Hybrid Convolution Reparameterization for Efficient Deep Learning-Based Nonprecipitation Echo Recognition and Removal

Weather radar reflectivity images play a critical role in reliable weather monitoring and forecasting. However, inherent factors such as ground clutter, sea clutter, and electromagnetic interference frequently introduce nonprecipitation echoes (NPEs) into these data, posing significant challenges for accurate precipitation detection. A promising solution is leveraging deep learning networks to identify and remove NPEs using satellite observations. To enhance the practicality of these models, recent advancements in reparameterization technology have shown potential for reducing computational complexity. However, existing methods focus on parallel multibranch reparameterization; the fusion of multiple convolutions into single equivalent convolutions, referred to as multiconvolution reparameterization, is still not explored. In this study, we propose a novel reparameterized NPE removal network (RepNPE-Net), designed to recognize and remove NPEs of radar reflectivity data using multichannel brightness temperature (BT) observations from a geostationary meteorological satellite. Reparameterized NPE removal network (RepNPE)-Net incorporates two innovative dual-stream convolution structure-based modules, including the reparameterized dual-stream convolutional module (RepDCM) and the reparameterized attention dual-stream convolutional module (RepADCM), which synergize standard and depthwise separable (DS) residual convolutional blocks to improve feature extraction and representation capabilities. Within the RepADCM module, a positional efficient local attention (PELA) block is designed to enable the network to focus on spatially significant positional features and enhance model's accuracy. Furthermore, to strengthen the practical application ability of the proposed RepNPE-Net, we introduce hybrid convolution reparameterization (HCR) technology, which consolidates multibranch and multiconvolution operations (e.g., depthwise (DW) and pointwise (PW) convolutions) into single equivalent convolutions during the inference stage, significantly reducing computational complexity without compromising network performance. Experimental results demonstrate that RepNPE-Net outperforms existing methods in both NPE removal accuracy and computational efficiency, highlighting its potential for improving radar data quality and advancing meteorological research and applications.