STRESZCZENIE

This study presents the development and evaluation of an algorithm designed to filter noise from the weather radar reflectivity data. The algorithm was tested on composite radar products from the POLRAD network operated by the Polish Institute of Meteorology and Water Management (IMGW). The algorithm’s steps include spatial data transformation, clustering using the DBSCAN algorithm, and classification based on the shapes of identified structures.
The filtering process focuses on the typical spatial patterns of radar noise, including RLAN interferences (spikes) and circular noise structures. It uses the radial and angular characteristics of clusters. The algorithm was evaluated on 22 data time-points with varying noise conditions. The algorithm achieved 82% overall accuracy. For the noise detection program reached 53.4% precision and 50% recall. While the method effectively detects isolated noise structures, its performance decreases when noise merges with precipitation in clustered regions.
Limitations of the current implementation are primarily related to the clustering stage. Misclassification occurs due to overlap of different structures. At the current stage, the algorithm provides a valuable baseline for further improvements. Possible next iterations may involve more advanced shape analysis or machine-learning–based clustering and classification.