Leveraging low-cost sensors and machine learning for air quality insights in an urban location of Zimbabwe: A case study

Ambient air pollution is a growing concern in Zimbabwe, particularly in urban and industrialised zones. The country faces increasing air quality challenges due to rapid urbanization, industrial growth, an increase in vehicular population and unsustainable energy practices. Despite efforts by the environmental authorities and the introduction of air pollution control regulations under the Environmental Management Act, challenges persist due to limited enforcement, outdated industrial technologies, insufficient public awareness of the risks associated with air pollution and a lack of monitoring data. This study explores the potential of using low-cost monitors complemented by machine learning. The efficacy of deploying low-cost air quality sensors coupled with machine learning algorithms to provide valuable air quality insights and predictive capabilities was investigated using data collected from a short-term monitoring campaign at the University of Zimbabwe campus. Key findings from our short-term monitoring campaign indicate that particulate matter (PM2.5 and PM10) concentrations frequently exceeded WHO guidelines, highlighting local air quality concerns. Six machine learning models were initially considered, with four deep learning models showing the most promise and being further evaluated. The Bidirectional Hybrid LSTM-CNN had the best performance in predicting air quality index, with R 2 = 0.57, RMSE=19.7 and MAE=15.4 when the models were evaluated on the test data set for a 6-hour forecasting horizon. This approach provides valuable insights for environmental management and public health planning in resource-constrained settings typical of developing countries and also provides practical insights into the application of emerging technologies for air quality management. While initial results with deep learning algorithms were promising, we discussed challenges encountered such as the impact of limited dataset size on model generalization.As a case study, this work demonstrates a viable framework for future long-term ambient air quality monitoring and prediction.