Abstract: Distributed acoustic sensing (DAS) has rapidly emerged as a transformative technology in seismic exploration, particularly in vertical seismic profiles (VSP). However, the acquired VSP data suffer from strong coherent DAS coupling noise and random noise. Current deep learning denoising methods, dependent on noise labels derived from conventional denoising techniques, fall short in addressing the unique noise properties inherent in DAS data. To address this challenge, we propose an exponential decay curve-constrained empirical mode decomposition (EDCC-EMD) analysis-based supervised denoising network. Our method begins with extracting the initial noise from the field DAS VSP data through the traditional EMD method. Despite containing some signal leakage, this noise is further processed through EMD to derive intrinsic mode functions (IMFs). We, then, analyze the correlation coefficients between these IMFs and the initial noise, applying an exponential decay curve (EDC) law to isolate pure noise. This refined noise data serves as accurate labels, enhancing the denoising network's precision. Meanwhile, most of the methods usually consider the t-x domain features and ignore the important frequency-domain features. Consequently, we train our network with frequency-domain data instead of time domain data, capitalizing on the more distinct separation of noise and signal characteristics, thereby facilitating more effective noise-signal discrimination. The experimental results demonstrate that our method significantly enhances the denoising performance and successfully recovers weak signals.