Automatic Alert Level
Alert level problematics
A rarely addressed problem is the computation of a simple activity number (alert level), which serves as first measure for volcanic unrest (or criticality). A different formulation of the latter problem results in the questions: What is the optimum alert level system at a volcano? How can geophysical activity parameters help to estimate the alert level automatically and therefore give a first rough estimate of imminent risk to the decision makers? If the physical mechanism leading to an eruption would be known in detail, the definition of different alert levels would be a comparable easy task to do. However, most of the physics of ascending magma columns or even in hydrothermal systems surrounding a magmatic body are not observed directly. Therefore, a direct estimate of warning levels out of the geophysical data will not be possible for most of the volcanic eruptions but might be possible using statistical measures of all data recorded.
Automatic alert level estimation
The focus of this task, in contrast to the deliverable of WP5 “Real-Time Classification” of seismic data, will be put on the statistical properties of all data recorded by a volcanic surveillance. Using this overall statistic pattern will make an objective, automatic alert level estimation more reasonable.
The use of Hidden Markov Models (HMM) together with Bayesian Belief Networks (BBN), the currently investigated approaches in the research field of automatic alert level estimation (e.g., DAAD funded VIGONI project Wassermann LMU München and Carniel, University of Udine), is exactly reflecting the balance of knowledge and ignorance about the physical state of an active volcano. The statistical method of HMM, originally designed for creating robust speech recognition tools, revealed good performance when classifying seismic signals of volcanic origin using seismic array as well as seismic single station analysis (Ohrnberger, 2001; Alasonati et al., 2005) and local seismic network data (master thesis LMU München) as it is based on a double stochastic process where one controlling parameter space (i.e., the physics of the processes) is completely hidden to the observer.
The BBN can in turn evaluate the output of HMM and combine them with the output of other methodologies in a probabilistic, automatic and easily updateable evaluation of different volcanic scenarios (Aspinall et al., 2003). The major advantage when using BBN as classifier is embedded in the possibility to include user defined conditional probability (i.e., weights) into the algorithm, which is not as easy when using classical HMM approaches. Additionally, the BBN is insensitive to non significant features also in contrast to HMM, where the classification performance will be lowered by including more and more insignificant parameters. This and the user defined weights (i.e., expert knowledge about the features: seismicity - high weight - meteorological parameters - low weight) will create a more robust “quick start” which is important when the fast response system is deployed the first time. However, in a later stage the BBN can also be trained driven by the data. A drawback of BBN classification is its time independence, where in real world problems volcanic activity shows strong time dependence. Time dependent BBN could be constructed in principle but only when accepting complicated networks and high computer costs. The goal is therefore to combine both methods in order to deliver a ready for use automatic alert level system.
Following points will be evaluated during the project:
- Objective estimation of activity levels (classes)
- Estimation of useful features (multi-parameter data sets);
- Estimation of states in the HMM;
- Construction of BBN
- Tests of robustness using only small training sets;
- Significance tests;
- Physical models;
- Problem of data gaps;
An important question covers the significance and robustness of the applied methods. The main reasons for the failure of existing early warning schemes in volcanology originate from inconsistencies within data catalogues, changes in the classification scheme over the years, as well as possibly not carefully extracted features and actual differences in eruptive behavior. To account for this, tests using a synthetic data set of a virtual volcano should be used first in order to evaluate the sources of errors in more detail (e.g., how much data we need to train a model sufficiently). Using these tests, the next step will be the application on real data (Azores, Merapi, Soufriere Hills, Stromboli, Teide etc.). The designed on-line tool forms again a special service, which can be hooked up to the expert system. A prototype version of this online module will be available in early 2009 to be tested in the framework of WP 4.