A DDUNet-Based Denoising Method for Improving Daytime Spaceborne Backscatter Lidar Data

Clouds and aerosols have numerous impacts on the Earth’s weather, air quality, and climate system. Spaceborne backscatter lidars provide critical data products for studying cloud and aerosols, such as feature top/base height and type (ice cloud, liquid water cloud, dust, smoke, etc.), as well as vertical profiles of properties such as depolarization ratio and extinction coefficient. However, large random errors in daytime retrievals of these parameters hinder their utility due to degradation of spaceborne lidar signals by solar background. While vertical and horizontal averaging can reduce these random errors, these data products must then be reported at courser spatial resolutions. We have developed a deep learning model for denoising daytime spaceborne lidar systems that employ photon counting detection, such as Cloud Aerosol Transport System (CATS) and Advanced Topographic Laser Altimeter System (ATLAS) on the ICESat-2 satellite, enabling improvements in signal-to-noise ratio (SNR) and data products at finer resolutions. In this presentation we apply a DDUNet autoencoder denoising method to CATS and ATLAS data, as well as spaceborne simulations using airborne data. These denoised datasets are compared to the standard processing technique (averaging) and the simulation input (truth) to estimate the performance of cloud detection frequencies, cloud top heights (CTH), depolarization ratios, and extinction coefficients. The DDUNet denoised dataset with a 2 km horizontal resolution increased the daytime SNR by a factor of 2.5 (on average), enabling better ice cloud detection and CTH estimates compared to averaging data to 60 km. Ice cloud detections are very similar to averaging to 60 km, but with 30X finer resolution, limiting the false positives created by “smearing” features. CTH errors are reduced from 368 m using 60 km averaging to 285 m using DDUNet denoised data. These improvements to layer detection and SNR then enable more retrievals of depolarization ratio and extinction coefficients during daytime with less random error.