Earthquake early warning (EEW) is a method to reduce loss from earthquake hazards. Sensors are placed in earthquake-prone areas, and using the property that electromagnetic waves travel faster than seismic waves, people can get earthquake warning information (earthquake magnitude [21], earthquake source [16], etc.) before the seismic waves reach the warning area [[1], [2], [3], [4]]. DAS uses optical fiber as the sensing unit, which can collect seismic signals in hard-to-reach places, such as underwater and glaciers [5,6]. In addition, as the deployment of optical fiber communication infrastructure in cities and oceans continues to expand, DAS can convert existing dark optical fibers in cities and oceans into dense seismic array networks, significantly reducing the arrival time of seismic waves to sensors [[7], [8], [9], [10]]. Thus, how to quickly and accurately identify seismic signals becomes the primary issue for DAS in EEW.

The weak backward Rayleigh scattering signal of DAS causes seismic signals to be easily overwhelmed by environmental and system noise, making accurate identification of seismic signals challenging [17,18]. In recent years, deep learning has been applied to earthquake recognition. In most representative works, Pablo D. Hernandez [11] et al. constructed three different one-dimensional neural network models for 60s earthquake samples collected by DAS to extract and classify one-dimensional signals, and the overall recognition rate reached more than 90%. Lv [12] et al., proposed the ADE-Net deep learning network to identify 5.12s of earthquake samples monitored by DAS, which converted 5.12s of seismic data into time-channel maps and its accuracy could reach 80.4%. However, the 5.12s consuming time is still difficult to meet the practical needs. How to achieve accurate seismic recognition in shorter time is an urgent problem to be solved in EEW with DAS.

In this paper, a DAS earthquake recognition method is proposed with QRE-net. The one-dimensional DAS samples are used for feature extraction and identification, the earthquake-duration requirement for samples is also reduced. In the experiment, three kinds of seismic signals with different effective durations (1s, 2s, 3s) are used to verify its effectiveness. The results show that the QRE network has an accuracy of 88.74% in the effective earthquake duration of 1s. It is believed that this method provides a fast DAS seismic identification method, which will promote DAS technology to play the great potential of existing communication cables covering the world, and is expected to realize global earthquake disaster detection and early warning.