Can a neural net be trained to solve this problem?

We are interested in exploring Matlab's deep learning toolbox to solve analog seismogram records. An example seismogram image with seismic traces is shown below. The traces are time series data. In all images, the center of the traces have been identified (the red lines) and saved elsewhere as x-y raster data. These data are necessarily disconnected from one another (i.e. there are always gaps in the data). In addition to the traces, the so-called mean lines - the zero energy line about which the traces oscillate - are also identified. The challenge is to associate each trace segment with its appropriate mean line. Some are easy, but during seismic activity it can be seen that traces can cross multiple mean lines.

We have access to a potentially large number of training set data, and would like to understand if this training data could be used to train a deep learning network.

Thanks for any suggestions as to how we might begin to investigate this problem. Please feel free to check out the image and data set at seismo.redfish.com.

Answer 1

[Swetha Polemoni]

Hi

It is my understanding that you want to categorise trace segment to its mean line(class). You can use Classification for these kind of problems. Hope you find this helpful.

[Lowell Bartlett]

Thanks for the response. Is there any particular type of classification method that you think would be most applicable so we can focus our research? Is it possible to establish correspondence with Mathworks staff on the issue to bounce ideas off, etc. ?