Estimation And Prediction Of PM2.5 And PM10 Concentration In Kathmandu Valley Using Remote Sensing And Machine Learning
TL;DR
The study demonstrates the potential of integrating satellite data with machine learning for air quality monitoring in data-scarce regions, supporting evidence-based environmental policy in Nepal and similar contexts.
Abstract
In recent years, public health and the environment have been seriously threatened by air pollution, especially in low and middle income nations like Nepal, where the challenge is compounded by sparse ground-level monitoring stations that provide limited spatial coverage. To address this data gap, this study uses machine learning algorithms and satellite data to meet the pressing demand for precise PM2.5 and PM10 prediction. Ground station data from the monitoring stations of the Department of Environment covering the period from January 2022 to December 2023 were combined with satellite-derived parameters including Aerosol Optical Depth (AOD), Land Surface Temperature (LST), Normalized Difference Vegetation Index (NDVI) and soil moisture to train the models. The Gradient Boosting algorithm performed consistently well, with average R² scores of 0.82 and 0.84 for PM2.5 and PM10 respectively, proving efficient in capturing non-linear relationships in air quality data. Likewise, Random Forest also showed reliable accuracy with R² values of 0.80 and 0.82 for PM2.5 and PM10 respectively. Spatial maps created from model predictions highlight pollution hotspots, providing actionable insights for targeted interventions. The study demonstrates the potential of integrating satellite data with machine learning for air quality monitoring in data-scarce regions, supporting evidence-based environmental policy in Nepal and similar contexts. Future research to incorporate high-resolution datasets and additional metrological variables to enhance model reliability and scalability.