Jensen–Shannon Distance-Based Filter and Unsupervised Evaluation Metrics for Polarimetric Weather Radar Processing

Abstract

An effective filtering technique is required for rainfall rate measurement by weather radar. A Jensen–Shannon Distance (JSD) based thresholding filter is proposed to mitigate non-meteorological signals, either in clear air or rain situations. This algorithm classifies range-Doppler bins into two classes, hydrometeors and non-hydrometeors, based on spectral polarimetric variable features. The result is a mask to be applied on the spectrograms. The variable selected here is the spectral co-polar correlation coefficient, available in dual-polarization and full polarimetric radars. The algorithm first does a global thresholding by finding an optimized threshold value based on the averaged clear air spectral polarimetric variable distribution. Next, classical filtering steps are carried out like a ground clutter notch filter around 0 ms^-1, a mathematical morphology to fill gaps in the hydrometeor areas and a removal of narrow Doppler power spectra. The second part of this article is the assessment of filtering techniques without ground truth. An assessment without ground-truth is useful to select optimal algorithm configurations from a large solution space. Criteria of good filtering are defined both in the spectral and time domain. Based on those criteria, subjective and objective unsupervised evaluation metrics are derived, with a focus on the objective ones. Data including clear air and rain collected from a full polarimetric Doppler X-band radar in urban area are used. With the proposed unsupervised evaluation metrics, the JSD-based thresholding filter is compared to two spectral polarimetric filters. Overall, the JSD-based filter performs very well considering both the subjective and the objective evaluation metrics.