| Popis: |
Information about the height and loading of sulfur dioxide (SO2) plumes from volcanic eruptions is crucial for aviation safety and for assessing the effect of sulfate aerosols on climate. While SO2 layer height has been successfully retrieved from backscattered Earthshine ultraviolet (UV) radiances measured by the Ozone Monitoring Instrument (OMI), previously demonstrated techniques are computationally intensive and not suitable for near real-time applications. In this study, we introduce a new OMI algorithm for fast retrievals of effective volcanic SO2 layer height. We apply the Full Physics Inverse Learning Machine (FP_ILM) algorithm to OMI radiances in the spectral range of 310–330 nm. This approach consists of a training phase that utilizes extensive radiative transfer calculations to generate a large dataset of synthetic radiance spectra for geophysical parameters representing the OMI measurement conditions. The principal components of the spectra from this dataset in addition to a few geophysical parameters are used to train a neural network to solve the inverse problem and predict the SO2 layer height. This is followed by applying the trained inverse model to real OMI measurements to retrieve the effective SO2 plume heights. The algorithm has been tested on several major eruptions during the OMI data record. The results for the 2008 Kasatochi, 2014 Kelud, 2015 Calbuco, and 2019 Raikoke eruption cases are presented here and compared with volcanic plume heights estimated with other satellite sensors. For the most part, OMI-retrieved effective SO2 heights agree well with the lidar measurements of aerosol layer height from Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) and thermal infrared retrievals of SO2 heights from the infrared atmospheric sounding interferometer (IASI). The errors in OMI retrieved SO2 heights are estimated to be 1–1.5 km for plumes with relatively large SO2 signals (> 40 DU). The algorithm is very fast and retrieves plume height in less than 10 min for an entire OMI orbit. This approach offers a promising prospect of using physics-based machine learning applications to other instruments. |
