A Neural Network Approach for Volcanic Monitoring of Sulphur Dioxide Using Hyperspectral Remote Sensed Data
Piscini, Alessandro¹; Carboni, Elisa²; Grainger, Roy^2
1INGV, ITALY; ²AOPP, UNITED KINGDOM

This paper describes an application of artificial Neural Networks for the simultaneous estimation of the columnar content and height of the SO2 plume from volcanic eruptions using hyperspectral remotely sensing data. The satellite remote sensing is a very effective and safe way to monitor volcanic eruption in order to safeguard the environment and the people affected by these natural hazards. Neural Networks are an effective and consolidated reversal technique for the estimation of geophysical parameters from satellite images. In addition, once the latter have been trained are proving very fast in the application stage. In our study two neural networks have been implemented to estimate the columnar content of SO2 and of its height. Neural Networks have been trained using all IASI channels between 1000 - 1200 and 1300 - 1410 cm -1, and the corresponding values of SO2 amount and height of the plume obtained from the same IASI channels using the Oxford SO2 retrieval scheme. As a case study we have chosen the Eyjafjallajokull volcano (Iceland), in particular the eruption took place during the months of April and May 2010, which had an enormous impact on the world economy. The neural network has been trained with a time series consisted of 58 hyperspectral eruption's images from 14 April to 14 May 2010 and validated on three independent data sets of images belonging to the same eruption, one in April and the other two in May. Furthermore they were applied to an independent data-set, a IASI image of Grìmsvotn volcanic eruption, occurred on May 2011. The results of the validation both for Eyjafjallajokull and Grìmsvotn independent data-sets have provided values of Root Mean Square Error (RMSE) between neural network outputs and targets always less than 20 DU for SO2 total column and 200 mb for plume height, respectively, so demonstrating the excellent performance in retrieval achieved by the neural network technique and its usefulness in near real time monitoring activity.