Improving SO2 Layer Height Retrieval for the OMI and OMPS Spectrometers with an Inverse Machine Learning Based Approach
Abstract
Accurate retrievals of SO2layer height in volcanic plumes are crucial for aviation safety and for determining the impact of volcanic eruptions on climate. Previous retrievals of SO2height with Ozone Monitoring Instrument (OMI) and similar hyperspectral UV instruments are mostly based on direct spectral fitting inversion techniques. These techniques have demonstrated the feasibility to derive effective plume height for large volcanic SO2amounts, but are too time consuming for near real time and direct readout processing. In this work, we aim to implement a previously developed full physics - inverse machine learning method for retrieving SO2layer height with OMI and OMPS instruments. The method consists of an offline training phase and an operational phase. In the training phase, a large dataset of synthetic top-of-the-atmosphere reflectance spectra is constructed accounting for multiple elastic (Rayleigh) and non-elastic Rotational Raman Scattering (RRS) scattering as well as ozone and sulfur-dioxide absorption using the LIDORT-RRS radiative transfer model simulations. The spectra are simulated for the 310-330 nm spectral window, using a high resolution synthetic solar flux dataset. In the next step, the simulated radiance spectra are processed and fed through a feed-forward neural network in order to create an inverse operator that maps the given TOA spectra and physical parameters to the SO2 layer height output. We apply the trained inverse model to OMI and OMPS measurements in the operational phase for highly efficient, near real time retrievals of volcanic SO2layer height with low demand for computational resources. The results obtained from the trained model are tested for the Kasatochi eruption in August 2008 and compared with previous studies on the Kasatochi plume height. Comparisons to the TROPOMI plume height retrievals are also foreseen. This data driven and machine learning based approach can be easily implemented with different hyperspectral UV sensors. In the future we will also utilize this technique to study both past eruption cases during the OMI/OMPS era, in addition to near-real-time operational applications.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.A51I2693F
- Keywords:
-
- 0305 Aerosols and particles;
- ATMOSPHERIC COMPOSITION AND STRUCTURE;
- 0370 Volcanic effects;
- ATMOSPHERIC COMPOSITION AND STRUCTURE;
- 4315 Monitoring;
- forecasting;
- prediction;
- NATURAL HAZARDS;
- 8409 Atmospheric effects;
- VOLCANOLOGY