Introduction

We are a group of physics engineering students from the University of São Paulo in Brazil. Our mission is to encourage the study of programming, software development, and physics through hands-on activities. By tackling real-world challenges, we aim to foster learning and share knowledge among our members, while pushing the boundaries of innovation.

The main languages used were Python and Jupyter Notebooks. All the libraries used are listed in the references. VSCode and Google Colab were used to deal with the programming. Also, GitHub was used for the versioning.

For our brainstorm and organization process, Miro was used. Canva was used for developing the presentation.

Flowchart elaborated:

upload_image.01:31:52.360883

Collecting Data

First and foremost, we downloaded the data available in NASA's dataset from Apollo's 12, 15, 16 and InSight's missions located in the Space Apps 2024 Seismic Detection Data Packet in the challenge's resources. However, the neural network training requires a larger dataset to be trained efficiently, therefore we accessed the SAGE dataset in order to provide with more data to get more accurate results. We got the aforementioned data from the PyWeed utility at the SAGE website. By selecting a period of time and limiting the affected area, we can choose which stations we want to get the quakes' signals from. Then, we get waveforms, that we manually had to remove due to incorrect data, which does not contain any earthquakes. The output file is in the ".mseed" format, and will be used by side of the lunar provided data.

upload_image.01:18:45.778999

Data Preprocessing

For the preprocessing of the raw seismological data collected in the “.mseed” format of earth’s seismological events, a Python system was created to apply a high-pass filter to the seismic wave data, enhancing the relevant signals while reducing noise.

Before filtering:

upload_image.01:23:09.645638

After filtering:

upload_image.01:24:09.981097

After filtering, the program utilizes five seismic event detection algorithms: Recursive STA/LTA, carlSTAtrig characteristic function, Classical STA/LTA, Delayed STA/LTA, and the Z-detector. These algorithms, implemented in the ObsPy open source library, identify and timestamp seismic events, ensuring a comprehensive and diverse dataset for the neural network’s training, thus improving the accuracy and efficiency of the model.

upload_image.01:29:14.101717

Fourier_View

A class was created in order to facilitate the data visualization in the Fourier Space. This was done because the spectrogram analysis was very useful in our work. The class has five instances: load_data, compute_spectrograms, plot_fourier, plot_2d_heatmaps and plot_3d_heatmaps.

The first two are necessary for loading the data from a ".mseed" file and computing it spectrograms, respectively. The instance plot_fourier was developed in order to calculate and to provide the data in the Fourier Space.

Event7_Fourier

The plot_2d_heatmaps provides the spectrogram of the event. This analysis is helpful specifically when it comes to Power analysis.

Ev7_Spectrogram

Finally, in order to get a complete overview, the plot_2d_heatmaps instance comes in scene. It provides a full map of the dataset and by using the zoom tool, more details can be observed easily.

Even7_3d The Use of CNN

For the prediction of the arrival time, we decided to use Convolutional Neural Network (CNN), a potent tool for visual feature analysis and learning. Specifically, CNNs are constructed and trained using 710 images, each with a resolution of 64 × 64 pixels. The chosen architecture incorporates convolutional layers intercalated with polling layers and the “ReLu” activation function. Following calibration, training and validation metrics (R² metric) of 82,44% and 46,21% have been achieved. The data analysis is provided below.

Data Analysis

upload_image.02:04:57.617077

At the left, we have the obtained loss function vs the number of epochs for both training and validation processes. At the right, we follow the progress of a metric called R². For the R², negative values means that the model is worse than outputting the average. Values between 0 and 1 describes how good is the model, where 0 is outputting the average and 1 means the model is perfect for leading the task.

upload_image.02:06:49.797828

Here we can see how close are the estimated for the test set of the arrival time. The best ideal case would be where all points match the red line.

Results

The results are available in the project's GitHub page, in the file "catalog.csv" it describes the results of the test dataset from the Space Apps 2024 Seismic Detection Data Packet in the following way: "Index, filename, time_rel(sec)".

2024 NASA Space Apps Challenge

Demarianos