Estimating surface-level fine particulate matter from satellite remote sensing data can expand the spatial coverage of ground-based monitors. This approach is particularly effective in assessing rapidly changing air pollution events such as wildland fires that often start far away from centralized ground monitors. We developed Deep Neural Network algorithm to bias correct hourly PM2.5 levels informed by GOES-R satellites, NOAA meteorology forecasts, and real-time PM2.5 observations from the Environmental Protection Agency (EPA) via AirNow. The surface-satellite-model collocated datasets for the period of 2020-2021 was used to assess the biases in GOES-GWR PM2.5 against AirNow measurements at hourly and daily scales. Then a deep neural network (DNN) based bias correction algorithm is used to improve the accuracies of GOES-GWR PM2.5. The DNN uses GOES-GWR PM2.5, GOES-R aerosol parameters, and HRRR meteorological fields as input and AirNow PM2.5 is used as target variable. The application of DNN reduced the PM2.5 biases as compared to GOES-GWR estimates. RMSE was also reduced to 6.55 µg/m3 from 8.72 µg/m3 in GOES-GWR estimates. The DNN model was also evaluated on two sets of independent datasets for its robustness. In the first independent dataset for the first half of 2020, ~89% of stations show an increase in correlation (r) and, ~76% and ~62% of stations show a reduction in bias. The IOA and r for the independent data were 0.77 and 0.61 (GWR: 0.68 and 0.53) and RMSE was 4.48 µg/m3 (GWR=6.13 µg/m3) for the same period.

Health and environmental hazards related to high pollutant concentrations have become a serious issue from the perspectives of public policy and human health. The objective of this research is to improve the estimation of grid-wise PM2.5, a criteria pollutant, by reducing systematic bias in estimating PM2.5 empirically from speciation provided by MERRA-2 using a ML approach. We present a unique application of machine learning (ML) for estimating hourly PM2.5 concentrations at grid points of Modern-Era Retrospective analysis for Research and Applications version 2 (MERRA-2). The model was trained using various meteorological parameters and aerosol species simulated by MERRA-2 and ground measurements from Environmental Protection Agency (EPA) air quality system (AQS) stations. monitors. The ML approach significantly improved performance and reduced mean bias in the 0-10 µg m-3 range. We also used the Random Forest ML model for each EPA region using one year of collocated datasets. The resulting ML models for each EPA region were validated and the aggregate data set has a Pearson correlation of 0.88 (RMSE = 4.8 µg m-3) and 0.82 (RMSE = 5.8 µg m-3) for training and testing, respectively. The correlation (and RMSE) increased to 0.89 (4.0), 0.95 (1.6), 0.94 (1.1) for daily, monthly, and yearly average comparisons. The results from initial implementation of the ML model for global region are encouraging but require more research and development to overcome challenges associated with data gaps in many parts of the world.