|
Titel |
Estimating conditional probability of volcanic flows for forecasting event distribution and making evacuation decisions |
VerfasserIn |
E. R. Stefanescu, A. Patra, M. F. Sheridan, G. Cordoba |
Konferenz |
EGU General Assembly 2012
|
Medientyp |
Artikel
|
Sprache |
Englisch
|
Digitales Dokument |
PDF |
Erschienen |
In: GRA - Volume 14 (2012) |
Datensatznummer |
250064473
|
|
|
|
Zusammenfassung |
In this study we propose a conditional probability framework for Galeras volcano, which is
one of the most active volcanoes on the world. Nearly 400,000 people currently live near the
volcano; 10,000 of them reside within the zone of high volcanic hazard. Pyroclastic flows
pose a major hazard for this population. Some of the questions we try to answer when
studying conditional probabilities for volcanic hazards are: “Should a village be evacuated
and villagers moved to a different location?", “Should we construct a road along this
valley or along a different one?", “Should this university be evacuated?" Here, we
try to identify critical regions such as villages, infrastructures, cities, university to
determine their relative probability of inundation in case of an volcanic eruption. In this
study, a set of numerical simulation were performed using a computational tool
TITAN2D which simulates granular flow over digital representation of the natural
terrain. The particular choice from among the methods described below can be
based on the amount of information necessary in the evacuation decision and on the
complexity of the analysis required in taking such decision. A set of 4200 TITAN2D
runs were performed for several different location so that the area of all probably
vents is covered. The output of the geophysical model provides a flow map which
contains the maximum flow depth over time. Frequency approach - In estimating the
conditional probability of volcanic flows we define two discrete random variables (r.v.) A
and B, where P(A =1) and P(B=1) represents the probability of having a flow at
location A, and B, respectively. For this analysis we choose two critical locations
identified by their UTM coordinates. The flow map is then used in identifying at
the pixel level, flow or non-flow at the two locations. By counting the number of
times there is flow or non-flow, we are able to find the marginal probabilities along
with the joint probability associated with an event. Logistic regression - Here,
we define A as a discrete r.v., while B is a continuous one. P(B) represents the
probability of having a flow -¥ hcritical at location B, while P(A) represents the
probability of having a flow or non-flow at A. Bayes analysis - At this stage of
the analysis we consider only the r.v. A, where P(A) represents the probability of
having a flow -¥ hcritical at location A. We are interested in observing how the
probability of having a flow -¥ hcritical at location A is changing when data from the
model is taken into consideration. We assume a Beta prior distribution for P(A) and
compute P(A/data) using Maximum Likelihood Estimation (MLE) approach. Bayesian
network for causal relationships - Here, we are interested in more than two critical
locations and we are able to incorporate using a directed acyclic graph the causal
relationship between all the chosen locations. Marginal probabilities along with
the joint probability associated with an event based on the “causal links" between
variables. |
|
|
|
|
|