How humans (tried) to forecast eruptions

There are a lot of indicators to analyze eruptions ( SO2, groundwater levels...)

For the sake of the contest, we will be focusing on seismic data here!

Events in seismic data before eruptions

Events in seismic data before eruptions

Volcano-tectonic earthquakes (VT)

• Caused by the movement of magma and fluids within the Earth's crust
• Weeks to months before the eruption
• Usually have frequencies in the range of 1 - 30 Hz                                                             

Low-frequency earthquakes (LFEs)

• Long earthquakes caused by fluid and earth crust movement 
• Hours to days before the eruption
• Usually have frequencies below 1 Hz                                                             

Events in seismic data before eruptions

Long-Period earthquakes (LP)

• High-frequency earthquakes caused by the sudden release of gas and magma
• Days to weeks before the eruption
• Usually have frequencies below 1 Hz          

Hybrid earthquakes (HYB)

• Has the characteristics of VT and LFEs
• Indicates the ascent of magma
• Usually have frequencies around 1-10 Hz     
• Can occur any time, usually accompanied
  with other types of events   

The forecast of Mount Pinatubo, 1991

April ~ Early May : Hybrid earthquakes

Early June : Volcano Tremors generated by the movement of magma, accompanied by an increase in seismic energy release. (started alerts and evacuations)

June 7 : Shift from VT to LP, indicating increasing gas pressure

June 12 : Sharp increase of number and size of LP events, 

meaning the release of gas and steam

June 15 : Eruption took place

Why current methods are unreliable?

• Hard to generalize to different volcanos
• Current methods have short warning times
• Most volcanos are NOT LINEAR
  -> doesn't behave like what we thought it should be ( ex: decrease of earthquake movements )

"Our capability to forecast eruptions is still limited, with ~20% of eruptions accurately forecasted. "

Introduction to INGV - Volcanic Eruption Prediction Contest

https://www.kaggle.com/competitions/predict-volcanic-eruptions-ingv-oe

Goal: Predict time_till_eruption for a volcano with ten 10 minutes of seismic data

Scoring method:

Mean Average Error (MAE)

\( \frac{\sum_{i=1}^{n}|y_i-f(x_i)|}{n}\)

My approach to the contest

Data Analyzing / Preprocessing

time_to_eruption range: 0.07 ~ 567 days

mean : 264 days

Data Analyzing / Preprocessing

Quite an average spread, only decline after ~4e7

Data Analyzing / Preprocessing

Data Analyzing / Preprocessing

Idea: aggregate the data and play with that to prevent dealing with time series!

Time series data is hard to deal with :( 

Data Analyzing / Preprocessing

Selected aggregated features:

1. sum 
2. min
3. max
4. mean 
5. standard deviation (std)
6. median
7. skewness (skew)
   -> how asymmetrical a set of data is
8. kurtosis
   -> checks for the "peakness" of the data

(of each sensor)

Data Analyzing / Preprocessing

Data Analyzing / Preprocessing

Data Analyzing / Preprocessing

Ready to train!

Trying out a regression approach

Linear Regression 

\(f(x) = w_0 + w_1x_1 + w_2x_2....+w_px_p\)

Linear Regression tries to find the best set of coefficients [\(w0,w1,w2....wp\)] that minimizes Sum of Squared Errors(SSE)

Trying out a regression approach

Linear Regression 

('sensor_1_mean', 59030887170688.34)
('sensor_1_skew', 4048663.024784088)
('sensor_10_skew', 2595411.6050720215)
('sensor_3_skew', 1270106.8636779785)
('sensor_8_skew', 978927.2803401947)

('sensor_7_skew', -165900.9322490692)
('sensor_9_skew', -1518843.438412428)
('sensor_6_skew', -2284296.6836452484)
('sensor_4_skew', -3355988.1515922546)
('sensor_1_sum', -983831722.500612)

Trained Coefficients

skew is generally helpful, while only certain mean and sums are useful

Trying out a regression approach

Linear Regression 

Trained Coefficients

median, mean (most of them), and sum (most of them) aren't very impactful

Trying out a regression approach

Linear Regression 

Private: 12002016 (~139 days)

Public: 11483340 (~132 days)

Trying out a regression approach

LightGBM Regressor 

Light weight gradient boosing machine proposed by Microsoft in 2017.

Often used in Kaggle competitions because of light weight and efficiency.

Trying out a regression approach

LightGBM Regressor 

Feature importances (Can't plot because too big)

std and kurtosis generally helped alot, while median and mean didn't have much importance 

Trying out a regression approach

LightGBM Regressor 

Private: 7192681 (~83 days)
Public: 7334062 (~85 days)

Things to improve

More Features

• 10, 25, 90 quantile data

• Get frequency data with FFT

Different Models

• XGBoost

• CatBoost 

Model Ensembling

Things to improve

CatBoost baseline + ts fresh

773 features vs 8 features!

Private : 4999330 (~57 days)

Public : 4989282 (~57 days)

Things to improve

[1st Place] Mel Spectrogram + 773 features

Private : 3438191 (~39 days)

Public : 3641873 (~42 days)

References

Power, J. A. (2001, December). Seismicity and Forecasting of the 1991 Eruption of Mount Pinatubo: A Ten-Year Retrospective. NASA/ADS. https://ui.adsabs.harvard.edu/abs/2001AGUFM.U31A..03P/abstract

Volcanic Earthquakes. (n.d.). PNSN. https://pnsn.org/outreach/earthquakesources/volcanic

Williams, R. G. (2021, June 1). A Burp or a Blast? Seismic Signals Reveal the Volcanic Eruption to Come. Quantamagazine. https://www.quantamagazine.org/seismic-data-helps-scientists-forecast-volcanic-explosions-20210601/

Volcano-Tectonic Event. (2002, October). NOVA. https://www.pbs.org/wgbh/nova/volcano/seis_vte.html

References

Long Period Event. (2002, October). NOVA. https://www.pbs.org/wgbh/nova/volcano/seis_lpe.html

How Do Scientists Forecast Eruptions? (n.d.). Smithsonian Institution National Museum of Natural History Global Volcanism Program. https://volcano.si.edu/faq/index.cfm?question=eruptionforecast

Jesper, D. S. (2021). Introduction to Volcanology, Seismograms and LGBM. Kaggle. https://www.kaggle.com/code/jesperdramsch/introduction-to-volcanology-seismograms-and-lgbm

Alexander, L. (2021). Ingv_catboost_baseline+tsfresh. Kaggle. https://www.kaggle.com/code/carpediemamigo/ingv-catboost-baseline-tsfresh

Jei, F. (2021). [1st Place] Mel Spectrogram + Blended ResNets. Kaggle. https://www.kaggle.com/competitions/predict-volcanic-eruptions-ingv-oe/discussion/211315

ChatGPT, personal communication, February 11, 2023