A Machine learning approach for predicting particulate matter concentration in Nairobi County

Abstract

Nairobi experiences elevated particulate matter(PM2.5) concentrations due to motor vehicle emissions, biomass fuel use, industrial processes, and open waste burning, leading to significant public health risks. This study developed a predictive tool for PM2.5 concentrations in Nairobi using machine learning, integrating real-time weather data and historical pollution trends. Using the CRISP-DM framework, the research incorporated data from low-cost air quality sensors, supplemented by meteorological data from Open- WeatherMap. After rigorous preprocessing such as missing value imputation, outlier treatment, and resampling to an hourly interval—exploratory data analysis revealed diurnal and seasonal patterns, with peak pollution levels observed during dry seasons and weekday rush hours. Correlation analysis showed weak negative relationships between PM2.5 and temperature (r = 0.07), wind speed (r = 0.10), and dew point (r = 0.05), indicating pollutant dispersion under warmer, windier conditions. Minor positive correlations with pressure (r = 0.03) and humidity (r = 0.01) suggested that stable atmospheric conditions slightly increase PM2.5 levels. Other factors like visibility, rainfall, and wind direction had minimal impact. Spatial analysis identified significant hotspots along high traffic corridors, notably the Nairobi Expressway/Southern Bypass (z = 2.80, p = 0.016) with average PM2.5 of 26.01 μg/m3, and coldspots in higher elevation areas like Kilimani (z = 2.61, p = 0.001), where PM2.5 averaged below 12 μg/m3. This reflects the combined influence of local emissions and topography on pollution distribution. A LightGBM quantile regression model achieved strong predictive performance (RMSE: 2.821, R²: 0.915). A Streamlit web app was developed to provide interactive forecasts and AQI categorizations, offering a valuable tool for air quality management and public health planning in Nairobi.