Start Registration
2021-01-01
Start Competition
2021-04-04
End
2021-10-31
Start Registration
2021-01-01
Start Competition
2021-04-04
End
2021-10-31

Competition Rules

Sponsor:

Organizer:

Abstract

"AETA Earthquake Prediction AI Algorithm Competition" aims at mining the correlation between precursor observation data and earthquake three elements through innovative algorithms, discovering abnormal signals and features related to impending earthquakes, and building earthquake prediction models based on historical observation data and earthquake catalogue, in the hope of promoting the solution of scientific problems of earthquake prediction and forecast. At the same time, we also hope that through this competition, more attention and participation of people from all walks of life are attracted and more new technology and new methods will be applied in earthquake prediction and forecast.

In this competition, 95 kinds of feature data of electromagnetic disturbance and geo-acoustic data from AETA seismic monitoring and prediction system in the Sichuan-Yunnan region in the past four years will be shared, as well as the earthquake catalogue from CENC. This competition focuses on evaluating the innovation and advancement of feature extraction and sample construction methods, as well as the applicability and accuracy of earthquake prediction models, and encourages innovative algorithms to solve scientific problems in earthquake prediction.

Schedule

There is no charge for participants to register online. The competition is planned to last 1 year. 3 months for registration and model construction, 7 months for earthquake prediction, 1 month for review and result publication, and the award ceremony will be held at the end. The specific arrangements and instructions are shown in the table below.

Competition Schedule

The competition scheduletimeinstructions
registrationJan. 1 2021 to Mar. 31 2021Registration, team building and model construction
Propaganda and trainingJan. 1 2021 to Mar. 31 2021Propaganda on campus, live streaming, release of Baseline, Case study
Competition periodApr. 4 2021 to Oct. 31 2021Make predictions for the coming week on each Sunday
Code and documentation SubmissionNov. 1 2021 to Nov. 14 2021Shortlisted teams submit code and documents for verification of model and predictions
Defense and ReviewNov. 15 2021 to Nov. 30 2021Shortlisted Teams do presentation to review committee.
Release competition resultsDec. 1 2021 to Dec. 7 2021Publish the list of 18 winning teams
Award ceremonyLate Dec. 2021The winning teams and guests attend the ceremony

Notes for participant

The competition is open to the whole society, except for the members of the organizing committee and the review committee.

(1) One or more people can be a team;

(2) Each participant can only join one team, otherwise he/she will be disqualified;

(3) Ensure that the registration information is accurate and valid.

Problem and Data

Problem

Earthquake prediction for the coming week on each Sunday based on historic data from AETA network in Sichuan and Yunnan region. The magnitude of target earthquake should be equal or larger than 3.5. The target region is 22°N -34°N, 98°E -107°E. For an earthquake with a magnitude of 3.5 or larger in the target area, it will not be counted if there is no AETA station within 100km.

A training data include precursor data and earthquake catalogue will be supplied for model construction once participants finish the registration. The real-time data for prediction will be updated weekly since Jan. 1th 2021. Starting from Apr. 4, 2021, all teams should start to make prediction for the coming week on each Sunday based on past data. 30 weeks prediction will be made.

Competition data

  • Training data for model construction

    95 kinds of feature data of electromagnetic disturbance and geo-acoustic will be given to all teams. The time interval of each data is 10 minutes which is marked by timestamp. The specification of 95 kinds feature data will be in a read me file. The time period of data is from Oct. 1th 2016 to Dec. 31th 2020. In addition, an earthquake catalogue of ≥3.5 seismic events in the target region is also provided.

  • Real-time data for prediction

    From Jan. 1th 2021, 95 kinds of feature data of electromagnetic disturbance and geo-acoustic will be updated each week. Teams can download data by week. There are two ways to download data. One is to download data from the website manually. The other one is to download data automatically through login into the data server by an executable program which will be supplied by the host.

Submission of prediction

The submission channel will open at 00:00 and close at 12:59 every Sunday. Each team can login the website and submit its prediction manually. And they also can login the sever and submit their prediction automatically. Each team will have a code and instruction guide for server login. During the open of submission channel, each team can change their prediction and submit several times. And the latest submission will replace the old one. The result of prediction will be released after 00:00 every Monday.

After Competition is end, each team should submit the model code and document (PPT form). It specifically includes the description of data processing methods, feature extraction methods, sample set and the architecture of prediction model, as well as a review of its 30 weeks prediction, etc. The submitted content includes design documents, algorithm and model description, source code, etc. The source code should be directly run in C/C++/Python/ Matlab environment and output the results. For the convenience of review, please provide the source code as well as the operating environment, input and output guide for the review committee to confirm the authenticity of the results.

Scoring rules

We updated Scoring Rules on March 20th. Compared to previous rules, a correct non-earthquake prediction is granted a fixed score for Part II. Contact us with e-mail or QQ group if you have trouble with Scoring Rules.

The competition organizers will invite famous professors and experts to set up a review committee. This committee is responsible for formulating review rules and review the algorithms and models of each team at the end of competition. The evaluation of each team by the committee is worth 10% of the final score. Here will illustrate how we calculate the prediction score of each team which is worth 90% of the final score.

The prediction score of our competition contains 2 parts:

Part I. Score of forecast on the occurrence of weekly earthquakes (Y/N)

Part II. Score of each week you predict correctly

Part I contains:

  1. Accuracy, ratio of correctly-predicted weeks to total. Higher Accuracy results in higher score.
  2. F1 Score, the harmonic mean of Precision and Recall. Higher F1 results in higher score.
  3. False Positive Rate, ratio of incorrectly-predicted non-earthquake weeks to total non-earthquake weeks. Lower FP Rate results in higher score.

Part II calculates total value of scores of each correctly-predicted earthquake week (each TP case). Score is decided by your prediction of the position and magnitude of the epicenter (P&M). Higher accuracy of P&M results in higher score. The weekly score:

  1. has a positive correlation with real magnitude. Higher RM results in higher score.
  2. has a negative correlation with deviation of position. Lower DoE results in higher score.
  3. has a negative correlation with deviation of magnitude. Lower DoM results in higher score.
  4. Highest score is picked if there are multiple eligible earthquakes (Magnitude ≥3.5 and Distance to nearest AETA Station <100km) in one week.

Part II also calculates total value of scores of each correctly-predicted non-earthquake week (each TN case). There is a fixed score for each non-earthquake week. So The score of Part II = Sum of scores of correctly-predicted earthquakes + Fixed score for non-earthquake * correctly-predicted non-earthquake weeks

Teams are regarded as submitting an incorrect prediction of Y/N if they miss submission:

  1. For earthquake weeks, they submit a non-earthquake prediction.
  2. For non-earthquake weeks, they submit an earthquake prediction.
  3. Score of Part II is not affected.

The distance calculation formula we use is as follows (lon1/lat1 are actual longitude/latitude of epicenter, lon2/lat2 are predicted longitude/latitude for epicenter):

from math import radians, cos, sin, asin, sqrt
def get_distance(lon1, lat1, lon2, lat2):
    lon1, lat1, lon2, lat2 = map(radians, [lon1, lat1, lon2, lat2])
    dlon = lon2 - lon1
    dlat = lat2 - lat1
    a = sin(dlat/2)**2+cos(lat1)*cos(lat2)*sin(dlon/2)**2
    c = 2*asin(sqrt(a))
    r = 6371
    distance = c*r
    return distance

Award setting

The results of this competition will be confirmed and approved by the review committee and posted on the official website. First prize, second prize, third prize and encouragement award are set in this competition, with a total bonus of 300,000 yuan.

  • First prize: 1 team, 100,000 yuan (pre-tax).

  • Second prize: 2 teams, 50,000 yuan (pre-tax).

  • Third prize: 5 teams, 20, 000 yuan (pre-tax).

  • Encouragement Award: 10 teams.

  • Participants Award: all teams that successfully complete this competition will be awarded as well.

The organizer will award trophies and certificates to all winning teams, and invite representative of all winning teams and partial participants award teams to the award ceremony for technical exchanges, and provide round-trip travel and accommodation expenses.