A Machine Learning Approach for Intelligent VAD Wind Estimation with S-band Polarimetric Weather Radars

The Velocity Wind Profile (VWP), one of the most widely used NEXRAD products, estimates the horizontal wind vector as a function of height using velocity azimuthal display (VAD) technique. Typically, VAD estimate the wind speed and direction by fitting a sinusoidal model to a ring of radial velocities from all the azimuth directions at a constant height of a PPI scan, under the assumption of a uniform wind field. It has been reported that VAD derived wind can be biased if a large amount of the radial velocities is contaminated by birds, that can generate independent flight velocities, such as during nocturnal bird migration. On the other hand, good wind tracers like insects, are more dependent on the wind velocity of interest, producing more reliable wind estimates. Thus, the overall accuracy of the VWP can be improved by excluding bird contaminated velocities in the VAD processing.

This work focuses on improving VAD wind estimation from clear air biological echoes by mitigating bird contamination. We propose an Intelligent VWP (IVWP) that adaptively filters out velocities from bird echoes in the VAD processing. IVWP operates in three main stages. First, the Hydrometeor Classification Algorithm (HCA) is used to detect biological echoes. The second stage involves the machine learning Bird-Insect Ridge Classifier (BIRC), developed to predict the probability of bird presence at each gate. BIRC is applied to biological gates and thresholds on its probabilities used to exclude bird echoes. The final stage is VAD wind estimation. The performance of IVWP is assessed using data from the operational NEXRAD radar, KOHX (Nashville, Tennessee) over the month of May 2018. This period was selected due to observed diurnal cycles of nocturnal bird migration and daytime insect domination. Wind bias is used as the evaluation metric, calculated as the deviation of the predicted VAD wind from the true wind. Moreover, Rapid Refresh (RAP) wind predictions are assumed to be the true wind. First, the profile of the biases from conventional VAD (i.e., without IVWP) are shown in Fig. 1. It can be observed that the biases match the expected diurnal migration cycle, with large nocturnal and low daytime values. Further verification using BIRC confirmed dominant bird presence in the region with the largest biases (>200 m height, and within 1-11 UTC), and dominant insect presence in the low bias region (11 – 24 UTC). On the other hand, the application of IVWP significantly reduces these biases as shown in Fig. 2 at those heights and times where nocturnal bird migration is expected. A unique feature of IVWP is its adaptability to different situations. Highly biased bird dominated VADs are automatically flagged as unreliable, while low bias insect dominated VADs are reliable. For mixed cases, containing near equal numbers of bird and insect echoes in the VAD estimation process, IVWP focuses on only insect gates, reducing VAD biases by up to 3 m/s. To the best of our knowledge, this is the first work that improves upon the conventional VWP by intelligently filtering out bird contamination.

link to the figure: https://www.dropbox.com/s/jqrvs80lg24z684/IVWP_AMS_conf2023.png?dl=0

200 m height, and within 1-11 UTC), and dominant insect presence in the low bias region (11 – 24 UTC). On the other hand, the application of IVWP significantly reduces these biases as shown in Fig. 2 at those heights and times where nocturnal bird migration is expected. A unique feature of IVWP is its adaptability to different situations. Highly biased bird dominated VADs are automatically flagged as unreliable, while low bias insect dominated VADs are reliable. For mixed cases, containing near equal numbers of bird and insect echoes in the VAD estimation process, IVWP focuses on only insect gates, reducing VAD biases by up to 3 m/s. To the best of our knowledge, this is the first work that improves upon the conventional VWP by intelligently filtering out bird contamination. \n\nlink to the figure: https://www.dropbox.com/s/jqrvs80lg24z684/IVWP_AMS_conf2023.png?dl=0\n"}" data-sheets-userformat="{"2":513,"3":{"1":0},"12":0}">