Identification of Infrasound Regimes at Mount Etna using Pattern Recognition Techniques

Overview

1. Introduction 2. Method 3. Results 4. Conclusion

Introduction

Mount Etna is an active volcano in a densely populated area.

It therefore needs constant monitoring.

Infrasound recordings play a vital role in this observation routine.

We apply a pattern recognition method to improve the analysis and interpretation of the signals in the infrasound data.

We select station ESLN at 1,700 m a.s.l. because of a high data availability and quality

Some Infrasound signals

Hover over the circles to learn more about the signals

Goal

Volcanic activity can manifest in many different infrasound scenarios.

For the human eye it is difficult to differentiate theses signals.

We apply a pattern recognition technique called Self-Organizing Maps to automatically classify infrasound patterns in the waveform.

Our approach simplifies the interpretation and analysis of infrasound data.

Method

We use an unsupervised pattern recognition technique called Self-Organizing Maps (SOMs) to classify infrasound waveform patterns.

We want to be sensitive to frequencies as low as 0.05 Hz. We need to ensure that for the lowest frequencies enough wave cycles fit into the window. We also do not want windows that are too long to avoid that the signal changes too much within the window. We therefore cut the waveform into windows of 250 s.

Step 2: Calculating the features

We calculate the spectrum to extract features that describe the waveform.

Step 2: Calculating the features

We simplify the spectrum by averaging in ten logarithmically spaced frequency bins.

These ten values built up our feature vector that represents the infrasound waveform within a 250 s window.

We call the parameter space of these features the Feature Space.

Step 3: Selection of a reference data set

We want to recognize typical infrasound patterns at Mount Etna.

For this, we need to compile a reference data set that contains feature vectors for every expected infrasound scenario.

We call these scenarios regimes and our reference data set should represent the parent population of possible regimes.

We have manually selected 8,600 feature vectors in the time period from December 2018 to March 2021 for our reference data set.

Step 4: Projection onto a 2D plane

We use Principal Component Analysis (PCA) to determine a 2D plane within the 10 dimensional feature space where we can project the feature vectors onto.

We call this projected feature space the Representation Space.

Step 5: Microclustering

1. Initialization

A lattice of Nodes is initialized in the feature space (i.e. they have the same dimensions as the feature vectors).

The nodes are arranged in a hexagonal configuration, e.g. each node has up to six neighboring nodes.

Step 5: Microclustering

2. Training

The nodes are iteratively shifted to best represent the feature vectors.

The node that is closest to any feature vector is called Best Matching Unit (BMU) for this particular vector.

In the training process the dispersion of all feature vectors that share the same BMU is minimzed and the dispersion of all nodes is maximized.

Step 6: Color coding

Based on their position in the representation space (i.e. after PCA projection) the nodes are assigned a color code.

Step 7: Predictions

Any feature vector (regardless whether it is included in the reference data set or not) can be assigned the color code of its BMU (i.e. the closest node in the feature space).

Similar feature vectors (representing a similar spectral content of the waveform samples) are represented by similar colors.

Step 8: Interpretation of the results

Up to this point the colors assigned to the waveform windows are completely arbitrary.
Using expert knowledge and information derived from other monitoring disciplines, we can assign labels to the colors.
Once these labels have been validated, newly acquired feature vectors can be interpreted based on these labels.
This is possible because new feature vectors are always evaluated on the nodes trained on the reference data set.
However, the prediction results are only valid as long as similar patterns are included in the reference data set.

Return to overview

Advance to next chapter

Results

To visualize the automated classification results we combine the color codes with the infrasound time series.

Displaying the raw waveform data is not practical when looking at time scales of more than a few hours

Result visualization

We therefore display the Root-Mean-Square (rms) value of the infrasound amplitude on a logarithmic y-axis.

Result visualization

We use the color information derived from our method to fill the area below the curve.

Infrasound Regimes

Use the buttons below to view the different regimes

Quiet phase

Low amplitude background noise. Absence of volcanic signals or wind noise. These patterns are represented by mint and light bluish colors.

Regime map

Quiet phase

Wind noise

Low volcanic activity

Medium volcanic activity

High volcanic activity

Intense volcanic activity

The color interpretation and knowledge of the most prominent regimes allows us to assign labels to the different region in the map.

We have validated the regime/color classification results with other volcanological observation as provided by the INGV monitoring routine. In the selected time frame from December 2018 to March 2021 we could not identify any false classifications.

Paroxysmal phase in February 2021

Use the buttons below to see the interpretation of the infrasound waveforms
for the first week of paroxysmal activity in February 2021

Return to overview

Advance to next chapter

Conclusion

We can utilize Self-Organizing maps to assign a color code to waveform samples of 250 s

These color codes are useful to get a fast and reliable analysis of the infrasound regime and can help non-experts at reading the infrasound data.

You have questions?

Contact

felix.eckel@ifg.uni-kiel.de

@FelixEckel

References

This presentation is powered by

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License