Machine Learning-based Spatial Mapping of Urban-Rural Air Quality: A Case Study of Mymensingh District in Bangladesh

Abstract

Air pollution is one of the most urgent environmental and social health issues in Bangladesh's rapidly urbanizing areas, but secondary cities still lack fine-scale evaluation and forecasting. This paper presents a combined PCA-machine learning-GIS model to quantify, model, and map air quality in urban and rural settings of the Mymensingh district. Department of Environment (DoE) monitoring records provided concentrations of SO2, NO2, CO, PM10, and PM2.5, supplemented with primary field measurements from an Aeroqual Series 500 sensor. To minimize multicollinearity, we used Principal Component Analysis (PCA) to create an Air Quality Index (PCA-AQI) composite as the target variable for machine learning modeling. We trained and evaluated Random Forest, Support Vector Machine, XGBoost, and Artificial Neural Network (ANN) models using standard performance measures. The ANN had the highest forecasting accuracy (R2 = 0.998), suggesting strong potential for capturing nonlinear interactions among pollutants. We used the primary datasets to train an ANN model optimized to predict PCA-AQI, and mapped the spatial distribution in ArcGIS using Inverse Distance Weighting (IDW) interpolation. Findings indicate that there are mostly poor air quality conditions in the urban areas, which are fueled by traffic and business operations, and the rural areas were generally good to moderate, with hotspots on highways and in markets. It is shown in the proposed framework that the fusion of PCA-based composite indices and machine learning can be effective in assessing air quality in data-scarce areas. The proposed framework shows that fusion of PCA-based composite indices with machine learning can effectively assess air quality management tool in the urban-rural setting and evidence-based environmental policy within developing countries.