Automatic Volcanic Ash Detection within MODIS Observations using a Back-Propagation Neural Network

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Sunday, 4 January 2015
Tami M. Gray, Vanderbilt University, Nashville, TN; and R. Bennartz and J. Rausch

Due to the climate effects of volcanic eruptions and the threat they pose to aviation, it is important to accurately determine the location of ash in the atmosphere. This study aims to explore the accuracy and reliability of training a neural network to identify cases of ash within observations from the Moderate Resolution Imaging Spectrometer (MODIS). Satellite images were obtained by MODIS for the following eruptions: Kasatochi, Aleutian Islands, August 7, 2008; Okmok, Aleutian Islands, July 12, 2008; GrÝmsv÷tn, northeastern Iceland, May 21, 2011; Chaitén, southern Chile, May 7, 2008; Puyehue-Cordón Caulle, central Chile, June 6, 2011; and Soufriere Hills, Montserrat, May 20, 2006. The following band combinations were created to differentiate between ambient clouds, surface features, ash, SO2, and ice: 12-11-μm brightness temperature difference to distinguish between ash and ice, 11-8.6-μm brightness temperature difference to identify SO2, and the 11-μm brightness temperature to identify Earth's surface features. As validation, the Hybrid Single Particle Lagrangian Integrated Trajectory Model (HYSPLIT) was used to obtain volcanic ash concentrations for the same archived eruptions. A back-propagation neural network was then trained with two random selections of pixels, one containing ash and the other excluding ash. The neural network used the following inputs and output, respectively: 12-11-μm brightness temperature difference, 11-8.6-μm brightness temperature difference, 11-μm brightness temperature, and the ash concentration integrated over the entire column (g/m2). When tested against MODIS granules containing ash, initial results show the neural network correctly placing ash, but overestimating in the presence of high clouds.