Addressing the Labelled Observation Data Gap for Biological and Meteorological Radar Echo Classification: A Transfer Learning Framework Using NEXRAD Data

ABSTRACT

Discriminating biological from meteorological radar echoes is essential for precipitation estimation, data quality control and ecological monitoring. In tropical regions, such as Rwanda, no labelled biological reference data exist to calibrate standard algorithms. We adapt three techniques to Rwanda's C‐band dual‐polarisation radar and quantify their transferability from S‐band NEXRAD: the physics‐based depolarisation ratio (DR) method, MistNet, a U‐Net convolutional network and RadxPID, a fuzzy‐logic hydrometeor classifier. Four volumes are evaluated: S‐band mixed echo and clear‐air migration, C‐band fall armyworm season and light rain. Skill is measured using fraction of echoes identified (FEI) and Heidke skill score (HSS); cross‐band transfer by distribution overlap (OVL), Hellinger distance and Cohen's kappa. On S‐band, DR reaches FEI = 75.3% (meteorological) and 99.4% (biological) with HSS = 0.75, exceeding MistNet (0.66) and RadxPID (0.52). Across bands, ρHV transfers best (OVL = 0.76) whereas ZH shifts most (OVL = 0.45), explaining why DR degrades gracefully at C‐band whereas MistNet's biological classification collapses below 2% of gates. Biological classification follows the boundary layer with range. Unanimous inter‐method agreement is 50%–61% on S‐band and 55% on C‐band. Physics‐based methods retain cross‐band coverage; data‐driven methods need retraining. DR is the most deployable option for data‐limited tropical environments.