Denoising of Coherent Doppler Wind Lidar Based on Parallel U-net of Denoising and Segmentation

The suppression of noise enhances the detection range and accuracy in low signal-to-noise ratio regions for coherent Doppler wind lidar (CDWL). This paper presents a denoising method for CDWL using a Parallel U-net. Specifically, the model features two paths: one for direct denoising and the other for integral-segmentation-differentiation denoising. By integrating the results from both branches, the Parallel U-net combines the high accuracy of the direct denoising branch with the low error rate benefit of the segmentation branch. To further validate the network's correctness and effectiveness, a method capable of generating random continuous wind fields was proposed. Consequently, tests utilizing this method revealed that the detection range of the Parallel U-net, using RMS equal to 1 m/s as a standard, was approximately 22% higher compared to the centroid method. Moreover, as an evaluation of the network's performance in real-world conditions, experiments conducted on an all-fiber CDWL indicated that the Parallel U-net nearly doubles the wind speed detection distance compared to the centroid method and improves the range by approximately 20% over a single-path U-net.