies reveal that the concentration of various pollutants, such as PM_2.5 and CO₂, is sometimes 2–5 times higher indoors than outdoors, and occasionally even more than 100 times higher, far exceeding the allowable limit established by the WHO [44]. Being exposed to pollution when staying in homes, offices, educational institutions, etc., can have a substantial negative influence on human health because the average person spends about 80–90% of their time indoors [42].

Because of their considerable effects on both air quality and human health, some essential pollutants stand out among the diverse range of contaminants. Among these significant contaminants are:

• PMs, including PM₁, PM_2.5, and PM₁₀, which refers to particles with a diameter less than 1 µm, 2.5 µm, and 10 µm, respectively, are linked to illnesses and fatalities. Reducing PM_2.5 levels from 35 µg/m³ to 10 µg/m³ could potentially decrease air pollution-related deaths by 15% [43,45]. PM_2.5 was the fifth-ranked mortality risk factor in the world and was responsible for 7.6% of all fatalities [46].
• Ozone (O3) is produced through photochemical reactions and plays a dual role in greenhouse gas emissions and its impact on human health and the environment. High concentrations of ground-level ozone can be particularly harmful. O3 exists as a gas both in the upper atmosphere (stratosphere) and at ground level. Stratospheric ozone is beneficial as it acts as a protective shield against ultraviolet rays. However, at the ground level and in the troposphere, ozone becomes a secondary air pollutant. It is formed through a series of intricate photochemical reactions involving solar radiation and ozone precursors [47].
• Nitrogen dioxide (NO₂) and sulfur dioxide (SO₂), produced from fuel burning [48], particularly in power plants and vehicles, are associated with respiratory issues and were responsible for 39% of NO_x emissions in Europe’s road transportation industry in 2017 [49]. These gasses are the primary acidic gases released by human activities. they not only contribute to the creation of acid rain and photochemical smog but also have detrimental effects on human health, vegetation, and materials [50].
• Carbon dioxide (CO₂), produced by burning fossil fuels, respiration, and natural processes, is a greenhouse gas contributing to global warming and pollution concentration, accounting for a significant percentage of emissions [35,51]. CO₂, as one of the greenhouse gasses (GHGs), plays a significant role in the global warming issue intertwined with industrial development in the globalized world. According to the current literature, the adoption of low-carbon practices is considered the most effective strategy for mitigating global warming. The combustion of fossil fuels by human activities is the primary source of CO₂ emissions, which greatly contributes to the creation of an environment conducive to global warming [52].
• Carbon monoxide (CO) is a hazardous gas emitted from various sources such as incineration, power plants, and urban road traffic. Inhalation of this gas can be fatal, as it converts to CO₂ in the atmosphere. CO poisoning is a prevalent form of toxicity in the modern world and is the leading cause of poisoning-related deaths in the United States. It is a highly toxic gas that lacks taste, odor, and irritants. Detecting CO is challenging due to these properties and the absence of a distinctive clinical signature, often mimicking other common disorders. CO is produced when hydrocarbons undergo incomplete combustion. Sources of CO include poorly ventilated garages with motor vehicle exhaust, as well as areas near garages. Combustion appliances can also generate CO when there is partial combustion of fuels like oils, coal, wood, kerosene, and others. A common scenario involves infrequently used and poorly maintained heating units [53].
• Methane (CH₄), mainly from natural gas and human activities like landfills and livestock, is another potent greenhouse gas. Methane contributes to the enhanced greenhouse effect. Methane production is a microbiological process, which is predominantly controlled by the absence of oxygen and the amount of easily [54]. CH₄ plays a significant role in intensifying the greenhouse effect, as it is approximately 20 times more potent than CO₂ on a molar basis. It is the second most influential greenhouse gas, following CO₂, and its overall impact, considering both direct and indirect effects on tropospheric ozone and stratospheric water vapor, is equivalent to about half of CO₂ [55].
• Volatile organic compounds (VOC), are considered significant contributors to air pollution, affecting the environment through both indirect and direct means. Indirectly, they act as precursors to the formation of ozone and smog. Directly, they pose toxicity risks to the environment. The rise of industrialization and urbanization has resulted in an increase in VOC emissions from various sources, both indoors and outdoors. These sources include the chemical industry, paper manufacturing, food processing, transportation, petroleum refineries, vehicle manufacturing, textiles, electronics, solvents, and cleaning products [56].

The World Health Organization (WHO) has taken significant steps by establishing thresholds and limits for the most significant pollutants (Figure 4).

The introduction of Air Quality Indices (AQIs) has simplified the process of assessing air quality. These indices offer a convenient way for individuals to understand the level of air pollution in their vicinity. AQIs are numerical values that correspond to different degrees of pollution, where higher numbers indicate poorer air quality. Although each country or region may have its own specific AQI standard, the Environmental Protection Agency (EPA) standard is widely adopted and encompasses six levels of health concern, ranging from” Good” to” Severe” [41] (Figure 5). The AQI is derived from measurements of five pollutants: SO₂, CO, O₃, NO₂, and PMs) [57]. The Indoor Air Quality Index (IAQ) is an important indicator used to assess indoor pollution. It takes into account various biological, chemical, and physical factors that contribute to indoor air quality, including CO₂, CO, VOC, and O₃, as well as physical factors such as temperature, humidity, and particulate matter. The IAQ provides a comprehensive measure of the overall quality of the air indoors by considering these different parameters [58]. This index serves as a measure of indoor pollution in enclosed spaces, where people spend approximately 90% of their time [59].

4.2. Supervised Learning Field

ML is a natural outgrowth of the intersection of Computer Science and Statistics [60]. A subfield of ML, supervised learning. It makes use of labeled datasets to train algorithms that can properly categorize data or predict outcomes. Its beginnings may be seen in the 1950s and 1960s when scientists created linear regression models to examine and forecast correlations between variables. In order to anticipate future results, supervised learning trains a model using input/output pairs or labeled data from the past. To accurately anticipate outputs for novel inputs, a function that can be roughly approximated must be developed. regression when we predict quantitative outputs, and classification when we predict qualitative outputs are two types of difficulties that might arise with supervised learning [61]. In the 1970s and 1980s, as computers became more accessible, researchers started to create increasingly complex supervised learning algorithms. As an illustration, decision trees (DT) were created in the 1970s as a technique to describe complicated decision-making processes [62]. An array of approaches exists, among which is the supervised ML algorithm developed by Vapnik, named SVM. This algorithm is versatile, catering to both classification and regression tasks. Another method involves the embedded technique called RF, which manifests as an assembly of unpruned classification or regression trees. In the context of the RF configuration, each separate tree produces a class prediction, and the ensemble makes its decision based on the class that receives the highest number of votes. ANN has made considerable progress in enhancing supervised learning by enabling the representation of complex non-linear relationships among variables. Various utilization techniques exist for ANN, including Recurrent Neural Network (RNN), LSTM networks, and Gated Recurrent Unit (GRU). These architectures, which fall under the umbrella of ANNs, are specifically tailored to handle continuous or sequential data. The resurgence of interest in supervised learning was ignited by the accessibility of large datasets and improved computational capabilities. Alongside this, DL algorithms were developed and applied across diverse fields by researchers, spanning domains like computer vision, and natural language processing. In supervised learning, a classification model uses known inputs to predict unknown outputs, specifically for categorical outputs. The dataset is divided into classes, and a classification algorithm learns from a training dataset to assign new data points to specific classes. Through a mapping function derived from the training data, the classification model can predict the class label for test data. The classification process involves data collection, preprocessing to eliminate noise and duplicates, and splitting the data into training and test sets using cross-validation. Once trained, the model can predict the class or label for new datasets, and its performance is evaluated using the test data. Classification can be binary, with two possible outcomes, or multi-label, dealing with multiple classes. In contrast, regression in supervised learning focuses on predicting continuous values based on variables, identifying correlations, and making predictions for continuous outputs. Regression can be categorized as simple linear regression, establishing a relationship between two variables with a straight line, or multiple regression, involving multiple variables and further classified as linear or non-linear regression [63].

4.3. Supervised Learning Approaches for Air Quality Analysis

The use of supervised ML in air quality analysis is widespread due to its effectiveness. There are two main approaches in this field: classification and regression models. Classification models categorize air quality based on pollutant readings, while regression models predict continuous values such as pollutant levels based on factors like time, location, and weather.

4.3.1. PM and Beyond: Exploring Pollutant Prediction in Air Quality Analysis

Particulate matter (PM) stands as a significant concern in the realm of air pollution, and numerous studies are devoted to employing ML models for PM-level forecasting. In one instance, high-resolution spatial air pollution maps of Charlotte, North Carolina were crafted utilizing linear regression and ML techniques [64]. In Ankara, Turkey, Bozdağ et al. introduced a hybrid model aimed at forecasting PM₁₀ concentrations. Multiple models, such as K-Nearest Neighbor (KNN), eXtreme Gradient Boosting (xGBoost), and Artificial Neural Networks (ANN), were employed in parallel to forecast PM10 concentrations [65]. In another study, a variety of techniques such as Adaptive Neuro-Fuzzy Inference System (ANFIS), Support Vector Regression (SVR), Classification and Regression Trees (CART), and Extreme Learning Machine (ELM) were employed to forecast PM10 and PM_2.5 levels in an industrial area. The results of this investigation indicated that ANFIS proved to be the most effective model [37]. Furthermore, a study that compared four algorithms for detecting peak PM_2.5 values found that Random Forest (RF) emerged as the most effective choice [46]. Expanding beyond PM, several studies have directed their focus toward other pollutants. One such study employed 16 alternative methods to predict annual average NO₂ concentrations and discovered that SVR and ANN exhibited lower accuracy in comparison to Generalized Boosted Machine (GBM), RF, and bagging [66]. Additionally, a dynamic indoor CO₂ model was developed using ML techniques like Support Vector Machines (SVM), Adaptive Boosting (AdaBoost), RF, GBM, Logistic Regression (LR), and Multilayer Perceptron (MLP) to manage the campus classroom’s HVAC system based on CO2 levels [67]. Finally, a novel non-linear air quality regression model, GaussODM, was introduced, drawing upon the Gaussian dispersion model. This model demonstrated superior performance when compared to both a simple regression model and a benchmark interpolation method for estimating geographic ambient NO_x concentrations [68].

4.3.2. Regression Techniques for Air Pollution Prediction

Regression techniques have proven to be highly effective in the domain of air pollution prediction. For example, indoor PM2.5 concentrations were predicted using both Multiple Linear Regression (MLR) and Random Forest Regression (RFR) [69]. In Colombia’s Aburrá Valley, a method for predicting pollutant concentrations utilized temporal features as input variables for an advanced Artificial Neural Network (ANN) and Support Vector Regression with Particle Swarm Optimization (SVR–PSO), achieving the best results [70]. More precise spatiotemporal Land Use Regression (LUR) models for pollutants like PM_2.5, PM₁₀, O₃, NO₂, CO, and SO₂ were created using mixed effect models and Least Absolute Shrinkage and Selection Operator (LASSO). These models outperformed previous LUR models, particularly at time scales of a day or longer [71]. ANN models have gained popularity in the prediction of air pollution and estimation of hospital admissions due to exposure to pollution. In a study focused on two major Brazilian cities, five ANNs were tested to estimate hospital admissions caused by PM₁₀ and meteorological factors [72]. Another study demonstrated improved one-step-ahead prediction results when combining ANNs and Multilayer Perceptron (MLP) models for PM10 and PM2.5 prediction [73]. Beyond ANN models, various other ML models, including Least-Squares Support Vector Machines (L-SVM), Gaussian Process Regression (GPR), RFR, and the PROPHET time series model, were evaluated in Bangladesh for monitoring PM and air quality. Among these models, GPR performed most favorably according to metrics such as R2, RMSE, and MAE [74]. Neural network models also showed lower error rates compared to linear regression models in forecasting CO concentrations [75].

4.3.3. Enhancing Air Quality Classification Methods

This study explores the advanced classification techniques for air quality analysis, addressing class imbalance and improving precision in assessing air quality. The introduction of the Adjusting Kernel Scaling (AKS) method, alongside classification algorithms like Adaboost, Multi-Layer Perceptron, GaussianNB, and SVM, demonstrates remarkable performance with an accuracy rate of 99.66% when applied to data from the Indian Central Pollution Control Board (CPCB), surpassing other classification methods [76]. Another study presents a laboratory study focusing on challenging multi-class classification problems related to indoor air contaminants. The study evaluates the proposed hybrid SVM (HSVM) model. It compares it with five existing methods, such as Euclidean distance to centroids (EDC), simplified fuzzy ARTMAP network (SFAM), multilayer perceptron neural network (MLP) based on back-propagation learning, individual FLDA, and single SVM, showcasing the HSVM model’s superior performance in addressing discrimination challenges in various electronic nose applications [77]. PCA was used to identify vehicular emissions and fuel combustion as significant pollution sources. Ensemble models, including Single Decision Tree (SDT), Decision Tree Forest (DTF), and Decision Tree Boost (DTB), are constructed and compared with Support Vector Machines (SVM). These models effectively discriminate seasonal air quality and predict air quality indices. The DT models show better performance in both classification and regression compared to SVM, credited to their use of bagging and boosting algorithms [78].

The authors of another study highlighted the use of machine learning algorithms, including Naïve Bayes, Random Forest, and K-Nearest Neighbor, for predicting PM₁₀ hotspots, a major air pollutant in Malaysia. These models offer effective tools for spatial PM₁₀ assessment, particularly in urbanized and industrialized areas with high PM₁₀ concentrations. The RF model’s output highlights high PM₁₀ concentrations in urbanized and industrialized areas, emphasizing the detrimental impact of air pollutants in urban regions. These models have the potential to support the Sustainable Development Goal (SDG) for Sustainable Cities and Communities by facilitating spatial PM₁₀ assessment and management [79]. Regarding the relationship between air pollution and COVID-19 infections, a Reduced-Space Gaussian Process Regression was employed to develop a classification model. This analysis reveals a correlation between high COVID-19 infection areas and elevated levels of NO₂ and PM₁₀, emphasizing the role of industrial factors and environmental conditions in the context of the pandemic [80].

4.3.4. Deep Learning’s Role in Reliable Air Pollution Forecasting

The significance of deep learning in the domain of air quality prediction has evolved into a transformative force, particularly within the specialized field of air pollution forecasting. The distinguishing hallmark of deep learning lies in its profound ability to uncover intricate patterns and interrelationships concealed within vast and complex datasets. This remarkable attribute has empowered deep learning (DL) models with an extraordinary capacity to not only predict air quality levels with precision but also to transcend conventional prediction methods. DL models bring a novel dimension to the table by extending their utility beyond mere accuracy; they contribute to the augmentation of forecast reliability. This is achieved through a unique capability—the assessment of uncertainty associated with each prediction, a feature that proves invaluable in air pollution forecasting.

To illustrate the practical application of DL in air quality prediction, consider the work of [81], which employed Long Short-Term Memory (LSTM), a modified RNN, to monitor PM_2.5 levels at various time stamps. Additionally, another study introduced the Convolutional Bidirectional Gated Recurrent Unit (CBGRU), a DL model tailored for short-term PM_2.5 forecasting [82]. Furthermore, the study [83] harnessed a Bayesian LSTM DL model to assess the impact of air quality regulations in Beijing, China. These instances underscore the versatility and effectiveness of DL models in addressing intricate challenges related to air quality prediction, with potential implications for mitigating the impact of air pollution on public health and the environment.

4.3.5. Enhancing Air Pollution Forecasting with Hybrid Models

The integration of hybrid models into air quality forecasting represents a promising approach, offering the potential to significantly enhance prediction accuracy and robustness in this critical field. These hybrid models ingeniously combine distinct methodologies, skillfully capitalizing on the complementary strengths of each component. A noteworthy illustration of this approach is demonstrated in [84], where researchers introduced the W-ANN hybrid model, skillfully merging the Wavelet transform method with traditional ANN models. The study’s findings revealed that the W-ANN model surpassed traditional ANN models in terms of forecasting accuracy, underlining the effectiveness of hybridization. Another compelling example of a hybrid model applied in the domain of air pollution forecasting is the ICEEMDAN-BPNN-ICA model, as presented by the authors in [85]. This multifaceted model seamlessly integrates three techniques: Intrinsic Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN), serving as a preprocessing step for data decomposition into intrinsic mode functions; Back Propagation Neural Network (BPNN), a versatile feedforward neural network employed to predict future values of each function; and Independent Component Analysis (ICA), a sophisticated signal processing technique adept at extracting independent components from the output of the BPNN. This comprehensive model was successfully employed to predict concentrations of various pollutants, including PM_2.5, SO₂, NO₂, CO, and O₃, thereby showcasing the immense potential of hybrid models in advancing air quality forecasting methodologies.

The findings of this section are listed in

Table 1. Supervised machine learning for air quality monitoring: a table overview.

	Source	ML Method	Predicted Value
01	Ravindra et al. [86]	RF, K-NN, LASSO, Decision Tree(DT), SVR Xgboost, DNN	Hospital admissions related to Acute Respiratory Infections
02	Dutta and Pal [87]	stacked-bidirectional long short-term memory (stacked-BDLSTM)	PM2.5, PM10
03	Van et al. [88]	DT, RF, XGBoost	AQI
04	Eren et al. [89]	LSTM, RNN, GRU	PM2.5
05	Barthwal [90]	Markov chain (DTMC) models	AQI
06	Wang et al. [91]	EML	Ozone
07	Persis and Amar [92]	NNs, SVM, DT, RF, XGboost.	AQI
08	Koo et al. [93]	DNN, RNN, Convolutional Neural Network (CNN),	PM2.5
09	Natsagdorj et al. [94]	Bayesian optimized LSTM, CNN-LSTM	PM2.5
10	Falah et al. [95]	RF, XGboost	PM2.5
11	Xie et al. [96]	Deep Learning-based Complex Trait Estimation Model(DL-CTEM)	NH3, CO2, H2S
12	Muthukumar et al. [97]	Convolutional Long Short-Term Memory (ConvLSTM), Graph Convolutional Network (GCN)	PM2.5
13	Abu El-Magd et al. [98]	RF	PM10
14	Huang et al. [99]	LR, RF, SVM, GPR, NN, ensemble tree	PM2.5, PM10
15	Gilik et al. [100]	CNN, LSTM	PM, NOx, SO2
16	Kumar and Pande [101]	Gaussian naive bayes (GNB), SVM XGBoost	AQI
17	Sethi and Mittal [102]	weighted naive bayes(WNB)	AQI
18	Abirami et al. [103]	SVR, Decision Tree Regression (DTR) RFR, MLR	AQI
19	Chen et al. [104]	DNN, LSTM	Ozone
20	Cheng et al. [105]	ResNet-LSTM	PM2.5
21	Cho and Moon [106]	ANN	CO2, PM10, PM2.5
22	Geetha et al. [107]	LSTM, RNN	SO2, CO2, NO2, CO, CFCs
23	Liu et al. [108]	SVM	AQI
24	Martín-Baos et al. [109]	LR, GPR, RF	AQI
25	Asha et al. [110]	Edited Nearest Neighbor (ENN)	NH3, CO, NO2, CH4, CO2, PM2.5
26	Ferreira et al. [111]	fuzzy ARTMAP	PMs
27	Choudhury et al. [112]	KNN, SVR, Hidden Markov Model(HMM) CNN,LSTM	NO2, O3
28	Qader et al. [113]	NNs, GPR	CO2
29	Magazzino et al. [38]	NNs, DT	Deaths
30	Mokhtari et al. [81]	CNN, LSTM	Propylene
31	De Mattos Neto et al. [73]	ANN, MLP	PM10, PM2.5
32	Wei et al. [114]	EMD-FUSION	SO2
33	Cihan et al. [37]	ANFIS, SVR, CART, RF, KNN, ELM	PM10, PM2.5
34	Taheri and Razban [67]	SVM, AdaBoost, RF, GBM, LR, MLP	CO2
35	K et al. [115]	LR	PM2.5
36	Cole et al. [39]	RF	PM2.5
37	Shahriar et al. [74]	L-SVM, GPR, RFR	PM2.5
38	Chang et al. [116]	Gradient Boosting Trees (GBT), SVR LSTM, LSTM2	PM2.5
39	Bozdag et al. [65]	LASSO, SVR, RF, kNN, xGBoost, ANN	PM10
40	Chang et al. [117]	LSTM, SVR, GBT	PM2.5
41	Magazzino et al. [118]	ANNs	Deaths
42	AlOmar et al. [84]	W-ANN	Ozone
43	Cazzolla Gatti et al. [40]	RF	Deaths
44	Han et al. [83]	LSTM	PM2.5
45	Shams et al. [75]	MLR, ANN	CO
46	Zeinalnezhad et al. [119]	ANFIS	SO2, O3, NO2, CO
47	Alyousifi et al. [120]	Multi-Wave Fuzzy Time Series (MWFTS)	API
48	Lu et al. [121]	Density-Based Spatial Clustering of Applications with Noise (DBSCAN), DNN	PM2.5
49	Tao et al. [82]	CBGRU	PM2.5
50	Chen et al. [66]	16 methods	NO2
51	Araujo et al. [72]	ELM, MLR, Radial Basis Function(RBF) Echo State Network(ESN),ENN	Hospitalizations
52	Murillo-Escobar et al. [70]	SVR–PSO	NO, NO2, O3, PM10, PM2.5
53	Zhou et al. [122]	GPM	NO2, HC
54	Yadav et al. [123]	CTSPD Algorithm	CO, Ozone, NO2, PM2.5, PM10
55	Jiang et al. [85]	ICEEMDAN-BPNN-ICA	PM2.5, SO2, NO2, CO, O3
56	Yuchi et al. [69]	MLR, RFR	PM2.5
57	Son et al. [71]	LASSO	PM2.5, PM10, O3, NO2, CO, SO2
58	Ketu et al. [76]	Adjusting Kernel Scaling (AKS)Adaboost, Multi-Layer Perceptron, GaussianNB, and SVM	AQI
59	Zhang, Lei, et al. [77]	hybrid SVM (HSVM)Euclidean distance to centroids (EDC), simplified fuzzy ARTMAP network (SFAM), multilayer perceptron neural network (MLP), individual FLDA, and single SVM	SO2, NO2, CO, CO2, NH3, O3, formaldehyde, benzene, toluene, inhalable particle, and VOCs
60	Singh et al. [78]	PCA, Single Decision Tree (SDT), Decision Tree Forest (DTF), Decision Tree Boost (DTB)SVM	AQI
61	Tella, Abdulwaheed, et al. [79]	Naïve Bayes, Random Forest, and K-Nearest Neighbor	PM10
62	Velásquez et al. [80]	Reduced-Space Gaussian Process Regression	NO2, PM10

. This table provides a comprehensive summary of the results obtained, emphasizing the key ML techniques utilized, and parameters predicted. Two graphical representations, namely Figure 6 and Figure 7, are provided in the study. They display the predicted shares of pollutants as outlined and the proportion of methods employed in the review, respectively.

5. Challenges and Future Directions

5.1. Findings, Limitations, and Challenges in Air Quality Research

The analysis of the existing research has highlighted the significance of exploring the link between COVID-19 fatalities, economic expansion, and air pollution. Several studies indicate that air pollution, particularly PM₁₀ and PM_2.5, may play a significant role in COVID-19-related deaths. The application of ML techniques was identified as an effective approach to forecasting air pollution and minimizing its negative impact on public health. However, it is worth noting that the studies included in the literature review were conducted in specific regions, and their conclusions may not apply to other locations. Further studies are necessary to gain a thorough understanding of the relationship between COVID-19 mortality, economic growth, and air pollution across various regions. The literature review also found that PM, NO_x, SO₂, O₃, CO, and CO₂ are frequently identified as pollutants of concern in air quality assessments. It indicates that these pollutants are causing significant apprehension and are expected to attract more attention and stricter regulations. This knowledge holds great significance for policymakers and researchers as it helps them direct their endeavors toward decreasing the concentrations of these pollutants and enhancing air quality on the whole. This review reveals also that in the domain of air quality research, certain algorithmic approaches have emerged as particularly prominent and widely employed. Among these, algorithms such as LSTM, RF, ANN, SVR, XGBoost, DT, Convolutional Neural Networks (CNN), and Gaussian Process Regression (GPR) stand out.

Temperature and humidity also play a vital role in air quality research. They influence how pollutants behave and their concentrations in the air. Factors like chemical reactions, dispersal patterns, the respiratory system of humans, and the functionality of air quality monitoring devices are all affected by temperature and humidity levels. It is important to consider that the choice of pollutants studied and predicted depends on the location and specific sources of pollution under investigation. For example, in heavily industrialized areas, studies may focus on predicting levels of PM and NO_x, while in rural regions, attention may be directed toward predicting levels of ozone and carbon monoxide. Therefore, it is crucial to account for the local context when designing and implementing IoT-based air quality monitoring systems to effectively address and mitigate pollution concerns.

A key challenge revolves around the necessity for more robust and reliable sensors. IoT sensor networks heavily rely on numerous low-cost devices, which can be susceptible to failures and inaccuracies. Consequently, this can lead to data gaps and inaccurate measurements, ultimately affecting the effectiveness and reliability of ML models. To overcome this obstacle, researchers should prioritize the development of more resilient and dependable sensors, alongside the formulation of approaches to handle missing data and sensor errors. Another significant challenge pertains to the concerns surrounding privacy issues and security [124]. IoT sensor networks collect and transmit vast amounts of sensitive data, making them vulnerable to hacking and various cyber-attacks. These breaches can lead to the compromise of sensitive information and even manipulation of data, ultimately impeding the effectiveness of ML models. To address this issue, researchers should concentrate on developing secure and privacy-preserving techniques for data collection, storage, and transmission. Researchers commonly rely on classification and regression methods when analyzing air quality data. This preference can be attributed to the fact that these approaches yield results that are more easily understood, in comparison to clustering and dimensionality reduction techniques. The ability to comprehend the contribution of various factors to air quality outcomes, as well as the impact of interventions, holds paramount importance in this field of study. One potential critique concerning the utilization of supervised ML techniques in air quality research is their potential inability to capture the intricacies and variability present in the data, potentially leading to inaccurate predictions. Moreover, these supervised techniques heavily rely on labeled data, which can be challenging to acquire, and they often require significant computational resources. Ensemble methods, including RF, XGBoost, AdaBoost, Gradient Boosting, GBT, and ensemble trees, have proven to be effective in air quality prediction. However, their effectiveness varies depending on the specific dataset and problem at hand. On the other hand, techniques such as decision tree-based methods (DT, CART) and Naive Bayes (Gaussian Naive Bayes) have shown less robust performance. DL techniques, such as Deep Learning-CTEM, ConvLSTM, GCN, and CNN+LSTM, were also proposed. Nonetheless, it is worth noting that these approaches can be computationally expensive and may not be practical for real-time predictions. Furthermore, these advanced techniques, such as neural networks (NN, MLP, NNs, LSTM, ELM) and hybrid models (EMD-FUSION, ANFIS), exhibit commendable performance in air quality analysis. However, their implementation and interpretation can pose challenges, alongside the need for larger datasets and computational resources. It is important to acknowledge that these approaches may lack interpretability and offer limited insights into the underlying mechanisms that drive air quality.

5.2. The Role of ML Models in Mitigating Climate Change and Air Pollution: A Sustainable Development

According to the Global Sustainable Development Report 2023, addressing air quality is a critical component of the Sustainable Development Goals (SDGs), particularly those related to environmental conservation and sustainability. The report emphasizes that efforts to protect and restore terrestrial ecosystems, a cornerstone of SDGs, are intrinsically linked to air quality improvements. Preserving air quality not only supports goals related to life on land, such as SDG 15, which focuses on halting biodiversity loss, but it is also essential for achieving SDG 3, which centers on good health and well-being. In the context of data and ML, the Global Sustainable Development Report underscores the importance of data-driven strategies and innovative technologies in enhancing air quality control. The report advocates for leveraging ML to advance data analysis and predictive modeling for more effective air quality monitoring and mitigation efforts. By embracing ML and data technologies, as highlighted in the report, it is possible to make well-informed decisions and implement policies that effectively address air pollution, mitigating its adverse effects on human health and the environment. As we progress through the second half of Agenda 2030, these novel partnerships and data technologies are integral in ensuring clean air for all, aligning with the SDGs, and enhancing the quality of life and the environment, as emphasized in the Global Sustainable Development Report 2023 [23].

Artificial intelligence and machine learning technologies play a pivotal role in addressing the challenges posed by climate change. Both climate change and the field of artificial intelligence are intricate subjects of academic exploration. The environmental consequences of ML, both in a positive and negative sense, are currently a subject of in-depth analysis. The United Nations has established a set of 17 Sustainable Development Goals (SDGs) that encompass environmental, social, and economic aspects. AI has the potential to reduce global emissions of greenhouse gasses by up to 4%, thereby assisting in alleviating the impact of climate change. AI methodologies have the capability to enhance the accuracy of forecasting tools utilized for predicting and evaluating extreme environmental occurrences and for analyzing long-term climate change data. Consequently, AI can serve as a crucial tool in the battle against climate change [125].

The reported potential effects of AI encompass both positive and negative ramifications on sustainable development. Nevertheless, as of now, there is no published research that systematically examines the extent to which AI could influence all facets of sustainable development, as defined in this investigation, encompassing the 17 Sustainable Development Goals (SDGs) and the 169 internationally agreed targets outlined in the 2030 Agenda for Sustainable Development. A substantial body of relevant evidence indicates that AI might serve as a facilitator for achieving 134 targets (approximately 79%) spanning all SDGs, primarily using technological enhancements that have the potential to overcome existing limitations. However, the development of AI may have an adverse impact on 59 targets (about 35%), also distributed across all SDGs [126].

Looking ahead, while ML offers substantial potential to mitigate pollution and carbon emissions through enhanced optimization of industrial processes and transportation systems, their implementation must be approached sustainably to prevent inadvertent increases in carbon footprints. ML, in particular, can have a negative impact on the environment through energy consumption in the training and use of large models, the operation of data centers, and the manufacturing of specialized hardware such as GPUs. These activities contribute to carbon emissions and other types of pollution [127]. The findings of various studies demonstrate that AI has a negative impact on carbon intensity. For example, a study in China presents the first evidence of a relationship between AI and carbon intensity, supporting the notion that AI has a positive effect on carbon emissions [128]. Another study reveals that CO₂ emissions increase with the number of Edge-AI G-IoT devices deployed and with the greenness of a country’s energy production [129]. The use of electronic devices and the infrastructure required to support them can contribute to more than CO₂ emissions as mentioned [130]. Data centers and high-performance computing facilities are estimated to contribute 100 megatonnes of CO₂ emissions per year, similar to the amount emitted by commercial aviation in the US [130]. This significantly contributes to climate change. Furthermore, the McKinsey Global Institute estimates that AI could potentially increase global GDP growth by 1.24% per year by 2030. However, there is a lack of consensus in the literature on the relationship between ICT and CO₂ emissions, with a viewpoint falling into three categories: ICT can promote economic development in an environmentally friendly manner, ICT is an energy consumer, and the relationship between ICT and CO₂ emissions is uncertain [128]. Studies have shown that energy savings in data centers can range from 154% to 604% depending on the specific scenario, reducing CO₂ emissions [131]. Therefore, it is crucial to evaluate the environmental consequences of integrating AI and IoT and take steps to mitigate their negative influence. The energy consumption, materials, electronic waste, and transportation associated with IoT contribute significantly to pollution. To alleviate the environmental impact of AI, it is essential to emphasize renewable energy, sustainable design, recycling of electronic waste, and energy efficiency. By adopting these strategies, we can reduce the ecological impact of these technologies while retaining their advantages.

5.3. Future Directions and Open Perspectives in Urban Planning for Air Quality Research

The future of air quality research holds great promise, with potential advancements in the field of machine learning (ML) and data-driven approaches. Beyond the current capabilities of supervised and unsupervised methods, one notable avenue for exploration is reinforcement learning, a concept receiving significant attention in various domains. This approach, known for its capacity to learn through interaction with the environment and optimize actions to maximize rewards, has the potential to revolutionize air quality research. By training agents to make real-time decisions based on air quality data and feedback loops, reinforcement learning can be employed to dynamically optimize interventions and policies, particularly in complex urban environments with variable pollution sources.

Semi-supervised learning offers another avenue for future research. Combining the strengths of labeled and unlabeled data, it is particularly well-suited for scenarios where acquiring large amounts of labeled data is challenging and cost-prohibitive. In the context of air quality research, where obtaining labeled data for various pollutants across diverse locations can be a formidable task, semi-supervised learning presents a viable solution. By effectively harnessing both labeled monitoring station data and unlabeled information from satellite imagery or sensor networks, semi-supervised learning models can potentially offer more precise and comprehensive insights into air quality patterns and the underlying drivers of pollution. These future research directions are poised to enhance our understanding and management of air quality, contributing to the well-being of both the environment and human health.

Regarding optimizing urban layouts, the potential of AI and data analytics to reshape urban landscapes is truly transformative. Urban planners now have the tools to conduct nuanced analyses aimed at reconfiguring the spatial arrangement of urban features. This includes a strategic reevaluation of the placement of residential zones, industrial areas, and traffic arteries. The overarching objective is to tactically reduce the proximity of sensitive receptors, such as residential communities, to pollution sources, thereby fostering cleaner urban environments. The impact of AI-driven algorithms extends to the relocation of industrial facilities away from densely populated areas, a move that not only mitigates harmful emissions exposure but also contributes to the enhancement of air quality. This facet underscores the profound influence of AI on urban planning, with air quality improvement at its core.

The integration of AI into the domain of smart transportation systems presents an auspicious path toward revolutionizing urban mobility and, consequently, elevating air quality. AI-infused traffic management solutions hold the potential to alleviate congestion, a well-documented contributor to air pollution in urban settings. Furthermore, the development of electric and autonomous vehicles, under the guidance of AI technologies, ushers in an era of reduced emissions, constituting a significant boon for air quality. Implementing predictive modeling further accentuates the positive impact of AI by facilitating traffic flow optimization, ultimately mitigating idling time and the pollution associated with congestion. As urban areas continue to grapple with the multifaceted challenge of air quality, AI-driven solutions in transportation systems stand out as a promising avenue for meaningful change, ensuring cleaner and healthier urban environments for residents.

6. Conclusions

In conclusion, this comprehensive review provides a nuanced analysis of the current landscape about the utilization of supervised machine learning in the fields of air quality management and climate change mitigation, with a keen focus on its alignment with the Sustainable Development Goals (SDGs). Through a meticulous examination of a variety of scholarly works, it becomes evident that these technological advancements hold great promise in elevating the precision and efficiency of air pollution monitoring and predictive capabilities. By harnessing advanced algorithms in conjunction with sensor networks, supervised learning has distinctly showcased its potential to uncover critical insights into the sources and intricacies of air pollution. The models generated through data-driven methodologies offer a heightened degree of accuracy, thereby facilitating timely interventions aimed at ameliorating environmental and public health risks. It is imperative to acknowledge, however, that certain challenges persist, most notably those about data quality and the interpretability of models. Despite these challenges, the integration of supervised learning undeniably represents a proactive and forward-looking approach to environmental stewardship. The confluence of cutting-edge technology and environmental imperatives has led to a more informed and efficient approach to the management of air quality.

In summary, the intersection of pioneering technology and environmental priorities has paved the way for a more enlightened and effective approach to air quality management and climate change mitigation. This journey towards a cleaner and healthier future remains reliant on sustained interdisciplinary collaboration and the continuous refinement of methodological approaches. The path forward, in alignment with the principles of the SDGs, holds the promise of a sustainable and improved future.