Neural-network approach to multispectral and hyperspectral data analysis for volcanic monitoring

The Artificial Neural Network (ANN) approach has demonstrated its effectiveness in Geophysics, considered as an universal approximator, being able to model physical nonlinear phenomena and to solve complex inversion problems in a very short time. Indeed once the training phase is completed, they can be applied in a very fast manner to new data, so that the computational burden required for the data processing is drastically reduced. This characteristic assumes an important role when considering the possible application to high revisit time sensors like Meteosat Second Generation (MSG) SEVIRI. Indeed the algorithms based on radiative transfer model simulations are generally time consuming making its application difficult in near real time.* *This dissertation continues a line of research started at the Earth Observation Laboratory of the Tor Vergata University in Rome for inverse modelling of volcanic ash mass retrieval from MODIS using ANNs reporting on the latest advances obtained in the development of neural network algorithms for volcanic ash parameters, such as ash particle size and Aerosol Optical Depth (AOD) and SO2 retrieval from MODIS data and extending the research activity including the first attempt of applying ANNs to hyperspectral remote sensed data, emulating an inverse model for simultaneous estimates of SO2 total columnar content and the cloud height. Attention has been paid also on an ANN pruning analysis in order to find significant spectral wavelength in multispectral data for parameter inversion.