ADE-Net: A Deep Neural Network for DAS Earthquake Detection Trained With a Limited Number of Positive Samples

Abstract

The distributed acoustic sensing (DAS) technology that has emerged in recent years, combined with existing optical cables for earthquake monitoring, has the advantages of low deployment cost and dense sampling capability and has been used in various environments. However, the massive data generated by DAS are difficult to process manually, and there is an urgent need for efficient automatic earthquake detection algorithm. The deep neural networks trained with manually processed earthquake data in the past decade have been introduced. However, available manually processed earthquake DAS data are very limited, which makes it difficult to train neural networks. A convolutional neural network called “APM-DAS-earthquake network (ADE-Net) ” was built and trained with only five earthquakes’ records in Tangshan, China. The neural network detected about 2.56 times earthquakes on a DAS dataset than the permanent seismic network’s catalog with an accuracy rate of 80.4% and a recall rate of 93.18%. Results of tests with synthetic datasets and another real dataset suggest the neural network works well with a variety of earthquake location, magnitude, type, and array geometry. The computational time cost with a PC for 10-s data is less than 2.5-s data, which demonstrates its feasibility on time-sensitive applications, for example, earthquake early warning.