Simultaneous retrieval of volcanic sulphur dioxide and plume height from hyperspectral data using artificial neural networks

Artificial neural networks (ANNs) are a valuable and well-established inversion technique for the estimation of geophysical parameters from satellite images; once trained, they help generate very fast results. Furthermore, satellite remote sensing is a very effective and safe way to monitor volcanic eruptions in order to safeguard the environment and the people affected by those natural hazards. This paper describes an application of ANNs as an inverse model for the simultaneous estimation of columnar content and height of sulphur dioxide (SO2) plumes from volcanic eruptions using hyperspectral data from remote sensing. In this study two ANNs were implemented in order to emulate a retrieval model and to estimate the SO2 columnar content and plume height. ANNs were trained using all infrared atmospheric sounding interferometer (IASI) channels between 1000-1200 and 1300-1410 cm(-1) as inputs, and the corresponding values of SO2 content and height of plume, obtained from the same IASI channels using the SO2 retrieval scheme by Carboni et al., as target outputs. The retrieval is demonstrated for the eruption of the Eyjafjallajokull volcano (Iceland) for the months of 2010 April and May and for the Grimsvotn eruption during 2011 May. Both neural networks were trained with a time series consisting of 58 hyperspectral eruption images collected between 2010 April 14 and May 14 and 16 images from 2011 May 22 to 26, and were validated on three independent data sets of images of the Eyjafjallajokull eruption, one in April and the other two in May, and on three independent data sets of the Grimsvotn volcanic eruption that occurred in 2011 May. The root mean square error (RMSE) values between neural network outputs and targets were lower than 20 Dobson units (DU) for SO2 total column and 200 millibar (mb) for plume height. The RMSE was lower than the standard deviation of targets for the Grimsvotn eruption. The neural network had a lower retrieval accuracy when the target value was outside the values used during the training phase.