NASA Space Apps 2024 entry from team Seismic Surge.
Data for this challenge can be found at the NASA Space App Challenge Page Seismic Detection Across the Solar System. Click on the Space Apps 2024 Seismic Detection Data Packet to download the data and demo notebook as a zipped folder. Due to the size of the data directory, it is missing from this repository as it is included in the gitignore. The data can be set up in a local or Google Colab environment with the steps below.
Once you have downloaded the data packet and unzipped the folder, the data directory can be copied into the main directory for this cloned repository to be used locally.
To access the data through a shared Google Colab notebook, the project repository can be uploaded to a shared Google drive and then
mounted to the notebook using the code below. This is utilized in the nasa2024_colab_notebook.ipynb.
from google.colab import drive
drive.mount('/content/drive')%cd /content/drive/My Drive/space_apps_2024_seismic_detectionWith your environment ready, you can install the dependencies by running the following command from your terminal:
pip install -r requirements.txt
To simplify working with the data, we've created python scripts to extract data in bulk from data sets of interest,
and perform data cleaning/transformations as needed. There are two separate scripts, each for working with the csv and
miniseed data. Each script can be found in the scripts directory.
The extract_csv.py script can be used to extract and returned the desired csv data as a single pandas DataFrame. The extract_data
function within this script loops through the data sets of interest and concatenates them together while performing data cleaning and
transformations as needed. Each data set of interest can be obtained with the code below:
import extract_csv
lunar_train = extract_csv.extract_data('lunar', 'train', 'csv')
lunar_catalog = extract_csv.extract_data('lunar', 'catalog', 'csv')
lunar_test = extract_csv.extract_data('lunar', 'test', 'csv')
mars_train = extract_csv.extract_data('mars', 'train', 'csv')
mars_catalog = extract_csv.extract_data('mars', 'catalog', 'csv')
mars_test = extract_csv.extract_data('mars', 'test', 'csv')The extract_mseed.py script can be used to extract and returned the desired miniseed data as a list of ObsPy streams.
This can be done by executing the follwoing commands. Currently it is only set up for returning the mars training data.
import extract_mseed
data = extract_mseed.extract_mseed()The mars_demo_notebook.ipynb takes the lunar demo_notebook.ipynb, provided by Space Apps Challenge, and adapts the code to be capable of
handling data in the mars files, such as accounting for differences between file structures and frequencies measured.
- Bandpass - a function in obspy that helps with filtering data. In this case it filters out data between the min and max frequencies given.
- Highpass - similar to bandpass but handles frequencies above 10 hz, which is bandpass cap. this is the one mars data uses to filter, but the notebook will recognize that bandpass can't handle it and automatically shift to highpass. frequencies above the specified frequency are allowed through.
- Lowpass - frequencies below the given frequency are allowed to come through.
There are many different options for model development, such as pre-built neural networks for earthquake detection, event detection, or even designing a model from scratch.
There is a python package, EQTransformer, for signal detection and phase picking using AI (Tensorflow). This model is an intriguing option as it is tuned for earthquake detection. As noted in their paper, Earthquake transformer—an attentive deep-learning model for simultaneous earthquake detection and phase picking, phase picking refers to measuring the arrival time of the seismic P and S wave phases. These phases are useful for estimating the location of the seismic activity. This could allow for more accuracy when it comes to the start and stop times of such seismic events, or for using either P or S wave to determine the general time of an event. Both the P and S waves can be seen in the image below from their paper.
When working to install the package, we kept running into a ModuleNotFoundError: No module named 'distutils.msvcompiler'. Further documentation for EQTransformer can be found at EQTransformer Docs.
EventDetector is a Python library for deep learning-based event detection in multivariate time series data which utilizes TensorFlow. More information regarding their model design can be found in their paper Universal Event Detection in Time Series.
When it comes to designing a model from scratch, the purpose of an event detection model is to focus on change point detection (CPD) as noted in the paper A survey of methods for time series change point detection. This refers to the problem of finding abrupt changes in data over a change of time. In the case of seismic detection, the abrupt change being velocity registered by the seismic instruments.
The lunar_model_notebook.ipynb implements a machine learning model to distinguish between seismic events and noise in lunar mission data. By filtering out noise and identifying significant seismic activity, the model intends to ensure that only valuable data is transmitted back to Earth from lunar detection devices.
- Objective: Automatically detect and differentiate seismic events from noise in lunar seismic data to optimize data transmission.
- Data Source: Seismic data collected from lunar missions, stored in CSV files containing time and velocity measurements.
- Model Architecture: An LSTM (Long Short-Term Memory) autoencoder designed for time series anomaly detection.
- Hyperparameters: Adjust sequence_length to change the size of data sequences. Modify epochs and batch_size for training. Set a custom anomaly detection threshold based on your data.
- Model Architecture: Alter the number of LSTM layers or units to optimize performance.
As discussed in the demo notebooks, an approach of assessing the short term average to long term average (STA/LTA) can be used for this
change point detection. Working to replicate this approach, the mars_model.py script found in the scripts directory initiates the
process for such a model. This script takes all of the mars training data CSVs as one dataframe, and builds a character function
that can be used to detect change points in the data. Visualizations show subsets of the character function and the detected change over
time. Due to the large gap in time frames for the Mars training set, these visualizations are best viewed in small intervals (by original
file).
Below is a visualization of the trigger event detection for our Mars model from the entire Mars training data set. It shows a subset of the data from 2022-02-03.
