Extracting Bird and Insect Migration Echoes From Single-Polarization Weather Radar Data Using Semi-Supervised Learning

Weather radar serves as a crucial tool for monitoring aeroecology by enabling the observation of migrating birds and insects. Although dual-polarization weather radar offers the possibility of classifying echoes, extracting migration echoes of birds and insects from historical single-polarization weather radar data remains challenging. The current deep-learning methods have been successfully extracting aerial migrations from single-polarization weather radar data. However, it still faces challenges in distinguishing between birds and insects at the pixel level, primarily due to the absence of distinct semantic features for each. To tackle this challenge, we propose a semi-supervised radar data processing framework, which generates a large number of single polarization training datasets from a small amount of dual polarization truth data and trains the image segmentation network of single polarization data to distinguish between bird and insect echoes. The framework comprises three components: an image classifier, an image generator, and an image segmentation model. Specifically, the image classifier and image generator leverage a small set of manually annotated dual-polarization radar data to generate the pixel-level single-polarization dataset for training the image segmentation model. The well-trained image segmentation model extracts migration echoes of birds and insects from radar images. Experimental results demonstrate that the proposed method achieves a mean intersection over union (IoU) of 97% for segmenting precipitation, bird, and insect targets. The proposed framework can utilize historical archived single-polarization weather radar data to provide large-scale, long-term, and repeatable monitoring data for birds and insects.