A027-05
Machine Learning for Efficient Prediction of High Spatial Resolution NO2 a Priori Profiles

Monday, 7 December 2020: 20:56
Virtual
Jennifer Grant, University of California, Berkeley, Department of Mathematics, Berkeley, CA, United States, Hannah Schaefer Kenagy, University of California, Berkeley, Department of Chemistry, Berkeley, CA, United States, Qindan Zhu, University of California, Berkeley, Department of Earth and Planetary Sciences, Berkeley, CA, United States and Ronald C Cohen, University of California, Berkeley, Dept. of Earth and Planetary Science, Berkeley, CA, United States; University of California, Berkeley, Dept. of Chemistry, Berkeley, CA, United States
Recorded Presentation
Abstract:
Nitrogen dioxide (NO2) is a toxic air pollutant involved in the formation of surface-level ozone and particulate matter. Tropospheric NO2 can be measured from space using satellite-based UV/VIS spectrometers (e.g., OMI, TROPOMI, TEMPO). However, because the sensitivity of space-based spectroscopic instruments varies with altitude, accurate high-resolution NO2 retrievals require high-resolution a priori vertical profiles. A priori profiles simulated using chemical transport models (CTM, e.g., WRF-Chem), which includes both a full chemistry mechanism and meteorological fields, are computationally expensive and time consuming, preventing routine development of profiles at the space and time resolution of the current generation of satellite instruments. We propose an efficient alternative that would use as inputs a meteorological forecast or analysis and a high resolution emission inventory to produce profiles at the native resolution of the measurements. Our method uses a random forest learning model trained on a long record of 12km profiles calculated with WRF- CHEM to predict a priori NO2 vertical profiles. We describe the accuracy of the predicted profiles and associated air mass factors and compare the computational expense of the two approaches.
<< Previous Abstract | Next Abstract >>