A Performance Evaluation of Sea Clutter and Chaff Classes for the WSR-88D Hydrometeor Classification Algorithm

The dual-polarimetric Weather Surveillance Radar–1988 Doppler (WSR-88D) network is capable of determining hydrometeor types via the Hydrometeor Classification Algorithm (HCA). While there are ten possible classes, the only non-hydrometeorological options are Ground Clutter and Biologicals. Several common targets do not fall into these classes, such as sea clutter, chaff, smoke, and radio frequency interference. Recently, a new HCA class known as Inanimate was developed based on the traditional HCA fuzzy logic technique, and is capable of isolating these four target types. However, more-advanced methods are necessary to separate these targets from each other due to significant overlap in their polarimetric statistics. In this study, a series of image processing steps are undertaken with Inanimate classifications, leading to areas of clustered detections. These clusters are then tested for 152 statistics that are used as input features to a Support Vector Machine (SVM), including all of the standard and polarimetric variables, a series of spatial statistics, and an approximation of circular depolarization ratio. Using a database of ground clutter, birds, smoke, insects, chaff, and sea clutter collected and human-truthed, including over 1,700 total cases, the SVM categorizes each cluster as sea clutter, chaff, or other Inanimate. The final HCA output uses these SVM categories to add two new classes to the existing HCA (Sea Clutter and Chaff), while the “other” category is returned to its original HCA classification. This study presents an exhaustive performance evaluation of the algorithm, with all results processed through the WSR-88D Open Radar Product Generator (ORPG) framework, demonstrating both operational readiness and significant skill. Discussions on feature selection, hyperparameter tuning, training/test/validation methods, and evaluation metrics are provided, including common skill scores, precision-recall and ROC curves, and holdout/cross-fold validations, which are both over 95% in the most-recent versions of the model. Applicability to real-time processing within the ORPG framework is detailed.