Abstract: Based on the “redness and swelling” hypothesis, the classification algorithm of data mining is applied to carry out earthquake prediction experiments on the sample set which is composed of microtremor data statistics and past earthquake cases. We chose seismic pairs that meet the requirements of magnitude, epicenter distance, occurrence time interval, and unaffected by typhoons conditions, and use the tail earthquakes as prediction targets. The standard deviation of microtremor data in time windows is calculated, and the standard deviation data is standardized by Z-score standardization method. Then, the median of the last group of standardized data from the nearest three stations to the epicenter is selected as the positive sample data, and the median of the seismic quiet period data of each station is selected as the negative sample data. Finally, the positive and negative sample data are constructed into the sample set. CART, GBDT and SVM methods are used to construct the prediction model on this sample set, and 5 fold cross validation method is used to evaluate the prediction model. The results show that: ① there is a relationship between earthquakes and microtremor changes, and there are microtremor anomalies before earthquakes (M≥6.0) occurred. ② The positive samples that magnitude over 6 have a greater influence on the model construct. ③ The SVM algorithm is more suitable for the small sample data environment of this paper.

Key words: Microtremor, Classification, Earthquake forecast