GNGTS 2022 - Atti del 40° Convegno Nazionale

512 GNGTS 2022 Sessione 3.3 Fig. 1 - The repeating module in an LSTM contains four interacting layers, modified from https://colah.github.io/ posts/2015-08-Understanding-LSTMs/ . Once we have a trained NN we can start working on the prediction of each single neuron from each HL. In Fig. 2 we show all the deep attributes produced by the trained NN during the horizons extraction task. Fig. 2 - Example of the results of Deep Attribute extraction from a real GPR data profile acquired on a perennially frozen lake containing brines, in Antarctica (further details can be found in Forte et al. , 2016). From left to right we can see the input GPR amplitude data, the Hidden Layer predictions and the final horizons extraction. Results. Starting from the HL shown in Fig. 2, we selected some interesting Deep Attributes for evaluation purposes: in Fig. 3 we can see the real data (Fig 3.A), Fig. 3.B is the first deep attribute (Deep Attribute B, from now on), and so on. A discussion on results and possible interpretation follows. • Deep Attribute B: in this attribute we can focus mainly on noisy areas, we can see the brines identified with green color, while we marked with number 1 a coherent dark blue area. The overall structure of attribute B can be considered a summary of the main features evident in the input data that can also be exploited for automatic extraction. In fact both the main reflectors and homogeneous zones are apparent. • Deep Attribute C: here we find some areas with higher values - marked with number 2, just above the main horizon. We can link them to low noise zones: they are divided by the intermediate brine area (labeled 5 in (E) below), where we have very high signal attenuation due to the high electrical conductivity of brines. • Deep Attribute D: this attribute is more linked to a phase attribute, and we can spot the closing of a dipping horizon - marked with number 3 - and some low amplitude but laterally continuous layering - marked with number 4-.