Application of model-based spectral analysis to wind-profiler radar observations

A classical way to reduce a radar's data is to compute the spectrum using FFT and then to identify the different peak contributions. But in case an overlapping between the different echoes (atmospheric echo, clutter, hydrometeor echo...) exists, Fourier-like techniques provide poor frequency resolution and then sophisticated peak-identification may not be able to detect the different echoes. In order to improve the number of reduced data and their quality relative to Fourier spectrum analysis, three different methods are presented in this paper and applied to actual data. Their approach consists of predicting the main frequency-components, which avoids the development of very sophisticated peak-identification algorithms. The first method is based on cepstrum properties generally used to determine the shift between two close identical echoes. We will see in this paper that this method cannot provide a better estimate than Fourier-like techniques in an operational use. The second method consists of an autoregressive estimation of the spectrum. Since the tests were promising, this method was applied to reduce the radar data obtained during two thunderstorms. The autoregressive method, which is very simple to implement, improved the Doppler-frequency data reduction relative to the FIT spectrum analysis. The third method exploits a MUSIC algorithm, one of the numerous subspace-based methods, which is well adapted to estimate spectra composed of pure lines. A statistical study of performances of this method is presented, and points out the very good resolution of this estimator in comparison with Fourier-like techniques. Application to actual data confirms the good qualities of this estimator for reducing radar's data.