Distributed acoustic sensing (DAS) technology is widely used in seismic monitoring, intrusion detection, and other fields due to its advantages of wide monitoring range and low cost. However, the problems of complex signal processing and large data volume limit its applications. This paper proposes a downsampling method based on short-time Fourier transform, which reduces the length of the time series while retaining high-frequency information. Experiments show that this method improves the efficiency and classification performance of the model, with an F1 value of 0.914 7 on a four-class private dataset and an accuracy of 0.994 4 on a two-class public dataset.