Air Quality and Pollution Assessment Using Machine Learning Techniques

1. Introduction

This study emphasizes the importance of mitigating the harmful effects of environmental pollution and conducting an in-depth examination of contributing factors. While the complete elimination of pollution is not feasible, understanding its causes and improving monitoring systems remains a key objective [11]. The research applies machine learning techniques to an updated pollution dataset. Among the tested models, Random Forest achieved the highest accuracy (95%), outperforming other classifiers [12,13,14]. The dataset was pre-processed to eliminate missing and irrelevant data. Performance was evaluated using accuracy, precision, recall, and F1-score. Results indicate that industrial proximity and population density significantly impact air quality.

2. Literature Review

To overcome extensive data gaps in satellite aerosol optical depth (AOD) retrievals, Li et al. proposed a multimodal data fusion system combining simulations, satellite, and ground-based measurements using optimal interpolation. Their model produced high-resolution AOD maps with strong correlation (R = 0.83). Bhatti et al. [1] examined PM2.5 and PM10 in Lahore using SARIMA and HYSPLIT to identify pollution sources, noting coal burning and traffic as key contributors.

Nguyen et al. [2] proposed a deep learning air quality forecasting framework integrating Wasp-based interfaces and cross-assorted DL models, achieving high accuracy across Australian regions. Abboud et al. [3] tackled limitations of static air quality monitoring by combining CNN-LSTM with mobile monitoring data, improving spatial coverage (83.3% in Paris).

Rodrigues et al. [4] investigated European cities’ compliance with air quality laws, analyzing 10 years of data and showing policy interventions reduced violations. Ntesat et al. [5] highlighted toxic metal pollution from fireworks in indoor air and used atomic spectroscopy to identify exposure risks to children.

Sannoh et al. [6] emphasized the persistent air pollution in Karachi due to insufficient monitoring and analyzed risk through PM2.5 levels and PMF models. Liu and Cui [9] evaluated control measures in China’s GBA region using WRF-CMAQ and SMAT-CE modeling, noting significant pandemic-related pollution declines. Liu et al. [10] proposed LightGBM and LSTM models to overcome forecasting limitations [8], achieving 97.5% accuracy. Table 1 summarizes key studies from literature [15,16,17,18].

3. Proposed Methodology

Our proposed methodology is shown in Figure 1. We implemented several ML models using RapidMiner:

(a)

KNN: Classifies data based on similarity to nearby data points.

(b)

Decision Tree: Tree-based model with branching decisions for multi-class classification.

(c)

Random Forest: Ensemble method using multiple decision trees.

(d)

Naive Bayes: Probabilistic classifier effective for binary and multi-class problems.

Figure 1. Paper Methodology.

Figure 1. Paper Methodology.

The dataset used in this research work includes 5000 records with 10 attributes. Table 2 shows dataset description.

4. Results

We used 10-fold cross-validation and stacking-based ensemble modeling. Five base models (Naive Bayes, Gradient Boosted Tree, KNN, Decision Tree, Random Forest) were combined. The meta-classifier (Random Forest) achieved the best results. Figure 2 shows machine learning Framework flow diagram.

Figure 3 illustrates a machine learning workflow for a cross-validation process, as displayed in the RapidMiner software. The workflow is split into training and testing phases.

Figure 4 illustrates a stacking ensemble learning process. The workflow begins with a set of “Base Learners” on the left, which are individual machine learning models. In this specific example, the base learners include Naive Bayes, Gradient Boosted Tree, Decision Tree, and Random Forest.

Figure 3. Cross-Validation Technique Diagram.

Figure 3. Cross-Validation Technique Diagram.

Figure 4. caption.

Figure 4. caption.

4.1. Dataset Description

The original dataset contained various air pollutant measures such as PM2.5, PM10, NO2, CO, SO2, and O3. The dataset also included weather-related parameters such as temperature, humidity, wind speed, and pressure. These attributes served as input features for the classification and regression models [19,20,21].

4.2. Performance Comparison of Algorithms

A comparison was made between several machine learning models including Decision Tree, Random Forest, KNN and SVM. The models were tested on both the original and preprocessed datasets. Preprocessing involved handling missing values, normalization, and feature selection to improve model performance [23,24].

Table 3 presents the classification accuracy [25] of different algorithms on the original and preprocessed datasets.

As shown in Table 3, all models improved after preprocessing. Random Forest consistently achieved the highest accuracy, followed by SVM. Decision Tree and KNN also showed improvement, but their performance lagged.

4.3. Confusion Matrix and Classification Report

Figure 5 illustrates the confusion matrix for the best-performing model, Random Forest, on the preprocessed dataset. The matrix reveals a strong balance between true positives and true negatives, indicating a reliable prediction performance. The classification report in Table 4 provides detailed precision, recall, and F1-score for each class.

5. Discussion

The results clearly show that preprocessing steps significantly improve the performance of all machine learning models. Random Forest emerged as the most effective model for predicting air quality, likely due to its ensemble nature which reduces overfitting and enhances generalization [26,27].

The improvements in accuracy, precision, and recall across models reinforce the importance of data preprocessing in environmental data analytics. These findings can be instrumental in guiding future implementations of predictive systems for air pollution monitoring and management.

6. Conclusions

Air pollution is a serious global problem that affects public health, ecosystems, and overall quality of life. This dataset is designed to analyze air quality by incorporating a wide range of environmental, industrial, and demographic factors. Consisting of 5,000 records and 10 features, the dataset includes key meteorological parameters (such as temperature and humidity) and pollution indicators (such as PM2.5, PM10, NO2, SO2, and CO). Additionally, factors such as proximity to industrial areas and population density help estimate potential pollution sources and levels of human exposure. The target variable, air quality, categorizes pollution levels into classes such as good and moderate, enabling qualitative analysis. This dataset supports descriptive analysis, predictive modeling, and policy development, offering researchers, environmentalists, and policymakers valuable insights to better understand air pollution dynamics and implement effective interventions. Its comprehensiveness makes it a valuable tool for promoting sustainable solutions and improving public health outcomes.