Support Vector Machines

For a proof of concept, we tried to train a support vector machine to distinguish between quiet time and quake time periods. The results are promising!

The support vector machine was trained on 7 different earthquake for a set of stations for the quiet time and the quake time. Currently, the first pick is considered as the boundary between the quiet time and quake time periods. A set of random data points- 3 random stations for each quake for both quiet time and quake time was chosen as the training data.

We had a dataset of 2240 data points in all.

Results:

True positive: Correctly predicted quake time as quake time

True Negative: Correctly predicted quiet time as quiet time

False Positive: Incorrectly predicted quiet time as quake time

False Negative: Incorrectly predicted quake time as quiet time

Training Data:

The Support Vector Machine seemed to be able to fit the training data well enough.

True positive: 78.93%

True Negative: 84.19%

False Positive: 15.81%

False Negative: 21.07%

Test Data:

The Support Vector Machine was also able to correctly classify the test data.

True positive: 80.00%

True Negative: 88.57%

False Positive: 11.43%

False Negative: 20.00%

The idea of detecting earthquakes by capturing the differences in the frequency domain properties, seems logical now. We will try changing a few parameters to figure the best possible way for the classification to get optimal results.