Predicting Earthquakes
Download
Report
Transcript Predicting Earthquakes
Predicting Earthquakes
By Lois Desplat
Why Predict Earthquakes?
To
minimize the loss of life and property.
Unfortunately, current techniques do not
have a high enough accuracy to be able to
accurately predict earthquakes.
Estimating earthquake probabilities
Scientists
study the histories of large
earthquakes in a specific area
The rate at which strain accumulates in
the rock
Methods to earthquake prediction
Need
to construct models based on:
Partial differential equations
Finite automata
Supervised learning techniques:
• Decision Tree
• Bayesian Classification
• Feed-Forward Neural Networks
Decision Tree
Tries
to generate rules with high accuracy
ID3, …
Bayesian Classifiers
They
are statistical classifiers
Only needs a small sample to find the
means and variances of the variables
necessary for classification
It can find the probability that a given
sample belongs to a certain class
(earthquake > 3.0)
Uses Bayes Theorem
Feed-Forward Neural Network
Network
given a set of input and
respective output to start learning
It connects each Perceptron and the
algorithm tries to minimize the weigths
between Perceptrons to the minimum so
that the input give the right output
The Bagging Method
Combine
the predictions of the past three
algorithms
You get a much more accurate prediction
Give different learning samples to each
algorithm
Some Problems
The
Data can have a lot of extra
information that adds noise
i.e. We might not want small scale
earthquakes that are really just
aftershocks of big earthquakes
We only look at the data in 1 dimension,
maybe if we plot the data in multiple
dimensions, we might some patterns
Not Good Enough!
Authors
claim that their bagging method
has 92% accuracy.
Highly doubt accuracy of that number but
even if true:
We still cannot predict earthquakes with
enough confidence
Solution
Do
short-term predictions instead of longterm
Analyze the data in multiple dimensions
over space, time and feature space.
Visualization of the Data Space
Data Space uses Magnitude, Epicentral
Coordinate, Depth and Time of occurrence
7D space uses:
NS: Degree of spatial non-randomness at short
distances
LS: Degree of spatial non-randomness at long
distances
CD: Spatial correlation dimension
SR: Degree of spatial repetitiveness
AZ: Average Depth
TI: Time Interval for the occurance of 100 events in
the sample space.
MR: Ratio of two events falling into different
magnitude ranges
Conclusion
This
method is able to find precursor
events just prior to an earthquake.
Unfortunately, it only works for short-term
predictions and cannot predict years or
months in advance.
Plenty of work can still be done!
References
“Predicting the Earthquake using Bagging Method in
Data Mining”, S.Sathiyabama, K.Thyagarajah, D.
Ayyamuthukumar
“A Bagging Method using Decision Trees in the Role of
Base Classifiers”, Kristína Machová, František Barčák,
Peter Bednár
“Cluster Analysis, Data-Mining, Multi-dimensional
Visualization of Earthquakes over Space, Time and
Feature Space”, Witold Dzwinel, David A. Yuen,
Krzysztor Boryczko, Yehuda Ben-Zion, Shoichi Yoshioka,
Takeo Ito
http://cse.stanford.edu/class/sophomorecollege/projects-00/neuralnetworks/Architecture/feedforward.html