Reconstruction for beam blockage of lidar based on generative adversarial networks

Doppler lidar is an active laser remote sensing instrument. However, beam blockage caused by low-altitude obstacles is a critical factor affecting the quality of lidar data. To reconstruct the line of sight velocities (LOSV) in areas with beam blockages and to evaluate the effectiveness of reconstruction results, the LOSV-filling network (LFnet) approach based on generative adversarial networks (GANs) and an evaluation scheme based on the degree of blockage are proposed in this paper. The LFnet comprises two adversarial models. The first adversarial model captures the structural features of LOSV to output the edge map, and the second adversarial fills in the blockage area using the edge map. We have built a packaged dataset consisting of training, validation and test datasets with mask sets. Then the sensitivity of the reconstruction effectiveness with different shielding conditions is studied, to reveal the mechanism of shielding influencing the reconstruction. A series of indicators were used to evaluate the model's performance, including the traditional indicators and the proposed indicator of root mean square error (RMSE). Finally, LFnet was demonstrated in a practical application in an airport. The complete process of an easterly gust front is reconstructed with RMSE less than 0.85 m/s, which has significance for flight safety.