Abstract. Cloud radar Doppler spectra are of particular interest for investigating cloud microphysical processes, such as ice formation, riming and ice multiplication. When hydrometeor types within a cloud radar observation volume have different terminal fall velocities, they can produce individual Doppler spectrum peaks. The peaks of different particle types can overlap and be further broadened and blended by turbulence and other dynamical effects. If these (sub-)peaks can be separated, properties of the underlying hydrometeor populations can potentially be estimated, such as their fall velocity, number, size and to some extent their shape. However, this task is complex and dependent on the operation settings of the specific cloud radar, as well as atmospheric dynamics and hydrometeor characteristics. As a consequence, there is a need for adjustable tools that are able to detect peaks in cloud radar Doppler spectra to extract the valuable information contained in them. This paper presents the synergistic use of two algorithms used for analyzing the peaks in Doppler spectra: PEAKO and peakTree. PEAKO is a supervised machine learning tool that can be trained to obtain the optimal parameters for detecting peaks in Doppler spectra for specific cloud radar instrument settings. The learned parameters can then be applied by peakTree, which is used to detect, organize and interpret Doppler spectrum peaks. The application of the improved PEAKO–peakTree toolkit is demonstrated in two case studies. The interpretation is supported by forward-simulated cloud radar Doppler spectra by the Passive and Active Microwave TRAnsfer tool (PAMTRA), which are also used to explore the limitations of the algorithm toolkit posed by turbulence and the number of spectral averages chosen in the radar settings. From the PAMTRA simulations, we can conclude that a minimum number of n = 20–40 spectral averages is desirable for Doppler spectrum peak discrimination. Furthermore, small liquid peaks can only be reliably separated for eddy dissipation rate values up to approximately 0.0002 m2 s−3 in the simulation setup which we tested here. The first case study demonstrates that the methods work for different radar systems and settings by comparing the results for two cloud radar systems which were operated simultaneously at a site in Punta Arenas, Chile. Detected peaks which can be attributed to liquid droplets agree well between the two systems, as well as with an independent liquid-predicting neural network. The second case study compares PEAKO–peakTree-detected cloud radar Doppler spectrum peaks to in situ observations collected by a balloon-based holographic imager during a campaign in Ny-Ålesund, Svalbard. This case demonstrates the algorithm toolkit's ability to identify different hydrometeor types but also reveals its limitations posed by strong turbulence and a low n. Despite these challenges, the algorithm toolkit offers a powerful means of extracting comprehensive information from cloud radar observations. In the future, we envision PEAKO–peakTree applications on the one hand for interpreting cloud microphysics in case studies. The identification of liquid cloud peaks emerges as a valuable asset, e.g., in studies on cloud radiative effects, in seeder–feeder processes, or for tracing vertical air motions. Furthermore, the computation of the moments for each subpeak enables the tracking of hydrometeor populations and the observation of growth processes along fallstreaks. On the other hand, PEAKO–peakTree applications could be extended to statistical evaluations of longer data sets. Both algorithms are openly available on GitHub, offering accessibility for the scientific community.
Read full abstract