High-resolution air quality forecasts for LMIC using a combination of machine learning and earth model simulations

Abstract

Background and aim: With the recent advancement of Earth observation, computer simulations, and low-cost monitoring technologies, the capacity to obtain accurate geospatial data has increased tremendously, particularly for locations without expansive in-situ monitoring networks. Here we present an optimized machine learning model to estimate near real-time air pollutant concentrations in selected locations in Africa and Latin America using a combination of NASA Goddard Earth Observing System Composition Forecasts (GEOS-CF) and low-cost sensor data. Methods: We use a machine learning approach to estimate near real-time air pollutants concentration at selected locations across Africa and Latin America. Several meteorological and chemical parameters are retrieved from NASA’s GEOS-CF to train a bias corrector model and predict corrected concentration estimates for these locations which are then validated against local monitoring data. We also conduct an explainability approach via SHAP Analysis to quantify the model contributing factors across these locations and track the model performance in extreme conditions. Results: The optimized machine learning model shows good agreement with ground air quality data, with R2 values from 0.61- 0.65 for locations in Mexico City (Mexico), Bogotá (Columbia) and Kigali (Rwanda). Via SHAP analysis, we demonstrate the need to conduct a measure of variance to determine model performance for different conditions and intervals, knowing that the model performance can be affected by various training conditions. Conclusion: Combining observations and optimized model simulations using machine learning techniques can significantly improve air quality forecasts in low- and middle-income countries, where the rapid pace of industrialization and communities are highly susceptible to air pollution health effects, and rarely have local air quality data and health risks alerting systems in place. These results are being used to assist air quality managers and environmental agencies in these locations to improve risk communication and reduce health burdens associated with outdoor air pollution.