In recent years, the industry's fast growth has been accompanied by air pollution, which kills millions of people yearly and gets widespread attention (Guo et al., 2020). According to the World Health Organization (WHO), about 90% of people breathe air that is contaminated and violates WHO air quality criteria (Bekkar et al., 2021; World Health Organization, 2021). Air pollution is a worldwide health issue, causing respiratory disorders, lung problems, eye problems, and skin diseases in people and affecting the ability of plants and animals to thrive. As a result, air pollution control and prevention have become major concerns. Factories' smoke exhaust, pollution caused by vehicles' exhaust, and power plants are the primary causes of air quality degradation (Sultana, 2019). (PM2.5, PM10), O3, SO2, CO, and NO2 are the five categories of air pollutants (Mao et al., 2021). PM2.5 is the most concerning air pollution component because these particles are small and light. They can stay in the atmosphere longer and easily bypass the filters in the human nose and throat (Akiladevi et al., 2020). PM2.5 is a standard air quality metric. However, it is usually measured with ground-based sensors (Jonathan et al., 2020). Many researchers focus on air pollution because of its increasing attention, and numerous important research papers are on it. Due to population and economic expansion, global energy consumption is steadily growing (Heydari et al., 2021).

Traditional statistical approaches have been frequently applied to solve air quality forecasting difficulties. These strategies are based on the principle of using historical data for learning; however, owing to the time-series data complexity and variance, they can produce poor estimates of air pollution. Several machine-learning algorithms have been developed during the last 60 years to aid in the resolution of complexity concerns (Ameer et al., 2019). Ensemble learning, MLR, SVM, RF, ANN, and other hybrid models are the primary machine learning approaches to combat air pollution (Bekkar et al., 2021). However, because the model selection is the focus of most prediction approaches and reasons for the change in air pollution concentrations are not analyzed by most present air quality prediction machine learning methods (Ameer et al., 2019). Furthermore, since contemporary deep learning frameworks are relatively adaptable, the model may need to be deep and sophisticated to match the Dataset. As a result, many weights in a deep neural network model may cause overfitting difficulties.

To assess and forecast the behavior of Particulate Matter (PM2.5), this study presents a Machine Learning approach that includes Cat Boost Regressor, Extreme Gradient Boosting Regressor, Random Forest Classifier, Logistic Regression, Support Vector Machine, K-Nearest Neighbor, and Decision Tree. This study summarizes the procedure of these methods to estimate the best solution for the corresponding requirement in any circumstance, to forecast air quality to raise public awareness about air quality degradation and its health effects.

The rest of the paper proceeds as follows: Section 2 presents the literature review, Section 3 presents the methodology used for the study, Section 4 presents the experiment and results, Section 5 shows the discussion of results, and Section 6 compares this work with existing research. Finally, Section 7 concludes the paper with a summary of the main points, future directions, and the study's limitations.

According to Liao et al. (2020), no studies with complete adequate long-time intervals that include pollutant measurements from all sources, CTM (Chemistry-Transport Models), data assimilation products, driving meteorological fields, and emission sources. As a result, to progress, it will be required first to create such extensive benchmark datasets for testing learning algorithms and designing deep network topologies. They examined studies on methods such as RNN, LSTM, GRU, CNN (Convolutional Neural Network), SAE (Sparse Autoencoder), and DBN (Deep Belief Network) for Air Quality Forecasts in this paper. Finally, they determined that dealing with meteorological factors and pollution measurements from ground-level monitoring networks limits deep-learning research for air quality forecasts. They looked at attempts to use deep learning techniques to overcome the limitations of standard air quality forecasting methods that use chemistry-transport models (CTMs) or shallow statistical methods.

Ameer et al. (2019) studied and compared four current methods for predicting air pollution in smart cities in Machine Learning Techniques for Predicting Air Quality comparative analysis. The methods were RF regression, GBR, DT (Decision Tree) regression, MLP (Multi-Layer Perceptron) regression, and RF regression emerged as the best. They identified which of the compared techniques used to predict Air Pollution is the best. They did not discuss data handling. Sultana compared air pollution detecting techniques using image processing, machine learning, and deep learning approaches, where they evaluated these three methods used to detect air pollution and better compare estimates, how they operate, and are processed in the air pollution detection (Sultana, 2019). Finally, they determined that the deep learning technique outperforms the other two regarding efficacy and accuracy. However, it necessitates a large dataset, and as the accuracy level rises, so does the total expenditure and cost. They considered three procedures (Image Processing, Machine Learning, and Deep Learning) used to detect air pollution and estimate a better comparison of how they work and are processed in air pollution detection. Data implementation was not discussed (Sultana, 2019).

Guo et al. developed an EN model to forecast PM2.5 concentrations based on previous PM2.5 concentrations, meteorological data, and time stamp data. RNN, GRU, LSTM, and NN (Neural Network) were among the optimum algorithms employed. Human activities and topographical data were missing from the study (Guo et al., 2020). The findings showed that the suggested technique beats existing algorithms in terms of performance. Mao et al. used graph convolution and LSTM networks to create and present a spatiotemporal modeling hybrid deep learning framework to forecast various air contaminants (Mao et al., 2021). Models such as MLR and LSTM networks were employed. The findings revealed that the distribution of errors in space, to some extent, corresponds to the spatiotemporal correlation strength distribution, highlighting the necessity of spatiotemporal dependency modeling for pollutant prediction. They did not discuss data implementation. Heydari et al. (2021) anticipated and assessed air pollution from Combined Cycle Power Plants by creating a novel hybrid intelligence model based on MVO (Multi-Verse Optimizer) algorithm and LSTM. They applied the method only to observe the correlation coefficient of NO2 and SO2 pollutants.

Xayasouk and Lee proposed a deep-learning-based technique for fine dust prediction. They utilized the deep-learning algorithm to construct a spatiotemporal prediction framework that considers the Dataset's temporal and geographical relationships during the modeling process (Xayasouk and Lee, 2018). To train and evaluate the data, they employed the Stacked Encoders model, which is unsuitable for learning and training the time series data (Xayasouk and Lee, 2018). Abdellatif et al. created a CNN-LSTM that can be utilized to estimate air quality and can efficiently conduct Spatiotemporal prediction (Bekkar et al., 2021). Deep learning models such as LSTM, CNN, GRU, CNN-GRU, CNN-LSTM, Bi-LSTM, and RNN were utilized (Bekkar et al., 2021). The model can efficiently extract data from temporal and spatial aspects using CNN and LSTM, and it also has excellent accuracy and stability, according to the findings of this work. They did not discuss the processing time. Aarthi et al. (2020) stated that Environmentalists and the government aided in framing air quality standards and regulations based on hazardous and pathogenic air exposure and health-related risks to human welfare. The processed datasets were used to generate a function that plots the training and validation data for several models, including SV (Support Vector), Lasso, Linear, and DT regression. The authors found that their project raised public awareness, assisted environmentalists and the government in creating air quality standards and regulations based on hazardous and pathogenic air exposure and health-related dangers to human welfare, and discussed the health effects of air quality degradation. They used a decision tree in this experiment, which is not a suitable classifier for time series data (Aarthi et al., 2020).

Aditya et al. (2018) suggested an approach that would assist ordinary people and meteorologists in detecting and forecasting pollution levels and responding appropriately. Logistic Regression and Autoregression were employed as machine-learning regression approaches. This will also assist individuals in establishing a data source for small towns, which are sometimes overlooked compared to major cities. Logistic Regression performed well on a prediction but failed to explain the constraints (Aditya et al., 2018). Balasubramanian et al. (2021) developed a technique to anticipate the following 5 h' Air Quality Index. They employed a Linear regression model, an SV regression Model, and RF regression Model for data analysis. According to the researchers, Machine Learning algorithms were used to anticipate the AQI (Air Quality Index) values for the following 5 h (Balasubramanian et al., 2021). The Stacking Ensemble model has the lowest RSME (Root Mean Squared Error) value when all the models' RMSE (Root Mean Square Error) values are compared. As a result, this model was picked to anticipate the following 5 h' Air Quality Index. They did not thoroughly discuss data handling. Dobrea et al. developed a technique that calculates the number of atmospheric pollutants (PM2.5 and PM10) (Dobrea et al., 2020). Support Vector Regression, Autoregression Integrated Moving Average, and LSTM are the models employed. After a comparison of data analysis methods and Machine Learning algorithms for estimating atmospheric pollutants (PM10 and PM2.5), it was determined that the Support Vector Regression and ARIMA (Auto Regressive Integrated Moving Average) algorithms are the most suitable for forecasting air pollutants concentrations, with correlation coefficients of 96.6% and 92.1% for PM10 and PM2.5, respectively (Dobrea et al., 2020). The experiment only focused on one factor of air pollution.

Akiladevi et al. (2020) proposed a technique for developing an air quality forecasting system that can anticipate main contaminants in various locations. To assess the Dataset's performance, ML (Machine Learning) methods such as LR (Linear Regression), NB (Naïve Bayes), SVM, RF, KNN (K-Nearest Neighbor), and DT were utilized. Performance measurement factors such as accuracy, recall, f1-score, Specificity, and Sensitivity were computed for each method. For each technique, confusion matrix parameters such as TP (True Positive), TN (True Negative), FP (False Positive), and FN (False Negative) were determined. LR had a 98% accuracy, NB had a 95% accuracy, RF had a 99% accuracy, SVM had a 70% accuracy, K-NN had a 97% accuracy, and DT had a 100% accuracy. Out of these six ML algorithms, the Decision Tree approach had the best accuracy (Akiladevi et al., 2020). The decision Tree was not a good time series data classifier, so it performed well in this research. Bui et al. (2018) proposed a deep learning technique for air quality index predictions. The Encoder-Decoder paradigm was employed, as well as Long Short-Term Memory units. Based on historical meteorological data, their suggested model produced substantial results in predicting PM2.5 AQI for the long term. The accuracy was discussed but not the processing time.

Taylan et al. (2021) mentioned that to minimize respiratory and cardiovascular deaths, researchers developed a method that is feasible, robust, and capable of evaluating pollutants' cumulative effect inside metropolitan areas. They employed the Non-linear Autoregressive with External (NARX) Input and the Levenberg–Marquardt (LM) Algorithm. They concluded that managing air pollution entails establishing capacity and monitoring ground-based networks and systems to make suitable strategic and operational decisions. Quality assurance and control, modeling methodologies, and institutional competencies are all required to implement these initiatives. The Dataset used was limited.

Kalajdjieski et al. (2020) developed a data fusion method for using multi-modal data such as weather and pollution measurements obtained by sensors and picture data collected by cameras. Basic Convolutional Neural Network, Residual Network Model, Inception Model, and Custom pre-trained Inception were among the predictive models tested. Their trials reveal that our bespoke pre-trained inception model, paired with their data preparation strategy, outperforms known state-of-the-art approaches in accuracy (Kalajdjieski et al., 2020). The model used was biased. Saleh et al. (2016) developed a model for predicting CO2 emissions from energy. The Support Vector Machine model was utilized. They concluded that a lower RMSE (Root Mean Square Error) value must be produced when the prediction model's accuracy is good. It can assist the management in developing policies or making decisions to limit the negative environmental impact throughout the manufacturing process by monitoring energy use. The experiment only focused on CO2 (Saleh et al., 2016).

Popa et al. developed a system model that forecasts temperature changes in a densely populated area of Bucharest, Romania. They employed LR, SVM with Gaussian kernel, and Gaussian process regression with the exponential kernel as well as other techniques (Popa et al., 2021). They concluded that future studies might combine the current findings with camera photos to assess and anticipate air pollution in various large cities or establish a platform to provide traffic suggestions based on air pollution predictions. They only used linear methods for classification.

Based on the reviewed literature on Machine Learning Applications in Air Pollution. To the best of our knowledge, no work was done involving the analysis and prediction of air pollution in South Africa. Many have been done in countries like China (Moursi et al., 2019; Guo et al., 2020; Harishkumar et al., 2020; Balasubramanian et al., 2021; Bekkar et al., 2021; World Health Organization, 2021), India (Aditya et al., 2018; Sultana, 2019; Aarthi et al., 2020; Akiladevi et al., 2020; Masood and Ahmad, 2020), Korea (Bui et al., 2018; Xayasouk and Lee, 2018; Yang et al., 2020), and Iran (Zamani Joharestani et al., 2019). The proposed method in this study will analyze and predict the behavior of PM2.5, monitor a period of historical levels and correlation analysis for future predictions of PM2.5 levels in cities of South Africa and evaluate the models used to find the best that will be used to measure the performance of the Dataset.

This study used the Anaconda Navigator (Jupyter Notebook) and an AMD Ryzen 7 5700U computer with 8GB of RAM and a 1.80 GHz Radeon graphics processor. Python 3.6 exposed the proposed machine learning models to data cleaning and feature extraction for training and testing models. This study aims to investigate various machine learning approaches to air pollution, to analyse and predict air pollution behavior in terms of air quality and air pollutants (PM2.5), detecting which areas are most affected in South African cities. All the graphs in this chapter are created using Python. The data was handled using Pandas, and the charts were plotted with Matplotlib and Seaborn.

Evaluation Models used for predicting PM2.5 concentrations for South African Cities.

The likelihood of PM2.5 surpassing the healthy level is predicted using regression models. Two regression models, Cat Boost Regressor and Extreme Gradient Boosting Regressor, were implemented for the PM 2.5 prediction. These models achieved reasonably good Accuracy scores of over 0.6 and were both often correct for over 0.9 of the time. To predict the Air Quality Status, K-Nearest Neighbor, Logistic Regression, Support Vector Machine, Decision Tree, and Random Forest Classifier, five classification models were implemented with an excellent Accuracy Score of over 0.98 to predict when provided the PM2.5 Level.

To assess the high-level relevance of traits, the Mean RMSE of all models used is compared, and the actual is compared to the predicted. The lagged inputs played a significant role in predicting the PM2.5 and the AQI status, as many of them were selected and used by models when predicting. According to the results, Cat Boost Regressor was the best model to predict PM2.5. Furthermore, for AQI status, Random Forest Classifier and Decision were equally the best.

In this study, SVM, Random Forest, and KNN performed better with accuracy of 98.88%, 100%, and 99.99%, respectively, compared to the same models by Akiladevi et al. (2020), which achieved the accuracy of 70%, 99%, and 97%, respectively. The Decision Tree performed best in both cases, with an accuracy of 100%.

Cross-validation, XGB, the second fold, had the highest RMSE of 39.86 compared to the XGB used by Zamani Joharestani et al. (2019), which achieved 13.58.

Gupta et al. included models with an accuracy of 99.88% for CatBoost regression, 92.40% for SVM, and 91.99% for Decision Tree. In contrast, this study has the accuracy of CatBoost regression, 98.88% for SVM and 100% for Decision Tree.

Generally, the models used in this work perform better on our datasets when compared to existing works using similar models, as shown in Table 4.

This study focused on predicting the concentration of PM2.5 pollutants in South African cities. The proposed machine learning models are intended to forecast the probability that PM2.5 would surpass the established threshold or not. At various heights above the ground along a vertical axis, meteorological data and air pollutant PM2.5 features are carefully considered. The forecasting ability of the models may be improved by incorporating other characteristics into Google Earth Engine that further extract meaningful information from the data. A higher forecast performance may be possible if more extensive and reliable data are provided. More complex models, like deep learning techniques, may improve prediction accuracy with a larger dataset.

Several models were used, and regression models used included Cat Boost Regressor and Extreme Gradient Boosting Regressor; the performance measure used is an RMSE (Root Mean Square Error). Classification models included K-Nearest Neighbor, Logistic Regression, Support Vector Machine, Decision Tree, and Random Forest Classifier, which were compared using the MSE (Mean Square Error), MAE (Mean Absolute Error), and RMSE (Root Mean Square Error) parameters for predicting the Air Quality Index (AQI) Status. The results show that the proposed hybrid model is more accurate than the solo models, proving its superiority. The suggested method can be used in the future to forecast data from other cities. Using prediction, we may also identify the polluted area and its root cause. Some pollutants pose a severe threat to human health in the future.

The data used in this investigation is static. Interestingly, the site offered daily updates to the data. Leveraging real-time data analysis through the cloud to create better results for improved performance shall be considered in the future extension of this work. Moreover, the models used in this work will be evaluated on more datasets from Nitrogen Dioxide (NO2), Ozone (O3), Sulfur Dioxide (SO2), Carbon Monoxide (CO) pollutants. Furthermore, Deep learning methods and Ensembled methods shall be consider for PM2.5, PM10 and other pollutants indicated above.

Not all the South African cities were included in the Dataset. This is because the ones included are the ones that are only having the stations. Even though it was possible to make predictions of the selected cities, the comparison could not be made for all the cities in South Africa since there are no recorded readings for some cities.

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author.

Study conception and design, analysis and interpretation of results, and draft manuscript preparation: IO and TM. Data collection: TM. All authors reviewed the results and approved the final version of the manuscript.