GOES-R PM2.5 Evaluation and Bias Correction - A Deep Learning Approach

National Oceanic and Atmospheric Administration (NOAA) provides near-real-time estimates of the PM2.5 using GOES east and west Advanced Baseline Imager (ABI). The PM2.5 was estimated from AOD using a geographically weighted regression (GWR) model. The model is updated dynamically for each hour using AirNow ground measurements to provide the best estimates. In this study, we used two years of hourly PM2.5 data and evaluated it against AirNow measured PM2.5. The biases are characterized under different conditions, and corrections are developed using deep learning algorithms. To further improve the PM2.5 estimates and reduce biases, we developed a deep neural network (DNN) model that was trained on hourly data for the continental US (CONUS). Several meteorological parameters from the NOAA High-Resolution Rapid Rate (HRRR) model, ABIs data of PM2.5, AOD, and smoke-dust mask are used as input to the DNN model. The applications of DNN improve the PM2.5 biases as compared to GWR estimated PM2.5. The DNN model increased the correlation (r) by 0.5-330% for 80% stations and reduce RMSE by 0-82% for 74% stations over GWR estimates. The aggregated r for the DNN was 0.86 (GWR=0.80) and RMSE was 6.99 µgm-3 (GWR=8.80 µgm-3) for the test dataset. The DNN model was also evaluated on an independent dataset for its robustness. The r for independent dataset was 0.67 (GWR=0.46) and RMSE was 4.026 µgm-3 (GWR=7.21 µgm-3). The bias-corrected value-added PM2.5 parameter is developed for EPA's AirNow portal supported by NASA Health and Air Quality Applied Science Tiger Team project.