Improving volcanic event recognition through new seismic signals representation

Many studies highlight the importance of automating some steps of volcano monitoring. In particular, the classification of the volcano seismic events is a basic but fundamental task for observatories. The literature shows that artificial neural networks (ANN) are currently the most widely used techniques to develop classifiers for seismic signals from volcanoes. When training neural models, in addition to the quality of the dataset, the representation of the seismic events that feed the ANN is crucial for obtaining good performance. In general the representation is carried out in the time, frequency and both (spectrograms) domains. Spectrograms often need to be normalised for a better comparison between classes, resulting in a loss of the information that time variable typically carries. In previous works we have incorporated the events’ duration at the input or in intermediate layers of the ANN, without significant improvements in classification. In this paper we present a method that incorporates the time variable by adding the signal trace as an image to the spectrogram. Additionally, we also tested adding the spectrum module. Class activation maps (CAM) were applied over the different inputs to observe the sensitive areas that contribute to the neural model’s decisions and determine if the newly added representations are relevant for discriminating between classes. The results show that, for some classes, this additional information is indeed relevant for discriminating between seismic events and CAM brings explainability to the models.