|
|
| |
| Earthquakes Prediction Using Deep Learning
Download PDF
|
Basel Y. El-Habil, Samy S. Abu-Naser
|
Abstract:
The goal of this project was to build deep learning model to predict earthquake location and magnitude and we also wanted to deploy the best model to predict whether there is an earthquake or not in year time. The model will be fast and more reliable earthquake detection as current methods depend on non-learning algorithms, so the data set we used contains 4,172 global maps generated from the data of the National Oceanic and Atmospheric Administration (NOAA), the earthquake data set includes Latitude, Longitude and Magnitude, from 2150 B.C.E. to 2022. Our deep learning model was trained and tested using the LSTM depending on the target the location and magnitude. We used the predication next farm architecture model for predicting earthquakes architecture where the input is 64x64 images (maps). There's one convolutional layer a max pooling layer drop out layer flattening layer and three dense layers so the second type of model. We used LSTM and the time series data to see if we could remove the need to generate images as was needed for the CNN so the model was trained and tested on the raw data. It is a map earthquake shown here and the black circle here shows the predicated earthquake, so the LSTM architecture input the seismic maps then there was a simple RNN layer an LSTM layer and dropout layer. Another LSTM layer and two dense layers so for the results here is a comparison of the confusion matrixes for the CNN versus the LSTM they both had good accuracy precision and recall but the RNN was a little better with 98 to 99 accuracy precision and recall versus about 99 so for the S wave arrival time prediction. There are some more results this is graphs of the predicted value versus the observed value for the RNN and for the LSTM a perfect prediction is indicated by black circles and you can see that the mean error value for the RNN is quite a bit lower than the loss error value for the LSTM. The RNN was the best model and overall here is a map of the four targets that we tried to predict which was classifying earthquakes versus noise predicting magnitude, location and time arrivals. We convert the earthquakes data to maps and sent this to one npy file. The python function runs the model and predicts the image automatically.
|
|
|
| | |