| dc.description.abstract | Air pollution, particularly particulate matter, poses significant health and environmental risks. Conventional particulate matter measurement systems provide only single-point instantaneous readings, limiting their ability to generate spatial distribution maps. This study applies machine learning techniques to estimate particulate matter distribution using data from a limited number of sensor nodes. These nsors collected data on PM1.0, PM2.5, and PM10 concentrations, alongside weather parameters (wind speed, temperature, and humidity) and spatial information (longitude, latitude). Various machine learning models, including Artificial Neural Networks, Support Vector Regression, Long Short-Term Memory, and Random Forest, were evaluated for particulate matter distribution prediction. The analysis revealed that Artificial Neural Networks consistently outperformed other models across various feature configurations for predicting particulate matter concentrations. When considering only geometric features (Euclidean_D and Orientation), the Artificial Neural Networks model achieved training scores ranging from 0.9746 to 0.9790(with features: wind speed, temperature, geometric features) for PM1.0 prediction. Similar trends were observed for PM2.5 and PM10, with Artificial Neural Network straining scores between 0.9766 (geometric features only) and 0.9805 (humidity, geometric features) for PM2.5, and 0.9792 (geometric features only) to 0.9798(wind speed, temperature, geometric features) for PM10. Validation R 2 scores remained impressive for Artificial Neural Networks, ranging from 0.9711 to 0.9756 (PM1.0), 0.9692 to 0.9789 (PM2.5), and 0.9775 to 0.9793 (PM10) under the respective particulate matter scales feature combinations. For generalizability on unseen data, Artificial Neural Networks also achieved the highest average prediction accuracy, exceeding 75% consistently across all feature combinations for PM1.0, PM2.5, and PM10 concentrations within a 160-meter radius from a central sensor. Further, the findings revealed that incorporating the weather parameters significantly improved model performance, reducing both RMSE and MAE. Specifically, PM1.0 RMSE decreased from 1.8719 µg/m3 to 1.7201 µg/m3, and MAE from 0.8125 µg/m3 to 0.7952µg/m3. For PM2.5, RMSE reduced from 2.5260 µg/m3 to 2.2139 µg/m3, and MAE from 1.1488 µg/m3 to 1.0012 µg/m3. Lastly, PM10 RMSE decreased from 2.231µg/m3 to 2.143 µg/m3, and MAE from 1.1120 µg/m3 to 1.0037 µg/m3. This research demonstrates a machine learning-based approach to overcome limitations of single- point particulate matter sensors and predict particulate matter distribution across aregion. The research results highlight the effectiveness of Artificial Neural Networks, achieving high accuracy and emphasizing the value of including weather data in model training for improved particulate matter distribution prediction. The proposed approach offers a practical solution for real-world air quality monitoring by enabling data-driven decision-making in pollution management, optimizing sensor placement strategies, and facilitating more accurate identification of pollution hotspots for targeted interventions. The study findings suggest a new approach to designing particulate matter sensors, overcoming limitations of single-point measurement, and emphasize the importance of including weather parameters in machine learning model training for spatial distribution prediction. | en_US |
