Ka-Band Cloud Radar Meteorological Echo Dataset With Complex Weather Coverage: Baseline Models for Deep Learning Applications

Hydrometeors are essential to climate change and radiative budget research, making its observation important. Ka-band radar, valued for high spatial resolution and strong penetration, is widely used for observing hydrometeors. But it is prone to interference from insects and turbulence, called clear-air echoes or clutter. To mitigate interference from clear-air echoes, a deep learning dataset using Ka-band radar and lidar observations is designed. It incorporates synchronized radar-lidar data and expert-reviewed manual corrections, covering diverse weather conditions to support robust algorithm evaluation and development. Analysis shows clear-air echoes account for 16.7% of echoes within a range of 0–15 km, causing significant errors in hydrometeors observation if unclassified. A UNet-based baseline model for meteorological echo recognition is presented. The model with squeeze-and-excitation (SE) modules achieves 95.6% mean intersection over union (mIoU), with inference time reduced to 26.9% of the previous algorithm and using just 3.2% GPU time. Sensitivity analysis indicates that reflectivity and Doppler velocity channels contribute the most to prediction accuracy (15.1% and 1.7%, respectively), while the linear depolarization ratio (LDR) adds less than 1.0%. Complementary statistical metrics—including Kullback-Leibler (KL) divergence and Kolmogorov-Smirnov (KS) statistic—further confirm the superior separability of spatial texture features derived from reflectivity. These findings demonstrate that LDR, although useful in manual identification, is not essential for automated classification. This study contributes an open-access, annotated radar-lidar dataset and strong baseline models, offering a reliable benchmark for future research on radar echo classification.