Spatiotemporal PM2.5 Estimation in Kathmandu Using Deep Learning with OpenMeteo and NASA MERRA-2 Data: Performance Benchmarking Against Machine Learning Model

Abstract

In rapidly urbanizing regions like the Kathmandu Valley, air pollution, specifically fine particulate matter (PM₂.₅), represents an increasing environmental and public health challenge. Traditional air quality monitoring systems are inherently limited in spatiotemporal scope, limiting the ability to conduct long-term air-quality assessments. This study aims to develop a framework for high-resolution historical PM₂.₅ concentration reconstruction from previously limited spatiotemporal PM₂.₅ datasets. To achieve this goal, we will incorporate historical Open-Meteo weather data and NASA MERRA-2 satellite reanalysis data as dependencies. The predictive models will be evaluated, namely the Deep Neural Network (DNN) and Extreme Gradient Boosting (XGBoost) two models using both hourly and daily datasets, based on qualitative and quantitative study metrics. Using the hourly Open-Meteo dataset, the DNN model achieved the highest accuracy (R² = 0.8725, RMSE = 18.23 µg/m³) and the XGBoost model performed best (R² = 0.7827, RMSE = 12.81 µg/m³) using the daily dataset. More generally, through evaluation, it appears that data quality and resolution can outweigh the effect of the algorithm in predicting PM₂.₅. This framework demonstrates considerable promise for capturing nonlinear and temporal dependencies within air pollution dynamics to conduct high-fidelity PM₂.₅ reconstruction. Findings support a data-informed basis for environmental and urban planning, forecasting of environmental issues, and public health intervention in the Kathmandu Valley.