Bridging the gap between data generation and data comprehension in seismology: Towards new insights and improved understandings in seismology via Intelligent Data Analysis
Final Report Abstract
Intelligent Data Analysis (IDA) provides a wealth of non-trivial ("intelligent") algorithms, techniques and methodologies aimed at analyzing empirical data. IDA is successfully being applied in various domains, resulting in new insights and improved understanding of many processes, in particular when uncertainty is inherently associated with the problem at hand. We applied IDA to various earth scientific problems, such as in the field of natural hazards, e.g., earthquakes, tsunamis, flood events, landslides, volcano eruptions. These domains all share (despite of their differences in causes and effects) many of the same model- and decision theoretic questions which can be dealt with using IDA methods. Usually the natural processes are complex and the number of potential influencing factors is large. The single and joint effects of the driving forces are not necessarily completely understood, potentially introducing a large degree of uncertainty which impacts the overall analysis and results. Additionally the observation on the basis of which inference is made is often sparse, inaccurate and/or incomplete, adding yet another layer of uncertainty on top. From an IDA point of view, many methods were applied where a principled handling of uncertainty is taken explicit into account: (Dynamic) Bayesian networks (D)BNs, learning (D)BNs from observations, Bayesian statistics, modeling and eliciting prior knowledge, mixture modeling, transfer learning, Bayesian non-parametrics, hierarchical modeling, visualization and dimensionality reduction. In particular worth mentioning is our use of socalled Bayesian Networks (BNs) in natural hazards, allowing for an intuitive, consistent, and rigorous way of quantifying uncertainties. The (in)dependencies between the involved variables relevant to a particular hazard domain are translated into a graph structure enabling for improved understandings and direct insights into the relationships and workings of a natural hazard system. Moreover, as BNs capture the joint distribution of all variables of a given domain, they may be used for expressing any conditional probability distribution of interest, thereby helping to answer quantitative questions on specific scenarios or process response chains. More specifically, we have looked at BNs in the context of tsunami early warning, probabilistic seismic hazard analysis (the ground-motion modeling part), landslide prediction and for flood damage assessments. For all domains, we took a data-driven stance, and were able to develop predictive models, where not only uncertainty of a prediction is taken into account, but we may also perform instant prediction with incomplete evidence/observations (useful for tsunami early warning) and at the same obtain insight into the workings of natural system system itself. Moreover, in comparative studies we showed that BNs perform better or on par with existing methods (usually based on simple predictive models, e.g.. regression models). The BNs were developed from data, however, often data in unavailable or scarce. In such situations prior knowledge needs to be incorporated in typical data-driven IDA methods. To this end we have investigated and developed a platform for eliciting expert knowledge, taking into account the typical biases and other problems with regard to consistently obtaining trustworthy domain knowledge. For this we have focused on probabilistic seismic hazard analysis, where expert knowledge of various ground motion models is required. Additionally, we have developed transfer-learning approaches for "borrowing" information form related ground motion models, hence augmenting data sets for regions where hardly any observations have been made, e.g., for geographically adjacent regions, where for one region we have plenty of data to derive a ground motion model, but for the other region we don't; in that case we implicitly borrow observations (via an intelligent mechanism) to derive a model even for the region with few observations.
Publications
-
2011. Learning Aggregations of Ground-Motion Models Using Data. International Conference on Applications of Statistics and Probability in Civil Engineering, ICASP'11, Taylor & Francis Group/CRC Press, Proceedings and Monographs in Engineering, Water and Earth Sciences, ISBN 978-0-415-66986-3. pp. 721-728
Riggelsen, C., Gianniotis, N., Scherbaum, F.
-
2012. A Machine Learning Approach for Improving the Detection Capabilities at 3C Seismic Stations. Recent Advances in Nuclear Explosion Minotoring Vol. 2 - Pure and Applied Geophysics, PAGEOPH
Riggelsen, Carsten & Ohrnberger, Matthias
-
2012. Learning Task Relatedness via Dirichlet Process Priors for Linear Regression Models. European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, Bruges, Belgium. ISBN 978-2-87419-047-6, pp. 293-298
Hermkes, M., Kühn, N., Riggelsen, C.
-
2013. An Interactive Tool for the Elicitation of Subjective Probabilities in Probabilistic Seismic Hazard Analysis. Bulletin of the Seismological Society of America, 103(5), pp. 2862-2874
Runge, A. K.; Scherbaum, F.; Curtis, A. & Riggelsen, C.
-
2013. Simultaneous Quantification of Epistemic and Aleatory Uncertainty in GMPEs using Gaussian Process Regression. Bulletin of Earthquake Engineering
Hermkes, Marcel; Kuehn, Nicolas M. & Riggelsen, Carsten
-
2013. Visualisation of High-Dimensional Data Using an Ensemble of Neural Networks. IEEE Symposium on Computational Intelligence and Ensemble Learning, IEEE SSCI/CIEL'13
Gianniotis, N., Riggelsen, C.
