Abstract:In view of the problems of difficulty and high cost of obtaining labeled seismic data required for the training of existing models, we propose a new machine learning model CCLSN for seismic detection suitable for small samples. Through jointly using continuous wavelet time-frequency transform and redesigned lightweight convolutional neural network, the model greatly reduces the amount of labeled seismic data required for training and improves the applicability. The results show that CCLSN achieves high-accuracy and stable recognition function using a small-scale dataset containing only hundreds of samples. The accuracy and recall rate of CCLSN are both above 98%. CCLSN can provide a new technical approach for automatic seismic detection in areas with fewer or weaker earthquakes, such as central and eastern China.