Machine Learning of A Priori Information in Optimal Estimation of Atmospheric Composition

The optimal estimation technique is commonly used in remote sensing to retrieve vertical profiles of atmospheric constitutes. This technique requires a priori information to stabilize the retrieval process. In cases when the measurements provide only limited information, accuracy of the a priori profiles can affect the retrieved quantity. As an example, measurements from ground-based Brewer spectrometers and multi-axis DOAS measurements have limited information about the vertical shape of ozone profiles, and accurate a priori information is needed. Often seasonal ozone climatology is used for this purpose. However, ozone vertical distribution can significantly change within shorter time scales, and seasonal climatology is unable to reflect those changes. In this study, we explore a new approach for predicting ozone profiles using machine learning. We train a neural network model with a set of more than 10,000 assimilated ozone profiles and corresponding meteorological parameters. The Modern-Era Retrospective analysis for Research and Applications version 2 (MERRA-2) assimilates satellite limb observations from NASA's Aura Microwave Limb Sounder (MLS) and Ozone Mapping Instrument (OMI). Using MERRA-2 data, we train the neural network to predict vertical ozone profiles over The NASA Goddard Space Flight Center from inputs of meteorological readings and temporal indicators. The neural network uses a convolutional architecture to capture correlations among ozone concentrations in nearby elevations. In preliminary experiments, we train the neural network using MERRA-2 data from January 2005 to December 2008, with eight snapshots spread throughout each day, and we evaluate its ability to predict ozone profiles for all of 2011 and 2012. The resulting model is able to predict more refined ozone profiles than seasonal climatology, reflecting the effects of time-of-day and the increased ozone variability in winter. These results are promising indicators of potential benefits of embedding modern machine learning methods into atmospheric remote sensing.