A Deep Learning-Based Image Segmentation Model for Extracting Migratory Birds and Insects Echoes Using Polarimetric Weather Radar Data

Weather radars are essential for monitoring airborne animals, including birds and insects. Despite the potential of dual-polarization radar moments for echo classification, extracting migration echoes from birds and insects at pixel-level remains challenging. In this study, six radar moments were rendered into two RGB images, and a deep learning-based image segmentation model was developed to extract bird and insect echoes at the pixel level. The model uses the typical encoder-decoder architecture. The ResNet-50 model was used to extract the features on different spatial scales, which was modified to accept six input channels, matching the six radar moments of polarimetric weather radar data during the encoding stages. The U-Net model was employed for multiscale feature fusion during the decoding stages. Training and validation on a manually labeled dataset demonstrated that the proposed method performs exceptionally well, achieving a mean intersection over union of 98%. The model's performance was further evaluated using two representative cases.