Bridging data gaps in meteor radar wind measurements: A hybrid Machine Learning approach for Atmospheric Dynamics

Specular meteor radars (SMRs) have been extensively used to investigate neutral winds in the mesosphere and lower thermosphere. Unlike other instruments, this type of radar has the advantage of operating continuously, regardless of the weather conditions. However, SMRs experience interruptions in their operation, which result in data gaps that can range from a few hours to several days, posing a challenge to continuous atmospheric analysis. These gaps, on the other hand, present an excellent opportunity to test and validate the predicting capabilities of advanced Machine Learning (ML) techniques for data imputation. In this study, we employ a robust three-step data imputation methodology to sequentially address different types of data gaps artificially introduced in the wind data from the MMARIA/SIMONe meteor radar networks in northern Germany and northern Norway. The three-step imputation protocol proceeds as follows: initially, isolated missing values are estimated using classical time-based interpolation methods. Subsequently, continuous missing values over a period, where concurrent data across neighboring heights are available, are addressed using traditional space-based machine learning methods, such as k-Nearest Neighbor (kNN) or Random Forest (RF). Finally, the most challenging gap type, continuous missing samples, defined as contiguous space-temporal data loss during a specified period, is predicted using a deep learning stepwise extrapolation model based on the Long Short-Term Memory (LSTM) network. The methodology is applied to a four-month period of multistatic SMR wind data encompassing the major sudden stratospheric warming (SSW) event of January 2024. Special attention will be given to comparing the predicted wind behavior against the observed wind data.