Development of Integrated Machine Learning Model for Estimation of Spatial Distribution of Particulate Matter Pollutant in Air

Abstract Accurate estimation of particulate matter (PM) concentration in air is crucial for understanding and mitigating air pollution. Conventional PM measurement systems often provide only single-point, instantaneous readings, limiting their ability to create comprehensive spatial pollution maps. This study aims to address this limitation by leveraging Machine Learning (ML) techniques to predict PM concentration using data from a limited number of sensor nodes. ML training spatiotemporal dataset was collected over a month, encompassing PM1.0, PM2.5, and PM10 concentrations, weather parameters, and spatial information. Four machine learning models namely; LSTM, ANN, SVR, and Random Forest, were evaluated for their ability to predict the spatial distribution of PM concentration at three different sites. The analysis revealed that ANNs consistently outperformed other models across different feature combinations. Within a 160‑meter radius of a central sensor node at (First Site), the ANN model achieved an average prediction accuracy above 86.0% for PM concentrations, with RMSE and MAE values of 1.72 µg/m³ and 0.80 µg/m³, respectively. The inclusion of weather parameters and feature engineering improved model performance by 12% to 18%, compared to models using only geometric features. When applied to (Second Site), the ANN model maintained an accuracy of 85%, demonstrating strong intra‑environment generalizability. However, performance declined at the (Third Site), located 3 kilometers away in a peri‑urban market setting, where microclimatic and topographic variability resulted in a lower prediction accuracy of 59.2%. This study demonstrates the effectiveness of a machine learning-based approach to overcome the limitations of single-point PM sensors and predict PM concentration across a region. The results highlight the superior performance of ANNs and emphasize the importance of incorporating weather parameters and feature engineering in model training. These findings are promising and point to a new approach of developing a practical model for spatial estimation of PM distribution from measurements obtained from sensor networks.