This study investigates the effectiveness of machine learning models in estimating ground-level PM2.5 concentrations using satellite-derived aerosol optical depth (AOD) in Kampala, a city with limited ground monitoring. Multivariate Linear Regression, Random Forest, and Multi-Layer Perceptron are compared using data from reference-grade and low-cost monitors. The study uses MODIS MAIAC AOD and PM2.5 measurements from January 2022 to December 2023 from two collocation sites. The results indicate that Random Forest consistently outperforms other methods, demonstrating strong performance despite limited data; achieving the lowest RMSE: (7.72 \(\upmu \hbox{g}/\hbox{m}^3\) ) and MAE: (5.5 \(\upmu \hbox{g}/\hbox{m}^3\) ), R2 of (0.72) and correlation of 0.85. Models trained on combined datasets from multiple devices and sites generally yielded better results than those trained on individual devices’ data, suggesting the benefits of data integration for robust PM2.5 estimation. Uniquely, the study demonstrates the viability of using low-cost sensor data as ground-truth for PM2.5 estimation, with a model trained on low-cost sensor data achieving a RMSE: 10.9 \(\upmu \hbox{g}/\hbox{m}^3\) , MAE: 8.6 \(\upmu \hbox{g}/\hbox{m}^3\) , R2: 0.59 and Correlation: 0.78, when validated against collocated reference-grade monitors. The results show promise for using low-cost sensor data as ground truth in PM2.5 estimation from satellite AOD, though further refinements are needed. This study contributes to the growing body of research on satellite-based air quality monitoring, particularly in cities in low- and middle-income countries where ground monitoring is limited. The results highlight the potential of using satellite data, machine learning techniques and low-cost sensors to improve PM2.5 monitoring in under-resourced regions.
