Home | deutsch  | Department | Legals | Data Protection | KIT

Presentation of Master's Theses: Arrival time detection with deep learning (Philipp Seither) & Parameter Sensitivity of Ash Fall Hazard Modelling (Regina Beckmann)

Presentation of Master's Theses: Arrival time detection with deep learning (Philipp Seither) & Parameter Sensitivity of Ash Fall Hazard Modelling (Regina Beckmann)



Philipp Seither, Regina Beckmann


Arrival time detection with deep learning (Philipp Seither)

Manual determination of P and S phase onset times can be a time-consuming process which takes up precious time better spent on other things. Automatic picking algorithms based on the ratio of short-term to long-term averaging exist but struggle, for example, with S wave arrivals in recordings from local events where the S coda is obscured.
In contrast to such auto pickers, neural networks are not limited to comparing human-engineered features like signal energy and others, rather they engineer their own features for detecting P and S phases from recordings when trained.
The choice of neural network architecture and availability of training data determines the quality of the final auto picker. Previous studies have successfully employed Convolutional Neural Networks trained on millions of event recordings to predict P and S phase onsets.
As a proof of concept, this study shows that Neural Networks can also be trained on parts of a single earthquake catalogue with only thousands of picks, and subsequently pick accurately on the remaining parts of the catalogue. Preprocessing of data and different varieties of CNNs are looked at and the most impactful choices in CNN training are identified. The transferability of CNNs trained on a regional catalogue to recordings from alien regions is tested and the trained networks are compared to a state-of-the-art CNN auto picker. Use cases for and the reliability of the trained network will be discussed.


Parameter Sensitivity of Ash Fall Hazard Modelling (Regina Beckmann)

Volcanoes can be found all around the world and frequently trigger fatal eruptions, which often represent a combination of various hazards. In contrast to magmatic and pyroclastic flows, explosions and lahars which generally are limited to the direct vicinity of the volcano, volcanic ash fall can concern much larger areas and cause serious human and economical damage. Thus, for any volcano hazard assessment, the modelling of the potential ash fall distribution is an essential component. Numerical models for this purpose demand various input parameters as properties of volcanic ash particles, mass flow rate, spatial and temporal behaviour of the eruption column, topography and weather conditions. In this work the fully computational model FALL3D is used. Unfortunately, collecting all required data is a tedious and time-consuming process. Therefore it is of major importance to identify the individual sensitivity of FALL3D to the most important input parameters, which is done in regards to spatial ash distribution. The study is done for several characteristic hazard scenarios at the volcanoes Santiaguito and Fuego in Guatemala which showed fatal activity since 1900 and therefore represent a serious threat to millions of people living within 100 km from both volcanoes. The results allow to distinguish between crucial parameters and ones which have only minor influence on spatial ash distribution so that future work can be concentrated on the former ones. Based on the findings of the sensitivity study ash fall hazard is assessed and evaluated for the investigated scenarios at both volcanoes of interest, revealing a distinct dependency on the month the eruption takes place. ​